{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.11","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":71549,"databundleVersionId":8561470,"sourceType":"competition"},{"sourceId":9729124,"sourceType":"datasetVersion","datasetId":5953414}],"dockerImageVersionId":31040,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Import and Config","metadata":{}},{"cell_type":"code","source":"import gc\nimport wandb\nfrom pytorch_lightning.loggers import WandbLogger\nimport os\nimport yaml\nimport sys\nimport cv2\nimport random\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport pandas as pd\nimport torch\nfrom glob import glob\nfrom torch.utils.data import Dataset\nfrom torch.utils.data import DataLoader\nfrom torch.optim import AdamW, Adam\nimport torch.nn as nn\nimport pytorch_lightning as pl\nfrom pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping, TQDMProgressBar\nimport torchvision.transforms as T\nimport albumentations as A\nimport pandas.api.types\nimport sklearn.metrics\nimport timm\nimport scipy\nimport albumentations as A\nfrom torchvision.transforms import v2\nfrom torchvision import models\nfrom tqdm.auto import tqdm\nfrom joblib import Parallel, delayed\nfrom torch.utils.data import default_collate\nimport pydicom as dcm\nimport transformers","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:08.807985Z","iopub.execute_input":"2025-06-07T01:05:08.808284Z","iopub.status.idle":"2025-06-07T01:05:08.814219Z","shell.execute_reply.started":"2025-06-07T01:05:08.808264Z","shell.execute_reply":"2025-06-07T01:05:08.813350Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"os.listdir('/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/test_images/44036939/2828203845')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:08.815345Z","iopub.execute_input":"2025-06-07T01:05:08.816008Z","iopub.status.idle":"2025-06-07T01:05:08.835951Z","shell.execute_reply.started":"2025-06-07T01:05:08.815989Z","shell.execute_reply":"2025-06-07T01:05:08.835296Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Anatomy & image Visual for more\nhttps://www.kaggle.com/code/abhinavsuri/anatomy-image-visualization-overview-rsna-raids#RSNA-Lumbar-Spine-Challenge.\n","metadata":{}},{"cell_type":"code","source":"# Đường dẫn đến folder chứa ảnh\nfolder = '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/test_images/44036939/2828203845'\n# Liệt kê các file DICOM trong thư mục\nfiles = [f for f in os.listdir(folder) if f.endswith('.dcm')]\n\n# Đọc một file (ví dụ: ảnh đầu tiên)\ndicom_path = os.path.join(folder, files[0])\ndemo_dcm = dcm.dcmread(dicom_path)\n\n# Chuyển pixel data sang numpy array\nimg = demo_dcm.pixel_array\n\n# Hiển thị ảnh\nplt.imshow(img, cmap='gray')\nplt.title(f\"File: {files[0]}\")\nplt.axis('off')\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:08.836614Z","iopub.execute_input":"2025-06-07T01:05:08.836840Z","iopub.status.idle":"2025-06-07T01:05:08.992166Z","shell.execute_reply.started":"2025-06-07T01:05:08.836824Z","shell.execute_reply":"2025-06-07T01:05:08.991267Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"SEED = 1710 # My birth day\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\ndef seed_everything(seed):\n random.seed(seed)\n os.environ['PYTHONHASHSEED'] = str(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed(seed)\n torch.backends.cudnn.deterministic = True # Fix the network according to random seed\n print('Finish seeding with seed {}'.format(seed))\n\nseed_everything(SEED)\nprint('Training on device {}'.format(device))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:08.993862Z","iopub.execute_input":"2025-06-07T01:05:08.994054Z","iopub.status.idle":"2025-06-07T01:05:09.003872Z","shell.execute_reply.started":"2025-06-07T01:05:08.994038Z","shell.execute_reply":"2025-06-07T01:05:09.003256Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%writefile config.yaml \ndata_path : '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification'\nout_put_dir : '/kaggle/working/models'\n\nseed : 1101\ndebug : False\ntrain_bs : 4\nvalid_bs : 4\ntest_bs : 8\nworker : 1\n\nprogress_bar_refresh_rate : 1\n\npseudo_train : 0\n\nsave_topk : 1\nfold : 5 # Cross Validation \n\ntask:\n kind: 'detect' # coordinate x, y\n #kind : 'classify' # severity -> label\n #kind : 'depth'\n condition: 'nfn'\n #condition: 'scs'\n #condition: 'scs'\n #condition: 'all'\n #direction: 'satg2'\n direction: 'ax'\n #direction: 'sagt1'\n position:\n - 'L1/L2'\n - 'L2/L3'\n - 'L3/L4'\n - 'L4/L5'\n - 'L5/S1'\n\nin_chans : 3\n\nimage_size : 384\n\nmodel:","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.004599Z","iopub.execute_input":"2025-06-07T01:05:09.004807Z","iopub.status.idle":"2025-06-07T01:05:09.014779Z","shell.execute_reply.started":"2025-06-07T01:05:09.004791Z","shell.execute_reply":"2025-06-07T01:05:09.014205Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os \nos.listdir('/kaggle/input/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.015455Z","iopub.execute_input":"2025-06-07T01:05:09.015661Z","iopub.status.idle":"2025-06-07T01:05:09.025268Z","shell.execute_reply.started":"2025-06-07T01:05:09.015641Z","shell.execute_reply":"2025-06-07T01:05:09.024757Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"with open(\"config.yaml\", \"r\") as file_obj:\n config = yaml.safe_load(file_obj)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.025948Z","iopub.execute_input":"2025-06-07T01:05:09.026181Z","iopub.status.idle":"2025-06-07T01:05:09.037672Z","shell.execute_reply.started":"2025-06-07T01:05:09.026167Z","shell.execute_reply":"2025-06-07T01:05:09.037078Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Sử dụng train data trong mở debug mode\nif config['debug']:\n IMAGE_PATH = '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/train_images/'\n series = pd.read_csv('/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/train_series_descriptions.csv')\nelse:\n IMAGE_PATH = '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/test_images/'\n series = pd.read_csv('/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/test_series_descriptions.csv')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.039841Z","iopub.execute_input":"2025-06-07T01:05:09.040051Z","iopub.status.idle":"2025-06-07T01:05:09.061716Z","shell.execute_reply.started":"2025-06-07T01:05:09.040032Z","shell.execute_reply":"2025-06-07T01:05:09.060985Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(IMAGE_PATH)\nseries.head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.062389Z","iopub.execute_input":"2025-06-07T01:05:09.062640Z","iopub.status.idle":"2025-06-07T01:05:09.081477Z","shell.execute_reply.started":"2025-06-07T01:05:09.062623Z","shell.execute_reply":"2025-06-07T01:05:09.080819Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Stage 0: intial (create meta file)\n* tạo metafile từ meta data của ảnh dicom: không chỉ chứa ảnh mà nhiều meta data khác\n* tạo một data frame - file chứa metadata quan trọng từ một series ảnh thuộc một study_id (phiên chụp bệnh nhân)","metadata":{}},{"cell_type":"code","source":"def create_dcm_df(study_id, series_id, series_description):\n try:\n # Find all .dcm to extract Instance_number từ tên file 1, 2, 3,... /* để tìm all file dcm\n path_list = glob(IMAGE_PATH + f'{study_id}/{series_id}/*.dcm')\n in_list = sorted([int(s.split('/')[-1].split('.')[0]) for s in path_list])\n\n # read file dicom by pydicom\n dcm_list = []\n for i in in_list:\n dcm_list.append(dcm.dcmread(IMAGE_PATH + f'{study_id}/{series_id}/{i}.dcm'))\n \n '''\n Extract metadata chính:\n ipp : vị trí x,y,z của ảnh trong cơ thể\n ioo : hướng trục của ảnh trong không gian\n '''\n ipp = np.asarray([d.ImagePositionPatient for d in dcm_list]).astype('float')\n iop = [d.ImageOrientationPatient for d in dcm_list]\n iop = [[float(d[0]), float(d[1]), float(d[2]), \n float(d[3]), float(d[4]), float(d[5])] for d in iop]\n ipp_x = ipp[:, 0]\n ipp_y = ipp[:, 1]\n ipp_z = ipp[:, 2]\n\n # Extract độ phân giải (Pixel Spacing) và kích thước\n shape = np.array([d.pixel_array.shape for d in dcm_list])\n sbs = np.asarray([d.SpacingBetweenSlices for d in dcm_list]).astype('float')\n ps = np.asarray([d.PixelSpacing for d in dcm_list]).astype('float') # x, y\n ps_x = ps[:, 0]\n ps_y = ps[:, 1]\n\n # create dataframe chứa metadata\n meta_dict = {\n 'instance_number' : in_list,\n 'ipp_x' : ipp_x,\n 'ipp_y' : ipp_y,\n 'ipp_z' : ipp_z,\n 'sbs' : sbs,\n 'ps_x' : ps_x, \n 'ps_y' : ps_y\n }\n meta_df = pd.DataFrame(meta_dict)\n\n #add other fearture\n meta_df['series_id'] = series_id\n meta_df['study_id'] = study_id\n meta_df['series_description'] = series_description\n meta_df['height'] = shape[:, 0]\n meta_df['width'] = shape[:, 1]\n meta_df['iop'] = pd.Series(iop)\n \n # xoá dữ liệu giải phóng bộ nhớ\n del dcm_list, ipp, iop, sbs, ps\n gc.collect()\n return meta_df[['study_id', 'series_id', 'series_description', 'instance_number', 'height', 'width', 'ipp_x', 'ipp_y', 'ipp_z', 'iop', 'sbs', 'ps_x', 'ps_y']]\n except:\n print(study_id, series_id, series_description)\n return None\n ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.082396Z","iopub.execute_input":"2025-06-07T01:05:09.082642Z","iopub.status.idle":"2025-06-07T01:05:09.091255Z","shell.execute_reply.started":"2025-06-07T01:05:09.082623Z","shell.execute_reply":"2025-06-07T01:05:09.090625Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"create_dcm_df(44036939, 2828203845, 'Sagittal T1').head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.092136Z","iopub.execute_input":"2025-06-07T01:05:09.092385Z","iopub.status.idle":"2025-06-07T01:05:09.904230Z","shell.execute_reply.started":"2025-06-07T01:05:09.092363Z","shell.execute_reply":"2025-06-07T01:05:09.903469Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"**Turn into .parquet file**","metadata":{}},{"cell_type":"code","source":"%%time\nif config['debug']:\n meta_df_list = []\n \n # Parallbel song song các \n meta_df_list = Parallel(n_jobs=-1)([delayed(create_dcm_df)(row.study_id, row.series_id, row.series_description) for _, row in series.iterrows()])\n\n # merge to big dataframe\n meta_df = pd.concat(meta_df_list)\n del meta_df_list\n gc.collect()\n\n #turn to file parquet\n meta_df.to_parquet('meta.parquet')\nelse:\n meta_df_list = []\n meta_df_list = Parallel(n_job = -1)([delayed(create_dcm_df)(row.study_id, row.series_id, row.series_description) for _, row in series.iterrows()])\n\n meta_df = pd.concat(meta_df_list)\n del meta_df_list\n gc.collect()\n meta_df.to_parquet('meta.parquet')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:09.904878Z","iopub.execute_input":"2025-06-07T01:05:09.905128Z","iopub.status.idle":"2025-06-07T01:05:12.535926Z","shell.execute_reply.started":"2025-06-07T01:05:09.905106Z","shell.execute_reply":"2025-06-07T01:05:12.535295Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Kiểm tra file parquet: file lớn thích hợp để lưu với nhiều dữ liệu hơn csv\nimport os \nos.listdir('/kaggle/working/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:12.536779Z","iopub.execute_input":"2025-06-07T01:05:12.537165Z","iopub.status.idle":"2025-06-07T01:05:12.541866Z","shell.execute_reply.started":"2025-06-07T01:05:12.537138Z","shell.execute_reply":"2025-06-07T01:05:12.541263Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# First stage (depth inference)\n* infer - suy luận depth từ ảnh sagittal T1 và T2\n* căn chỉnh độ sâu suy luận theo luật thuật toán - align infered depth using rule base algorithims","metadata":{}},{"cell_type":"markdown","source":"## Custom dataset\n* Hẹp lỗ liên hợp trái (left nfn)\n* Hẹp lỗ liên hợp phải (right nfn)\n* Hẹp dưới khớp trái (left ss)\n* Hẹp dưới khớp phải (right ss)\n* Hẹp ống sống (scs)\n","metadata":{}},{"cell_type":"markdown","source":"condition: xác định loại ảnh MRI nào dùng, gồm:\n\n'scs': Sagittal T2/STIR\n\n'nfn': Sagittal T1","metadata":{}},{"cell_type":"code","source":"# # Bước đầu của pipeline \n# class DepthDetectDataset(Dataset):\n# def __init__(self, meta, condition, usage = 'sub'):\n# if condition == 'scs':\n# meta = meta.loc[meta.series_description == 'Sagittal T2/STIR']\n# else:\n# meta = meta.loc[meta.series_description == 'Sagittal T1']\n# self.id = list(meta.study_id.unique())\n# if 3637444890 in self.id:\n# self.id.remove(3637444890)\n# self.meta = meta\n# self.condition = condition\n# self.usage = usage\n# self.resize = v2.Resize(384, 384)\n\n# def for_scs(self, study_id):\n# depth = 32\n# #lọc dữ liệu với 1 study_id\n# meta = self.meta.loc[(self.meta.study_id == study_id) & (self.meta.series_decription == 'Sagittal T2/STIR')]\n \n# #sắp xếp trục theo không gian x\n# meta = meta.sort_values('ipp_x', ascending = True).reset_index(drop = True)\n \n# #đọc ảnh dicom thành mảng 2d numpy\n# img = [self.load_dicom(IMAGE + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm') for _, row in meta.iterrows()]\n \n# #resize, chuyển ảnh thành tensor rồi stack lại\n# volume = self.normalize(torch.cat([self.resize(torch.tensor(i.astype(np.float32))[None, ...]).to(torch.float32) for i in img]).contiguous())\n \n# #chuẩn hoá độ sâu depth\n# if volume.shape[0] < depth:\n# volume = torch.cat([volume, torch.zeros(depth-volume.shape[0], volume.shape[1], volume.shape[2])])\n# elif volume.shape[0] > depth:\n# volume = torch.nn.fuctional.interpolate(volume[None, None, ...], (depth, volume.shape[1], volume.shape[2])).squeeze()\n# return volume.to(torch.float32)\n \n# def for_nfn(self, study_id):\n# depth = 32\n# #lọc dữ liệu với 1 study_id\n# meta = self.meta.loc[(self.meta.study_id == study_id) & (self.meta.series_decription == 'Sagittal T1')]\n \n# #sắp xếp trục theo không gian x\n# meta = meta.sort_values('ipp_x', ascending = True).reset_index(drop = True)\n \n# #đọc ảnh dicom thành mảng 2d numpy\n# img = [self.load_dicom(IMAGE + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm') for _, row in meta.iterrows()]\n \n# #resize, chuyển ảnh thành tensor rồi stack lại\n# volume = self.normalize(torch.cat([self.resize(torch.tensor(i.astype(np.float32))[None, ...]).to(torch.float32) for i in img]).contiguous())\n \n# #chuẩn hoá độ sâu depth\n# if volume.shape[0] < depth:\n# volume = torch.cat([volume, torch.zeros(depth-volume.shape[0], volume.shape[1], volume.shape[2])])\n# elif volume.shape[0] > depth:\n# volume = torch.nn.fuctional.interpolate(volume[None, None, ...], (depth, volume.shape[1], volume.shape[2])).squeeze()\n# return volume.to(torch.float32)\n \n# def __getitem__(self, index):\n# study_id = self.id[index]\n\n# if self.condition == 'scs':\n# volume = self.for_scs(study_id)\n# elif self.condition == 'nfn':\n# try:\n# volume = self.for_nfn(study_id)\n# except:\n# print(study_id)\n# return volume, torch.tensor([study_id])\n\n# def normalize(self, x):\n# upper = torch.quantile(x, torch.tensor([0.99]))\n# lower = torch.quantile(x, torch.tensor([0.01]))\n# x = torch.clip(x, lower, upper)\n# x = x - torch.min(x)\n# x = x / (torch.max(x)+1e-6)\n\n# def __len__(self):\n# return len(self.id)\n\n# def load_dcm(self, path):\n# dicom = dcm.read_file(path)\n# data = dicom.pixel_array\n# return data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:12.542648Z","iopub.execute_input":"2025-06-07T01:05:12.543430Z","iopub.status.idle":"2025-06-07T01:05:12.553024Z","shell.execute_reply.started":"2025-06-07T01:05:12.543409Z","shell.execute_reply":"2025-06-07T01:05:12.552357Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class DepthDetectDataset(Dataset):\n def __init__(self, meta, condition, usage='sub'):\n if condition == 'scs': \n meta = meta.loc[meta.series_description=='Sagittal T2/STIR']\n else: \n meta = meta.loc[meta.series_description=='Sagittal T1']\n self.id = list(meta.study_id.unique())\n if 3637444890 in self.id: \n self.id.remove(3637444890)\n self.meta = meta\n self.condition = condition\n self.usage = usage\n \n self.resize = v2.Resize((384, 384))\n \n def __getitem__(self, index):\n study_id = self.id[index]\n #print(study_id)\n #try:\n if self.condition == 'scs':\n volume = self.for_scs(study_id)\n elif self.condition == 'nfn':\n try: \n volume = self.for_nfn(study_id)\n except: \n print(study_id)\n return volume, torch.tensor([study_id])\n\n def for_scs(self, study_id):\n depth = 32\n meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T2/STIR')]\n meta = meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n img = [self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm') for _, row in meta.iterrows()]\n volume = self.normalize(torch.cat([self.resize(torch.tensor(i.astype(np.float32))[None, ...]).to(torch.float32) for i in img]).contiguous())\n if volume.shape[0] < depth:\n volume = torch.cat([volume, torch.zeros(depth-volume.shape[0], volume.shape[1], volume.shape[2])])\n elif volume.shape[0] > depth:\n volume = torch.nn.functional.interpolate(volume[None, None, ...], (depth, volume.shape[1], volume.shape[2])).squeeze()\n return volume.to(torch.float32)\n\n def for_nfn(self, study_id):\n depth = 32\n meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T1')]\n meta = meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n img = [self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm') for _, row in meta.iterrows()]\n volume = self.normalize(torch.cat([self.resize(torch.tensor(i.astype(np.float32))[None, ...]).to(torch.float32) for i in img]).contiguous())\n if volume.shape[0] < depth:\n volume = torch.cat([volume, torch.zeros(depth-volume.shape[0], volume.shape[1], volume.shape[2])])\n elif volume.shape[0] > depth:\n volume = torch.nn.functional.interpolate(volume[None, None, ...], (depth, volume.shape[1], volume.shape[2])).squeeze()\n return volume.to(torch.float32)\n \n def normalize(self, x):\n upper = torch.quantile(x, torch.tensor([0.99]))\n lower = torch.quantile(x, torch.tensor([0.01]))\n x = torch.clip(x, lower, upper)\n x = x - torch.min(x)\n x = x / (torch.max(x)+1e-6)\n return x\n\n def __len__(self):\n return len(self.id)\n\n def load_dicom(self, path):\n dicom = dcm.dcmread(path)\n data = dicom.pixel_array\n return data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:12.555748Z","iopub.execute_input":"2025-06-07T01:05:12.555935Z","iopub.status.idle":"2025-06-07T01:05:12.573133Z","shell.execute_reply.started":"2025-06-07T01:05:12.555920Z","shell.execute_reply":"2025-06-07T01:05:12.572416Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"demo_path = '/kaggle/working/meta.parquet'\nmeta_df = pd.read_parquet(demo_path)\ndataset_demo = DepthDetectDataset(meta=meta_df, condition = 'nfn')\nprint(len(dataset_demo))\ndataset_demo.__getitem__(0)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:12.573863Z","iopub.execute_input":"2025-06-07T01:05:12.574091Z","iopub.status.idle":"2025-06-07T01:05:13.920033Z","shell.execute_reply.started":"2025-06-07T01:05:12.574069Z","shell.execute_reply":"2025-06-07T01:05:13.919283Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Model\nsử dụng 3D ConvNext","metadata":{}},{"cell_type":"code","source":"from torchvision.ops import StochasticDepth \nfrom typing import List, Dict\nfrom torch import Tensor","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.920828Z","iopub.execute_input":"2025-06-07T01:05:13.921116Z","iopub.status.idle":"2025-06-07T01:05:13.924911Z","shell.execute_reply.started":"2025-06-07T01:05:13.921092Z","shell.execute_reply":"2025-06-07T01:05:13.924218Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Stem block: để sử lý đầu vào ảnh 3D và chuẩn hoá\n# 3D Conv - dùng kernel_size [D, H, W] - (1, 2, 2) - giữ nguyên depth, chỉ tích chập trên H, W. Tương tự với stride\n# Output h/2, w/2 với stride = 2\nclass ConvNextStem(nn.Sequential):\n def __init__(self, in_features: int, out_features : int):\n super().__init__(\n nn.Conv3d(in_features, out_features, kernel_size = (1, 2, 2), stride = (1, 2, 2)),\n nn.GroupNorm(num_groups = 1, num_channels = out_features)\n )\n \n# Tránh vanish: nhân hệ số x vơi gamma (learnable scaling) học được\nclass LayerScaler(nn.Module):\n def __init__(self, init_value : float, dimensions : int):\n super.__init__()\n self.gamma = nn.Parameter(init_value * torch.ones((dimensions)), requires_grad = True)\n\n def forward(self, x):\n return self.gamma[None, ..., None, None]* x\n\n# Residual bottleneck block trong các kiến trúc hiện đại: \n# Học được nhiều đặc trưng hơn thông qua: expansion, ...\nclass BottleNeckBlock(nn.Module):\n def __init__(\n self, \n in_features : int,\n out_features: int,\n expansion : int = 4,\n drop_p : float = .0,\n layer_scaler_init_value : float = 1e-6,\n ):\n super().__init__()\n expanded_features = out_features * expansion \n \n # 3 lần Conv để lấy được nhiều feature đặc trưng hơn\n self.block = nn.Sequential(\n nn.Conv3d(\n in_features, in_features, kernel_size = (2, 7, 7), padding = 'same', bias = False, groups = in_features\n ),\n\n # chuẩn hoá theo channel\n nn.GroupNorm(num_groups = in_features, num_channels = in_features),\n nn.Conv3d(in_features, expanded_features, kernel_size = 1),\n nn.GELU(),\n nn.Conv3d(expanded_features, out_features, kernel_size = 1),\n )\n\n def forward(self, x : Tensor) -> Tensor:\n res = x\n x = self.block(x)\n x += res\n return x","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.925850Z","iopub.execute_input":"2025-06-07T01:05:13.926092Z","iopub.status.idle":"2025-06-07T01:05:13.938569Z","shell.execute_reply.started":"2025-06-07T01:05:13.926072Z","shell.execute_reply":"2025-06-07T01:05:13.937928Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# tăng channel không đồng nghĩa với học được nhiều feature hữu ích\nclass ConvNexStage(nn.Sequential):\n def __init__(\n self, in_features: int, out_features: int, depth: int, **kwargs\n ):\n super().__init__(\n # Downsampler giảm độ phân giải không gian (D, H, W), tăng số lượng channel.\n nn.Sequential(\n nn.GroupNorm(num_groups=in_features, num_channels=in_features),\n nn.Conv3d(in_features, out_features, kernel_size=(2, 2, 2), stride=(2, 2, 2))\n ),\n # BottleNeckBlock: áp dụng nhiều khối tích chập (Conv3D) nâng cao biểu diễn đặc trưng.\n *[\n BottleNeckBlock(out_features, out_features, **kwargs)\n for _ in range(depth)\n ],\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.939203Z","iopub.execute_input":"2025-06-07T01:05:13.939360Z","iopub.status.idle":"2025-06-07T01:05:13.952977Z","shell.execute_reply.started":"2025-06-07T01:05:13.939348Z","shell.execute_reply":"2025-06-07T01:05:13.952368Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class ConvNextEncoder(nn.Module):\n def __init__(\n self,\n in_channels: int,\n stem_features: int,\n depths: List[int],\n widths: List[int],\n drop_p: float = .0, # regularization\n ):\n super().__init__()\n # giảm khích thước ảnh đầu vào (spatial downsampling), tăng số lượng features - channel\n self.stem = ConvNextStem(in_channels, stem_features)\n\n in_out_widths = list(zip(widths, widths[1:]))\n # create drop paths probabilities (one for each stage)\n drop_probs = [x.item() for x in torch.linspace(0, drop_p, sum(depths))]\n\n # sử dụng các NexStage để trích xuất được feature trừu tượng\n self.stages = nn.ModuleList(\n [\n ConvNexStage(stem_features, widths[0], depths[0], drop_p=drop_probs[0]),\n *[\n ConvNexStage(in_features, out_features, depth, drop_p=drop_p)\n for (in_features, out_features), depth, drop_p in zip(\n in_out_widths, depths[1:], drop_probs[1:]\n )\n ],\n ]\n )\n\n\n def forward(self, x):\n # qua stem để tiền sử lý\n x = self.stem(x)\n\n #rồi đi qua từng stage để extract\n for stage in self.stages:\n x = stage(x)\n return x","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.953642Z","iopub.execute_input":"2025-06-07T01:05:13.953801Z","iopub.status.idle":"2025-06-07T01:05:13.965990Z","shell.execute_reply.started":"2025-06-07T01:05:13.953788Z","shell.execute_reply":"2025-06-07T01:05:13.965264Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class ClassificationHead(nn.Sequential):\n def __init__(self):\n super().__init__(\n nn.AdaptiveAvgPool3d((1, 1, 1)), # (B, C, 1, 1, 1)\n nn.Flatten(1), # (B, C)\n nn.LayerNorm(512),\n nn.Linear(512, 3)\n )\nclass Flatten(nn.Sequential):\n def __init__(self):\n super().__init__(\n nn.AdaptiveAvgPool3d((1, 1, 1)),\n nn.Flatten(1),\n nn.LayerNorm(512)\n )","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.966744Z","iopub.execute_input":"2025-06-07T01:05:13.966953Z","iopub.status.idle":"2025-06-07T01:05:13.981063Z","shell.execute_reply.started":"2025-06-07T01:05:13.966930Z","shell.execute_reply":"2025-06-07T01:05:13.980263Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# #test phase\n# x = torch.randn(4, 512, 32, 384, 384)\n# test = ClassificationHead()\n# print(test(x).shape) \n# run to long must to gpu\nx = torch.randn(4, 512, 8, 8, 8)\ntest = Flatten()\nprint(test(x).shape) ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:13.981714Z","iopub.execute_input":"2025-06-07T01:05:13.981901Z","iopub.status.idle":"2025-06-07T01:05:14.008679Z","shell.execute_reply.started":"2025-06-07T01:05:13.981887Z","shell.execute_reply":"2025-06-07T01:05:14.008015Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# dự đoán độ sâu của đốt sống bên trái và bên phải\nclass ConvNextSSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n # Extract feature từ 3D ConvNextEncoder \n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n # Output là 10 branches: mỗi branch là 1 đoạn đốt sống\n self.ll1 = nn.Linear(512, 96)\n self.ll2 = nn.Linear(512, 96)\n self.ll3 = nn.Linear(512, 96)\n self.ll4 = nn.Linear(512, 96)\n self.ll5 = nn.Linear(512, 96)\n self.rl1 = nn.Linear(512, 96)\n self.rl2 = nn.Linear(512, 96)\n self.rl3 = nn.Linear(512, 96)\n self.rl4 = nn.Linear(512, 96)\n self.rl5 = nn.Linear(512, 96)\n # trả về 10 keys: mỗi key là tensor [B, 96]\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n # [B, 96]\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\n\n# Bổ sung thông tin về vị trí trong chuỗi vào embedding vector\nclass PositionalEncoding(nn.Module):\n \n def __init__(self, d_model: int, dropout: float = 0.1, max_len: int = 5000):\n super().__init__()\n self.dropout = nn.Dropout(p=dropout)\n \n # tạo vector position từ 0 - max_len-1\n position = torch.arange(max_len).unsqueeze(1)\n \n # tấn suất sin/cos \n div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))\n \n # Ánh xạ vào sin cos\n pe = torch.zeros(max_len, 1, d_model)\n pe[:, 0, 0::2] = torch.sin(position * div_term)\n pe[:, 0, 1::2] = torch.cos(position * div_term)\n \n # lưu pe và buffer \n self.register_buffer('pe', pe)\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x: Tensor, shape [batch_size, seq_len, embedding_dim]\n \"\"\"\n # Thêm encoding vào x. Sau đó chuyển lại về [B, T, D]\n x = x.permute(1, 0, 2)\n x = x + self.pe[:x.size(0)]\n return self.dropout(x.permute(1, 0, 2))\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.009351Z","iopub.execute_input":"2025-06-07T01:05:14.009598Z","iopub.status.idle":"2025-06-07T01:05:14.020382Z","shell.execute_reply.started":"2025-06-07T01:05:14.009577Z","shell.execute_reply":"2025-06-07T01:05:14.019681Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## pipeline 3 use pretrain model \n* Ở AttentionSSDepthDetect có thể bỏ unsqueeze trong forward","metadata":{}},{"cell_type":"code","source":"# pipeline 3 usepretrain model \nclass AttentionSSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n # load pretrained model ConvNeXt-base\n self.encoder = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=True, num_classes=0)\n \n self.flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n \n self.in_features = self.encoder.num_features\n\n # tạo position encoding dạng sin cos\n self.rpe = PositionalEncoding(self.in_features, dropout=0., max_len=64)\n \n self.transformer0 = nn.TransformerEncoderLayer(d_model=self.in_features, nhead=8, activation='gelu',dropout=0.1, batch_first=True)\n self.transformer1 = nn.TransformerEncoderLayer(d_model=self.in_features, nhead=8, activation='gelu',dropout=0.1, batch_first=True)\n\n # L1/L2 -> L5/S1\n self.ll1 = nn.Linear(512, 64)\n self.ll2 = nn.Linear(512, 64)\n self.ll3 = nn.Linear(512, 64)\n self.ll4 = nn.Linear(512, 64)\n self.ll5 = nn.Linear(512, 64)\n self.rl1 = nn.Linear(512, 64)\n self.rl2 = nn.Linear(512, 64)\n self.rl3 = nn.Linear(512, 64)\n self.rl4 = nn.Linear(512, 64)\n self.rl5 = nn.Linear(512, 64)\n def forward(self, x, label=None):\n # Đảm bảo là x có 3 kênh đầu vào, nếu chỉ có 1 thì thêm 1 dims vào\n # hơi dư thừa nếu về mặt Conv2D\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x) # [B, 64]\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n\n # Trả về dictionary chứa 10 output embeddings\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.021033Z","iopub.execute_input":"2025-06-07T01:05:14.021237Z","iopub.status.idle":"2025-06-07T01:05:14.033897Z","shell.execute_reply.started":"2025-06-07T01:05:14.021223Z","shell.execute_reply":"2025-06-07T01:05:14.033147Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ConvNeXtNFNDepthDetect sử dụng backbone của ConvNeXt 3D để phát hiện bệnh nfn cột sống ở thắt lưng\nclass ConvNextNFNDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n # backbone (encoder)\n # Chức năng chính là trích xuất các đặc trưng từ 1 ảnh - volumn 3D MRI kích thước [B, 1, D, H, W]\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n # left l1 -> l5\n self.ll1 = nn.Linear(512, 32)\n self.ll2 = nn.Linear(512, 32)\n self.ll3 = nn.Linear(512, 32)\n self.ll4 = nn.Linear(512, 32)\n self.ll5 = nn.Linear(512, 32)\n\n # right l1 -> l5\n self.rl1 = nn.Linear(512, 32)\n self.rl2 = nn.Linear(512, 32)\n self.rl3 = nn.Linear(512, 32)\n self.rl4 = nn.Linear(512, 32)\n self.rl5 = nn.Linear(512, 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n # 10 Linear header dự đoán đặc trung cho từng kh đốt sống\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\nclass ConvNextNFNDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(widths[-1]))\n self.ll1 = nn.Linear(widths[-1], 32)\n self.ll2 = nn.Linear(widths[-1], 32)\n self.ll3 = nn.Linear(widths[-1], 32)\n self.ll4 = nn.Linear(widths[-1], 32)\n self.ll5 = nn.Linear(widths[-1], 32)\n self.rl1 = nn.Linear(widths[-1], 32)\n self.rl2 = nn.Linear(widths[-1], 32)\n self.rl3 = nn.Linear(widths[-1], 32)\n self.rl4 = nn.Linear(widths[-1], 32)\n self.rl5 = nn.Linear(widths[-1], 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.034564Z","iopub.execute_input":"2025-06-07T01:05:14.034776Z","iopub.status.idle":"2025-06-07T01:05:14.048688Z","shell.execute_reply.started":"2025-06-07T01:05:14.034753Z","shell.execute_reply":"2025-06-07T01:05:14.048114Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class ConvNextSCSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.l1 = nn.Linear(512, 32)\n self.l2 = nn.Linear(512, 32)\n self.l3 = nn.Linear(512, 32)\n self.l4 = nn.Linear(512, 32)\n self.l5 = nn.Linear(512, 32)\n def forward(self, x, label = None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n x = self.l1(x)\n x = self.l2(x)\n x = self.l3(x)\n x = self.l4(x)\n x = self.l5(x)\n \n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}\n \nclass ConvNextSCSDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.l1 = nn.Linear(widths[-1], 32)\n self.l2 = nn.Linear(widths[-1], 32)\n self.l3 = nn.Linear(widths[-1], 32)\n self.l4 = nn.Linear(widths[-1], 32)\n self.l5 = nn.Linear(widths[-1], 32)\n def forward(self, x, label = None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n x = self.l1(x)\n x = self.l2(x)\n x = self.l3(x)\n x = self.l4(x)\n x = self.l5(x)\n \n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.049255Z","iopub.execute_input":"2025-06-07T01:05:14.049486Z","iopub.status.idle":"2025-06-07T01:05:14.064776Z","shell.execute_reply.started":"2025-06-07T01:05:14.049466Z","shell.execute_reply":"2025-06-07T01:05:14.064079Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class RegConvNextNFNDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(1024))\n self.ll1 = nn.Linear(1024, 3)\n self.ll2 = nn.Linear(1024, 3)\n self.ll3 = nn.Linear(1024, 3)\n self.ll4 = nn.Linear(1024, 3)\n self.ll5 = nn.Linear(1024, 3)\n self.rl1 = nn.Linear(1024, 3)\n self.rl2 = nn.Linear(1024, 3)\n self.rl3 = nn.Linear(1024, 3)\n self.rl4 = nn.Linear(1024, 3)\n self.rl5 = nn.Linear(1024, 3)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x).sigmoid()\n ll2 = self.ll2(x).sigmoid()\n ll3 = self.ll3(x).sigmoid()\n ll4 = self.ll4(x).sigmoid()\n ll5 = self.ll5(x).sigmoid()\n rl1 = self.rl1(x).sigmoid()\n rl2 = self.rl2(x).sigmoid()\n rl3 = self.rl3(x).sigmoid()\n rl4 = self.rl4(x).sigmoid()\n rl5 = self.rl5(x).sigmoid()\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\n\nclass RegConvNextSCSDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(1024),\n )\n self.l1 = nn.Linear(1024, 3)\n self.l2 = nn.Linear(1024, 3)\n self.l3 = nn.Linear(1024, 3)\n self.l4 = nn.Linear(1024, 3)\n self.l5 = nn.Linear(1024, 3)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n l1 = self.l1(x).sigmoid()\n l2 = self.l2(x).sigmoid()\n l3 = self.l3(x).sigmoid()\n l4 = self.l4(x).sigmoid()\n l5 = self.l5(x).sigmoid()\n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.065421Z","iopub.execute_input":"2025-06-07T01:05:14.065656Z","iopub.status.idle":"2025-06-07T01:05:14.079234Z","shell.execute_reply.started":"2025-06-07T01:05:14.065638Z","shell.execute_reply":"2025-06-07T01:05:14.078526Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Base Models\n\n\nclass ConvNextStem(nn.Sequential):\n def __init__(self, in_features: int, out_features: int):\n super().__init__(\n nn.Conv3d(in_features, out_features, kernel_size=(1, 2, 2), stride=(1, 2, 2)),\n nn.GroupNorm(num_groups=1, num_channels=out_features)\n )\n\nclass LayerScaler(nn.Module):\n def __init__(self, init_value: float, dimensions: int):\n super().__init__()\n self.gamma = nn.Parameter(init_value * torch.ones((dimensions)),\n requires_grad=True)\n\n def forward(self, x):\n return self.gamma[None,...,None,None] * x\n\nclass BottleNeckBlock(nn.Module):\n def __init__(\n self,\n in_features: int,\n out_features: int,\n expansion: int = 4,\n drop_p: float = .0,\n layer_scaler_init_value: float = 1e-6,\n ):\n super().__init__()\n expanded_features = out_features * expansion\n self.block = nn.Sequential(\n # narrow -> wide (with depth-wise and bigger kernel)\n nn.Conv3d(\n in_features, in_features, kernel_size=(2, 7, 7), padding='same', bias=False, groups=in_features\n ),\n # GroupNorm with num_groups=1 is the same as LayerNorm but works for 2D data\n nn.GroupNorm(num_groups=in_features, num_channels=in_features),\n # wide -> wide\n nn.Conv3d(in_features, expanded_features, kernel_size=1),\n nn.GELU(),\n # wide -> narrow\n nn.Conv3d(expanded_features, out_features, kernel_size=1),\n )\n #self.layer_scaler = LayerScaler(layer_scaler_init_value, out_features)\n #self.drop_path = StochasticDepth(drop_p, mode=\"batch\")\n\n\n def forward(self, x: Tensor) -> Tensor:\n res = x\n x = self.block(x)\n #x = self.layer_scaler(x)\n #x = self.drop_path(x)\n x += res\n return x\n\nclass ConvNexStage(nn.Sequential):\n def __init__(\n self, in_features: int, out_features: int, depth: int, **kwargs\n ):\n super().__init__(\n # add the downsampler\n nn.Sequential(\n nn.GroupNorm(num_groups=in_features, num_channels=in_features),\n nn.Conv3d(in_features, out_features, kernel_size=(2, 2, 2), stride=(2, 2, 2))\n ),\n *[\n BottleNeckBlock(out_features, out_features, **kwargs)\n for _ in range(depth)\n ],\n )\n\nclass ConvNextEncoder(nn.Module):\n def __init__(\n self,\n in_channels: int,\n stem_features: int,\n depths: List[int],\n widths: List[int],\n drop_p: float = .0,\n ):\n super().__init__()\n self.stem = ConvNextStem(in_channels, stem_features)\n\n in_out_widths = list(zip(widths, widths[1:]))\n # create drop paths probabilities (one for each stage)\n drop_probs = [x.item() for x in torch.linspace(0, drop_p, sum(depths))]\n\n self.stages = nn.ModuleList(\n [\n ConvNexStage(stem_features, widths[0], depths[0], drop_p=drop_probs[0]),\n *[\n ConvNexStage(in_features, out_features, depth, drop_p=drop_p)\n for (in_features, out_features), depth, drop_p in zip(\n in_out_widths, depths[1:], drop_probs[1:]\n )\n ],\n ]\n )\n\n\n def forward(self, x):\n x = self.stem(x)\n for stage in self.stages:\n x = stage(x)\n return x\n\nclass ClassificationHead(nn.Sequential):\n def __init__(self):\n super().__init__(\n nn.AdaptiveAvgPool3d((1, 1, 1)),\n nn.Flatten(1),\n nn.LayerNorm(512),\n nn.Linear(512, 3)\n )\nclass Flatten(nn.Sequential):\n def __init__(self):\n super().__init__(\n nn.AdaptiveAvgPool3d((1, 1, 1)),\n nn.Flatten(1),\n nn.LayerNorm(512)\n )\n\n\nclass ConvNextSSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.ll1 = nn.Linear(512, 96)\n self.ll2 = nn.Linear(512, 96)\n self.ll3 = nn.Linear(512, 96)\n self.ll4 = nn.Linear(512, 96)\n self.ll5 = nn.Linear(512, 96)\n self.rl1 = nn.Linear(512, 96)\n self.rl2 = nn.Linear(512, 96)\n self.rl3 = nn.Linear(512, 96)\n self.rl4 = nn.Linear(512, 96)\n self.rl5 = nn.Linear(512, 96)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\nclass PositionalEncoding(nn.Module):\n\n def __init__(self, d_model: int, dropout: float = 0.1, max_len: int = 5000):\n super().__init__()\n self.dropout = nn.Dropout(p=dropout)\n\n position = torch.arange(max_len).unsqueeze(1)\n div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))\n pe = torch.zeros(max_len, 1, d_model)\n pe[:, 0, 0::2] = torch.sin(position * div_term)\n pe[:, 0, 1::2] = torch.cos(position * div_term)\n self.register_buffer('pe', pe)\n\n def forward(self, x: Tensor) -> Tensor:\n \"\"\"\n Args:\n x: Tensor, shape [batch_size, seq_len, embedding_dim]\n \"\"\"\n x = x.permute(1, 0, 2)\n x = x + self.pe[:x.size(0)]\n return self.dropout(x.permute(1, 0, 2))\n\nclass AttentionSSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=True, num_classes=0)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.in_features = self.encoder.num_features\n self.rpe = PositionalEncoding(self.in_features, dropout=0., max_len=64)\n self.transformer0 = nn.TransformerEncoderLayer(d_model=self.in_features, nhead=8, activation='gelu',dropout=0.1, batch_first=True)\n self.transformer1 = nn.TransformerEncoderLayer(d_model=self.in_features, nhead=8, activation='gelu',dropout=0.1, batch_first=True)\n\n self.ll1 = nn.Linear(512, 64)\n self.ll2 = nn.Linear(512, 64)\n self.ll3 = nn.Linear(512, 64)\n self.ll4 = nn.Linear(512, 64)\n self.ll5 = nn.Linear(512, 64)\n self.rl1 = nn.Linear(512, 64)\n self.rl2 = nn.Linear(512, 64)\n self.rl3 = nn.Linear(512, 64)\n self.rl4 = nn.Linear(512, 64)\n self.rl5 = nn.Linear(512, 64)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\nclass ConvNextNFNDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.ll1 = nn.Linear(512, 32)\n self.ll2 = nn.Linear(512, 32)\n self.ll3 = nn.Linear(512, 32)\n self.ll4 = nn.Linear(512, 32)\n self.ll5 = nn.Linear(512, 32)\n self.rl1 = nn.Linear(512, 32)\n self.rl2 = nn.Linear(512, 32)\n self.rl3 = nn.Linear(512, 32)\n self.rl4 = nn.Linear(512, 32)\n self.rl5 = nn.Linear(512, 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\nclass ConvNextSCSDepthDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=32, depths=[3,3,9,3], widths=[64, 128, 256, 512])\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(512))\n self.l1 = nn.Linear(512, 32)\n self.l2 = nn.Linear(512, 32)\n self.l3 = nn.Linear(512, 32)\n self.l4 = nn.Linear(512, 32)\n self.l5 = nn.Linear(512, 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n l1 = self.l1(x)\n l2 = self.l2(x)\n l3 = self.l3(x)\n l4 = self.l4(x)\n l5 = self.l5(x)\n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}\n \nclass ConvNextSCSDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(widths[-1]),\n )\n self.l1 = nn.Linear(widths[-1], 32)\n self.l2 = nn.Linear(widths[-1], 32)\n self.l3 = nn.Linear(widths[-1], 32)\n self.l4 = nn.Linear(widths[-1], 32)\n self.l5 = nn.Linear(widths[-1], 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n l1 = self.l1(x)\n l2 = self.l2(x)\n l3 = self.l3(x)\n l4 = self.l4(x)\n l5 = self.l5(x)\n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}\n\nclass ConvNextNFNDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(widths[-1]))\n self.ll1 = nn.Linear(widths[-1], 32)\n self.ll2 = nn.Linear(widths[-1], 32)\n self.ll3 = nn.Linear(widths[-1], 32)\n self.ll4 = nn.Linear(widths[-1], 32)\n self.ll5 = nn.Linear(widths[-1], 32)\n self.rl1 = nn.Linear(widths[-1], 32)\n self.rl2 = nn.Linear(widths[-1], 32)\n self.rl3 = nn.Linear(widths[-1], 32)\n self.rl4 = nn.Linear(widths[-1], 32)\n self.rl5 = nn.Linear(widths[-1], 32)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x)\n ll2 = self.ll2(x)\n ll3 = self.ll3(x)\n ll4 = self.ll4(x)\n ll5 = self.ll5(x)\n rl1 = self.rl1(x)\n rl2 = self.rl2(x)\n rl3 = self.rl3(x)\n rl4 = self.rl4(x)\n rl5 = self.rl5(x)\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n \nclass RegConvNextNFNDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(1024))\n self.ll1 = nn.Linear(1024, 3)\n self.ll2 = nn.Linear(1024, 3)\n self.ll3 = nn.Linear(1024, 3)\n self.ll4 = nn.Linear(1024, 3)\n self.ll5 = nn.Linear(1024, 3)\n self.rl1 = nn.Linear(1024, 3)\n self.rl2 = nn.Linear(1024, 3)\n self.rl3 = nn.Linear(1024, 3)\n self.rl4 = nn.Linear(1024, 3)\n self.rl5 = nn.Linear(1024, 3)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n ll1 = self.ll1(x).sigmoid()\n ll2 = self.ll2(x).sigmoid()\n ll3 = self.ll3(x).sigmoid()\n ll4 = self.ll4(x).sigmoid()\n ll5 = self.ll5(x).sigmoid()\n rl1 = self.rl1(x).sigmoid()\n rl2 = self.rl2(x).sigmoid()\n rl3 = self.rl3(x).sigmoid()\n rl4 = self.rl4(x).sigmoid()\n rl5 = self.rl5(x).sigmoid()\n return {'left_L1/L2': ll1,'left_L2/L3': ll2,'left_L3/L4': ll3, 'left_L4/L5': ll4, 'left_L5/S1': ll5,\n 'right_L1/L2': rl1, 'right_L2/L3': rl2, 'right_L3/L4': rl3, 'right_L4/L5': rl4, 'right_L5/S1': rl5}\n\nclass RegConvNextSCSDepthDetect(nn.Module):\n def __init__(self, widths):\n super().__init__()\n self.encoder = ConvNextEncoder(in_channels=1, stem_features=widths[0]//2, depths=[3,3,9,3], widths=widths)\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool3d((1,1,1)),\n nn.Flatten(1),\n nn.LayerNorm(1024),\n )\n self.l1 = nn.Linear(1024, 3)\n self.l2 = nn.Linear(1024, 3)\n self.l3 = nn.Linear(1024, 3)\n self.l4 = nn.Linear(1024, 3)\n self.l5 = nn.Linear(1024, 3)\n def forward(self, x, label=None):\n x = x.unsqueeze(1)\n x = self.encoder(x)\n x = self.flatten(x)\n l1 = self.l1(x).sigmoid()\n l2 = self.l2(x).sigmoid()\n l3 = self.l3(x).sigmoid()\n l4 = self.l4(x).sigmoid()\n l5 = self.l5(x).sigmoid()\n return {'L1/L2': l1, 'L2/L3': l2, 'L3/L4': l3, 'L4/L5': l4, 'L5/S1': l5}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.080078Z","iopub.execute_input":"2025-06-07T01:05:14.080308Z","iopub.status.idle":"2025-06-07T01:05:14.124485Z","shell.execute_reply.started":"2025-06-07T01:05:14.080285Z","shell.execute_reply":"2025-06-07T01:05:14.123894Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Lightning Module","metadata":{}},{"cell_type":"code","source":"# Là module wrapper cho các model khác, sử dụng pytorch_lightning \nclass DepthDetectModule(pl.LightningModule):\n def __init__(self, condition, widths=None, model_type='regression'):\n super().__init__()\n self.config = config\n if condition == 'scs':\n if model_type == 'regression': \n self.model = RegConvNextSCSDepthDetect(widths)\n else: \n self.model = ConvNextSCSDepthDetect(widths)\n elif condition == 'nfn':\n if model_type == 'regression': \n self.model = RegConvNextNFNDepthDetect(widths)\n else: \n self.model = ConvNextNFNDepthDetect(widths)\n def forward(self, batch):\n preds = self.model(batch)\n return preds","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.125189Z","iopub.execute_input":"2025-06-07T01:05:14.125423Z","iopub.status.idle":"2025-06-07T01:05:14.138250Z","shell.execute_reply.started":"2025-06-07T01:05:14.125396Z","shell.execute_reply":"2025-06-07T01:05:14.137702Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Inference","metadata":{}},{"cell_type":"markdown","source":"### instance prediction\n* dự đoán độ sâu (depth) cho từng instance / study\n* Đây là bước inference (suy diễn) từ các checkpoint .ckpt đã được huấn luyện trước đó","metadata":{}},{"cell_type":"code","source":"%%time\nprefix = ''\nimport warnings\nwarnings.filterwarnings(\"ignore\")\ndepth_predict = {'scs': {\n 'L1/L2':[], \n 'L2/L3': [], \n 'L3/L4': [], \n 'L4/L5': [], \n 'L5/S1': []\n }, \n 'nfn': {\n 'left_L1/L2': [], \n 'left_L2/L3': [], \n 'left_L3/L4': [], \n 'left_L4/L5': [], \n 'left_L5/S1': [], \n 'right_L1/L2': [], \n 'right_L2/L3': [], \n 'right_L3/L4': [], \n 'right_L4/L5': [], \n 'right_L5/S1': [], \n }\n }\nmodel_path_dict = {\n 'scs': [\n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_l1_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_l1_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_l1_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_l1_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_depth_1024_ssr_l1_4.ckpt', \n ], \n 'nfn': [\n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_l1_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_l1_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_l1_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_l1_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_depth_1024_ssr_l1_4.ckpt', \n ]\n}\n##############DEPTH DETECT#########################\nfor condition in ['nfn', 'scs']:\n print(condition)\n model_path_list = model_path_dict[condition]\n for model_path in model_path_list:\n _meta_df = meta_df.copy()\n #_series = series.copy()\n dataset_test = DepthDetectDataset(_meta_df, condition, 'sub')\n data_loader_test = DataLoader(\n dataset_test,\n batch_size=config[\"test_bs\"], \n shuffle=False,\n num_workers=4,\n pin_memory=False\n )\n model_name = model_path.split('/')[-1]\n if '1024' in model_name: \n widths = [128, 256, 512, 1024]\n else: \n widths = [64, 128, 256, 512]\n if 'l1' in model_name: \n model_type = 'regression'\n else: \n model_type = 'classification'\n model = DepthDetectModule.load_from_checkpoint(model_path, condition=condition, widths=widths, model_type=model_type)\n model.eval()\n model.zero_grad()\n model.to(device)\n\n pred_temp = {}\n for k in depth_predict[condition].keys(): \n pred_temp[k] = []\n study_id_list = []\n with torch.no_grad():\n for data in tqdm(data_loader_test, total=len(data_loader_test)):\n images, study_id = data\n images = images.to(device)\n preds = model.forward(images)\n #print(preds)\n if model_type == 'regression': \n for k, v in preds.items(): \n pred_temp[k].append((v[:, -1]*32).to('cpu').detach().numpy())\n else: \n for k, v in preds.items(): \n pred_temp[k].append(torch.argmax(v, dim=1).to('cpu').detach().numpy())\n study_id_list.append(study_id.to('cpu').reshape(-1).detach().numpy())\n del images, study_id, preds\n gc.collect()\n for k, v in pred_temp.items(): \n depth_predict[condition][k].append(np.concatenate(v))\n study_id = np.concatenate(study_id_list)\n del pred_temp, study_id_list\n gc.collect()\n \n for k, v in depth_predict[condition].items(): \n depth_predict[condition][k] = np.median(np.array(depth_predict[condition][k]), axis=0)\n depth_predict[condition]['study_id'] = study_id\n del study_id\n gc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:05:14.138876Z","iopub.execute_input":"2025-06-07T01:05:14.139088Z","iopub.status.idle":"2025-06-07T01:10:04.647652Z","shell.execute_reply.started":"2025-06-07T01:05:14.139063Z","shell.execute_reply":"2025-06-07T01:10:04.647004Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Create label coordinate % align","metadata":{}},{"cell_type":"code","source":"def create_label_ins(study_id, depth, level, condition, desc): \n coor_dict = {'study_id': [], 'series_id': [], 'instance_number': []}\n _meta = meta_df.loc[meta_df.series_description==desc]\n for s, d in zip(study_id, depth): \n sub_meta = _meta.loc[_meta.study_id==s]\n sub_meta = sub_meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n if len(sub_meta) > 32: \n d = (d/32)*len(sub_meta)\n try: \n row = sub_meta.iloc[round(d)]\n except: \n if condition == 'Spinal Canal Stenosis': \n row = sub_meta.iloc[int(len(sub_meta)//2)]\n elif condition == 'Left Neural Foraminal Narrowing': \n row = sub_meta.iloc[int(2*(len(sub_meta)//3))]\n elif condition == 'Right Neural Foraminal Narrowing': \n row = sub_meta.iloc[int(len(sub_meta)//3)]\n print(s)\n coor_dict['study_id'].append(s)\n coor_dict['series_id'].append(row.series_id)\n coor_dict['instance_number'].append(row.instance_number)\n coor_dict['condition'] = condition\n coor_dict['level'] = level.split('_')[-1]\n return pd.DataFrame(coor_dict)\n\nscs_study_id = depth_predict['scs']['study_id']\nscs_coor_list = []\nfor k, v in depth_predict['scs'].items(): \n if k != 'study_id': \n scs_coor_list.append(create_label_ins(scs_study_id, v, k, 'Spinal Canal Stenosis', 'Sagittal T2/STIR'))\n\nnfn_study_id = depth_predict['nfn']['study_id']\nnfn_coor_list = []\nfor k, v in depth_predict['nfn'].items(): \n if k != 'study_id': \n if k.split('_')[0] == 'left': \n condition = 'Left Neural Foraminal Narrowing'\n else: \n condition = 'Right Neural Foraminal Narrowing'\n nfn_coor_list.append(create_label_ins(nfn_study_id, v, k, condition, 'Sagittal T1'))\nscs_coor = pd.concat(scs_coor_list)\nnfn_coor = pd.concat(nfn_coor_list)\npred_coor = pd.concat([scs_coor, nfn_coor]).sort_values(['study_id', 'series_id', 'level'])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:04.648408Z","iopub.execute_input":"2025-06-07T01:10:04.648627Z","iopub.status.idle":"2025-06-07T01:10:04.682004Z","shell.execute_reply.started":"2025-06-07T01:10:04.648608Z","shell.execute_reply":"2025-06-07T01:10:04.681331Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"del scs_coor, nfn_coor, scs_coor_list, nfn_coor_list, depth_predict\ngc.collect()\npred_coor.head()\npred_coor.to_csv('stage1_coor.csv', index=False)\nos.listdir('/kaggle/working/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:04.682906Z","iopub.execute_input":"2025-06-07T01:10:04.683156Z","iopub.status.idle":"2025-06-07T01:10:04.952769Z","shell.execute_reply.started":"2025-06-07T01:10:04.683132Z","shell.execute_reply":"2025-06-07T01:10:04.951931Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Second Stage (xy inference)\n* infer xy-coordinate of locations of sagittal t1 & t2\n* ensemble or align (rule base)","metadata":{}},{"cell_type":"markdown","source":"## Coordinate prediction dataset","metadata":{}},{"cell_type":"code","source":"class CoorDetectDataset(Dataset):\n def __init__(self, coor, meta, condition, usage='train'):\n if condition == 'scs':\n coor = coor.loc[coor.condition=='Spinal Canal Stenosis']\n elif condition == 'ss':\n coor = coor.loc[(coor.condition=='Left Subarticular Stenosis') | (coor.condition=='Right Subarticular Stenosis')]\n elif condition == 'nfn':\n coor = coor.loc[(coor.condition=='Right Neural Foraminal Narrowing') | (coor.condition=='Left Neural Foraminal Narrowing')]\n #g_coor = coor.groupby('study_id').count()\n #if condition == 'scs':\n # self.id = g_coor.loc[g_coor.series_id==5].reset_index().study_id.unique()\n #else:\n # self.id = g_coor.loc[g_coor.series_id==10].reset_index().study_id.unique()\n self.id = coor.study_id.unique()\n self.coor = coor\n self.meta = meta\n self.condition = condition\n self.usage = usage\n if 3637444890 in self.id: \n self.id.remove(3637444890)\n #self.id = [2773343225]\n #self.id = [1782095928]\n\n self.resize = v2.Resize((384, 384))\n \n def __getitem__(self, index):\n study_id = self.id[index]\n #print(study_id)\n #try:\n if self.condition == 'scs':\n volume = self.for_scs(study_id)\n elif self.condition == 'nfn':\n volume = self.for_nfn(study_id)\n if self.condition == 'ss':\n volume = self.for_ss(study_id)\n return volume, torch.tensor(study_id)\n\n def for_scs(self, study_id):\n meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T2/STIR')]\n meta = meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n #img = [self.normalize(self.load_dicom(f'/content/train_images/{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in meta.iterrows()]\n coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Spinal Canal Stenosis')]\n meta_list = []\n for _, row in coor.iterrows():\n series_id, instance_number = row.series_id, row.instance_number\n meta_list.append(meta.loc[(meta.series_id==series_id) & (meta.instance_number==instance_number)])\n sub_meta = pd.concat(meta_list)\n idx = meta.loc[meta.ipp_x == sub_meta.ipp_x.median()].index[0]\n #print(old_idx)\n img_row = meta.iloc[idx]\n before_img_row = meta.iloc[idx-1]\n after_img_row = meta.iloc[idx+1]\n img = self.normalize(self.load_dicom(IMAGE_PATH + f'{img_row.study_id}/{img_row.series_id}/{img_row.instance_number}.dcm'))\n bimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{before_img_row.study_id}/{before_img_row.series_id}/{before_img_row.instance_number}.dcm'))\n aimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{after_img_row.study_id}/{after_img_row.series_id}/{after_img_row.instance_number}.dcm'))\n img = self.resize(torch.tensor(img[None, ...]))\n bimg = self.resize(torch.tensor(bimg[None, ...]))\n aimg = self.resize(torch.tensor(aimg[None, ...]))\n img = torch.cat([bimg, img, aimg]).to(torch.float32)\n return img\n def for_ss(self, study_id):\n meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Axial T2')]\n meta = meta.sort_values('ipp_z', ascending=False).reset_index(drop=True)\n img = [self.normalize(self.load_dicom(f'/content/train_images/{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in meta.iterrows()]\n coor = self.coor.loc[(self.coor.study_id==study_id)]\n coor_dict = {}\n for _, row in coor.iterrows():\n series_id, instance_number = row.series_id, row.instance_number\n target_row = meta.loc[(meta.series_id==series_id) & (meta.instance_number==instance_number)]\n idx = target_row.index[0]\n #print(row.level, idx, idx/len(img))\n #plt.title(row.level)\n #plt.imshow(img[idx])\n #mask = torch.zeros(img[idx].shape)\n #mask[int(row.y)-10:int((row.y))+10, int(row.x)-10:int((row.x))+10] = 1\n #plt.imshow(mask, alpha=0.5)\n #plt.show()\n height, width = img[idx].shape\n z = idx/depth if len(img) < depth else idx/len(img)\n x = row.x/width\n y = row.y/height\n if row.condition == 'Right Subarticular Stenosis':\n coor_dict['right_' + row.level] = torch.tensor([x, y, z]).to(torch.float32)\n else:\n coor_dict['left_' + row.level] = torch.tensor([x, y, z]).to(torch.float32)\n volume = torch.cat([self.resize(torch.tensor(i)[None, ...]).to(torch.float32) for i in img]).contiguous()\n if volume.shape[0] < depth:\n volume = torch.cat([volume, torch.zeros(depth-volume.shape[0], volume.shape[1], volume.shape[2])])\n elif volume.shape[0] > depth:\n volume = torch.nn.functional.interpolate(volume[None, None, ...], (depth, volume.shape[1], volume.shape[2])).squeeze()\n return volume, coor_dict\n\n def for_nfn(self, study_id):\n meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T1')]\n meta = meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n #img = [self.normalize(self.load_dicom(f'/content/train_images/{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in meta.iterrows()]\n coor = self.coor.loc[(self.coor.study_id==study_id)]\n right_meta_list = []\n left_meta_list = []\n for _, row in coor.iterrows():\n series_id, instance_number = row.series_id, row.instance_number\n if row.condition == 'Right Neural Foraminal Narrowing':\n right_meta_list.append(meta.loc[(meta.series_id==series_id) & (meta.instance_number==instance_number)])\n else: \n left_meta_list.append(meta.loc[(meta.series_id==series_id) & (meta.instance_number==instance_number)])\n\n right_sub_meta = pd.concat(right_meta_list)\n left_sub_meta = pd.concat(left_meta_list)\n ridx = meta.loc[meta.ipp_x == right_sub_meta.ipp_x.median()].index[0]\n lidx = meta.loc[meta.ipp_x == left_sub_meta.ipp_x.median()].index[0]\n right_img_row = meta.iloc[min(max(ridx, 0), len(meta)-1)]\n #display(right_img_row)\n right_before_img_row = meta.iloc[min(max(ridx-1, 0), len(meta)-1)]\n rightafter_img_row = meta.iloc[min(max(ridx+1, 0), len(meta)-1)]\n left_img_row = meta.iloc[min(max(lidx, 0), len(meta)-1)]\n left_before_img_row = meta.iloc[min(max(lidx-1, 0), len(meta)-1)]\n leftafter_img_row = meta.iloc[min(max(lidx+1, 0), len(meta)-1)]\n rimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{right_img_row.study_id}/{right_img_row.series_id}/{right_img_row.instance_number}.dcm'))\n rbimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{right_before_img_row.study_id}/{right_before_img_row.series_id}/{right_before_img_row.instance_number}.dcm'))\n raimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{rightafter_img_row.study_id}/{rightafter_img_row.series_id}/{rightafter_img_row.instance_number}.dcm'))\n limg = self.normalize(self.load_dicom(IMAGE_PATH + f'{left_img_row.study_id}/{left_img_row.series_id}/{left_img_row.instance_number}.dcm'))\n lbimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{left_before_img_row.study_id}/{left_before_img_row.series_id}/{left_before_img_row.instance_number}.dcm'))\n laimg = self.normalize(self.load_dicom(IMAGE_PATH + f'{leftafter_img_row.study_id}/{leftafter_img_row.series_id}/{leftafter_img_row.instance_number}.dcm'))\n \n rimg = torch.cat([self.resize(torch.tensor(i)[None, ...]).to(torch.float32) for i in [rbimg, rimg, raimg]])\n limg = torch.cat([self.resize(torch.tensor(i)[None, ...]).to(torch.float32) for i in [lbimg, limg, laimg]])\n img = torch.stack([limg, rimg]).to(torch.float32).contiguous()\n return img\n\n def normalize(self, x):\n lower, upper = np.percentile(x, (1, 99))\n x = np.clip(x, lower, upper)\n x = x - np.min(x)\n x = x / np.max(x)\n return x\n\n def __len__(self):\n return len(self.id)\n\n def load_dicom(self, path):\n dicom = dcm.dcmread(path)\n data = dicom.pixel_array\n return data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:04.953654Z","iopub.execute_input":"2025-06-07T01:10:04.953900Z","iopub.status.idle":"2025-06-07T01:10:04.980154Z","shell.execute_reply.started":"2025-06-07T01:10:04.953883Z","shell.execute_reply":"2025-06-07T01:10:04.979422Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Coordinate prediction models","metadata":{}},{"cell_type":"code","source":"class ConvNextSCSDetect(nn.Module):\n def __init__(self, encoder):\n super().__init__()\n #self.size = 384\n if encoder == 'convnext': \n self.encoder = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n elif encoder == 'efficientnetv2-l': \n self.encoder = timm.create_model('tf_efficientnetv2_l.in21k_ft_in1k', in_chans=3, pretrained=False, num_classes=0, drop_rate=0.)\n self.in_features = self.encoder.num_features\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1),\n #nn.LayerNorm(self.in_features)\n )\n self.l1 = nn.Linear(self.in_features, 2)\n self.l2 = nn.Linear(self.in_features, 2)\n self.l3 = nn.Linear(self.in_features, 2)\n self.l4 = nn.Linear(self.in_features, 2)\n self.l5 = nn.Linear(self.in_features, 2)\n def forward(self, x, label=None):\n #for loc, img in x.items():\n #print(img.shape)\n # img = self.encoder.forward_features(img)\n # img = self.flatten(img)\n # x[loc] = img\n x = self.encoder.forward_features(x)\n x = self.flatten(x)\n l1 = self.l1(x)\n l2 = self.l2(x)\n l3 = self.l3(x)\n l4 = self.l4(x)\n l5 = self.l5(x)\n return {'L1/L2': l1.sigmoid(), 'L2/L3': l2.sigmoid(), 'L3/L4': l3.sigmoid(), 'L4/L5': l4.sigmoid(), 'L5/S1': l5.sigmoid()}\n\nclass ConvNextNFNDetect(nn.Module):\n def __init__(self, encoder):\n super().__init__()\n if encoder == 'convnext': \n self.encoder = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n elif encoder == 'efficientnetv2-l': \n self.encoder = timm.create_model('tf_efficientnetv2_l.in21k_ft_in1k', in_chans=3, pretrained=False, num_classes=0, drop_rate=0.)\n self.in_features = self.encoder.num_features\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1),\n #nn.LayerNorm(self.in_features)\n )\n self.ll1 = nn.Linear(self.in_features, 2)\n self.ll2 = nn.Linear(self.in_features, 2)\n self.ll3 = nn.Linear(self.in_features, 2)\n self.ll4 = nn.Linear(self.in_features, 2)\n self.ll5 = nn.Linear(self.in_features, 2)\n self.rl1 = nn.Linear(self.in_features, 2)\n self.rl2 = nn.Linear(self.in_features, 2)\n self.rl3 = nn.Linear(self.in_features, 2)\n self.rl4 = nn.Linear(self.in_features, 2)\n self.rl5 = nn.Linear(self.in_features, 2)\n def forward(self, x, label=None):\n shape = x.shape\n x = x.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n x = self.encoder.forward_features(x)\n x = self.flatten(x)\n x = x.reshape(shape[0], shape[1], -1)\n x_left = x[:, 0, :]\n x_right = x[:, 1, :]\n ll1 = self.ll1(x_left)\n ll2 = self.ll2(x_left)\n ll3 = self.ll3(x_left)\n ll4 = self.ll4(x_left)\n ll5 = self.ll5(x_left)\n rl1 = self.rl1(x_right)\n rl2 = self.rl2(x_right)\n rl3 = self.rl3(x_right)\n rl4 = self.rl4(x_right)\n rl5 = self.rl5(x_right)\n return {'left_L1/L2': ll1.sigmoid(),'left_L2/L3': ll2.sigmoid(),'left_L3/L4': ll3.sigmoid(), 'left_L4/L5': ll4.sigmoid(), 'left_L5/S1': ll5.sigmoid(),\n 'right_L1/L2': rl1.sigmoid(), 'right_L2/L3': rl2.sigmoid(), 'right_L3/L4': rl3.sigmoid(), 'right_L4/L5': rl4.sigmoid(), 'right_L5/S1': rl5.sigmoid()}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:04.980909Z","iopub.execute_input":"2025-06-07T01:10:04.981099Z","iopub.status.idle":"2025-06-07T01:10:05.001708Z","shell.execute_reply.started":"2025-06-07T01:10:04.981084Z","shell.execute_reply":"2025-06-07T01:10:05.000998Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Coordinate detection lightning module","metadata":{}},{"cell_type":"code","source":"class DetectModule(pl.LightningModule):\n def __init__(self, condition, encoder):\n super().__init__()\n self.config = condition\n if condition == 'scs':\n self.model = ConvNextSCSDetect(encoder)\n elif condition == 'nfn':\n self.model = ConvNextNFNDetect(encoder)\n elif condition == 'ss': \n pass\n #self.ema = ExponentialMovingAverage(self.model.parameters(), decay=0.995)\n #self.ema.to(device)\n\n #self.model = torch.optim.swa_utils.AveragedModel(self.model,\n # multi_avg_fn=torch.optim.swa_utils.get_ema_multi_avg_fn(0.999))\n\n def forward(self, batch):\n preds = self.model(batch)\n return preds","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:05.005473Z","iopub.execute_input":"2025-06-07T01:10:05.005738Z","iopub.status.idle":"2025-06-07T01:10:05.026145Z","shell.execute_reply.started":"2025-06-07T01:10:05.005722Z","shell.execute_reply":"2025-06-07T01:10:05.025452Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Coordinate inference","metadata":{}},{"cell_type":"code","source":"%%time\nprefix = ''\nimport warnings\nwarnings.filterwarnings(\"ignore\")\ncoor_predict = {'scs': {\n 'L1/L2':[], \n 'L2/L3': [], \n 'L3/L4': [], \n 'L4/L5': [], \n 'L5/S1': []\n }, \n 'nfn': {\n 'left_L1/L2': [], \n 'left_L2/L3': [], \n 'left_L3/L4': [], \n 'left_L4/L5': [], \n 'left_L5/S1': [], \n 'right_L1/L2': [], \n 'right_L2/L3': [], \n 'right_L3/L4': [], \n 'right_L4/L5': [], \n 'right_L5/S1': [], \n }\n }\nmodel_path_dict = {\n 'scs': [\n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_effv2l_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_effv2l_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_effv2l_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_effv2l_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/scs_detect_pre_effv2l_4.ckpt', \n ], \n 'nfn': [\n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_4.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_effv2l_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_effv2l_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_effv2l_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_effv2l_3.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_detect_pre_effv2l_4.ckpt', \n ]\n}\n##############COOR DETECT#########################\nfor condition in ['nfn', 'scs']:\n print(condition)\n model_path_list = model_path_dict[condition]\n for path in model_path_list:\n if 'effv2l' in path.split('/')[-1]: \n encoder = 'efficientnetv2-l'\n else: \n encoder = 'convnext'\n _meta_df = meta_df.copy()\n _coor = pred_coor.copy()\n dataset_test = CoorDetectDataset(_coor, _meta_df, condition, 'sub')\n data_loader_test = DataLoader(\n dataset_test,\n batch_size=config[\"test_bs\"],\n shuffle=False,\n num_workers=4,\n pin_memory=False\n )\n print(path, encoder)\n model = DetectModule.load_from_checkpoint(path, condition=condition, encoder=encoder)\n model.eval()\n model.zero_grad()\n model.to(device)\n\n pred_temp = {}\n for k in coor_predict[condition].keys(): \n pred_temp[k] = []\n study_id_list = []\n with torch.no_grad():\n for data in tqdm(data_loader_test, total=len(data_loader_test)):\n images, study_id = data\n images = images.to(device)\n preds = model.forward(images)\n #print(preds)\n for k, v in preds.items(): \n pred_temp[k].append(v.to('cpu').detach().numpy())\n study_id_list.append(study_id.to('cpu').reshape(-1).detach().numpy())\n for k, v in pred_temp.items(): \n coor_predict[condition][k].append(np.concatenate(v))\n del pred_temp\n gc.collect()\n study_id = np.concatenate(study_id_list)\n for k, v in coor_predict[condition].items(): \n coor_predict[condition][k] = np.mean(np.array(coor_predict[condition][k]), axis=0)\n coor_predict[condition]['study_id'] = study_id\n del study_id\n gc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:10:05.026867Z","iopub.execute_input":"2025-06-07T01:10:05.027060Z","iopub.status.idle":"2025-06-07T01:13:47.245271Z","shell.execute_reply.started":"2025-06-07T01:10:05.027038Z","shell.execute_reply":"2025-06-07T01:13:47.244458Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"pred_coor.head(1)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.246439Z","iopub.execute_input":"2025-06-07T01:13:47.246727Z","iopub.status.idle":"2025-06-07T01:13:47.254944Z","shell.execute_reply.started":"2025-06-07T01:13:47.246705Z","shell.execute_reply":"2025-06-07T01:13:47.254352Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%time\ndef create_label_coor(study_id, coor_df, coor, level, condition, desc): \n _meta = meta_df.loc[meta_df.series_description==desc].copy()\n _coor = coor_df.loc[coor_df.condition == condition]\n _coor_df = {'study_id': [], 'series_id': [], 'x': [], 'y': []}\n for s, c in zip(study_id, coor): \n sub_meta = _meta.loc[_meta.study_id == s]\n sub_coor = _coor.loc[(_coor.study_id==s) & (_coor.level==level.split('_')[-1])].squeeze(axis=0)\n #display(sub_coor)\n meta_row = sub_meta.loc[(sub_meta.instance_number==sub_coor.instance_number) & (sub_meta.series_id==sub_coor.series_id)].squeeze(axis=0)\n x = round(meta_row.width * c[0])\n y = round(meta_row.height * c[1])\n _coor_df['study_id'].append(s)\n _coor_df['series_id'].append(sub_coor.series_id)\n _coor_df['x'].append(x)\n _coor_df['y'].append(y)\n _coor_df['level'] = level.split('_')[-1]\n _coor_df['condition'] = condition\n del _meta, _coor, sub_meta, sub_coor, meta_row\n return pd.DataFrame(_coor_df)\n\nscs_study_id = coor_predict['scs']['study_id']\nscs_coor_list = []\nfor k, v in coor_predict['scs'].items(): \n if k != 'study_id': \n scs_coor_list.append(create_label_coor(scs_study_id, pred_coor, v, k, 'Spinal Canal Stenosis', 'Sagittal T2/STIR'))\n\nnfn_study_id = coor_predict['nfn']['study_id']\nnfn_coor_list = []\nfor k, v in coor_predict['nfn'].items(): \n if k != 'study_id': \n if k.split('_')[0] == 'left': \n condition = 'Left Neural Foraminal Narrowing'\n else: \n condition = 'Right Neural Foraminal Narrowing'\n nfn_coor_list.append(create_label_coor(nfn_study_id, pred_coor, v, k, condition, 'Sagittal T1'))\nscs_coor = pd.concat(scs_coor_list)\nnfn_coor = pd.concat(nfn_coor_list)\n_pred_coor = pd.concat([scs_coor, nfn_coor]).sort_values(['study_id', 'series_id', 'level'])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.255699Z","iopub.execute_input":"2025-06-07T01:13:47.255951Z","iopub.status.idle":"2025-06-07T01:13:47.305306Z","shell.execute_reply.started":"2025-06-07T01:13:47.255926Z","shell.execute_reply":"2025-06-07T01:13:47.304671Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"pred_coor_stage2 = pd.merge(pred_coor, _pred_coor, on=['study_id', 'series_id', 'level', 'condition'], how='inner')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.306032Z","iopub.execute_input":"2025-06-07T01:13:47.306280Z","iopub.status.idle":"2025-06-07T01:13:47.316312Z","shell.execute_reply.started":"2025-06-07T01:13:47.306260Z","shell.execute_reply":"2025-06-07T01:13:47.315668Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"display(pred_coor_stage2.head())\npred_coor_stage2.to_csv('stage2_coor.csv', index=False)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.316985Z","iopub.execute_input":"2025-06-07T01:13:47.317158Z","iopub.status.idle":"2025-06-07T01:13:47.325884Z","shell.execute_reply.started":"2025-06-07T01:13:47.317144Z","shell.execute_reply":"2025-06-07T01:13:47.325390Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"os.listdir('/kaggle/working/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.326571Z","iopub.execute_input":"2025-06-07T01:13:47.326866Z","iopub.status.idle":"2025-06-07T01:13:47.338270Z","shell.execute_reply.started":"2025-06-07T01:13:47.326840Z","shell.execute_reply":"2025-06-07T01:13:47.337728Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Third Stage (calc. location of axial t2)¶\n* calcurate depth of axial t2 for each location roughly, using xyz-coordinate (refered to @hengck's transformation from sagittal t2 to axial t2)\n* roughly separate each locations\n* infer instance number\n* infer xy-coordinate","metadata":{}},{"cell_type":"markdown","source":"### Calculate axial slice","metadata":{}},{"cell_type":"code","source":"def project_to_3d(row):\n sx, sy, sz = row.ipp_x, row.ipp_y, row.ipp_z\n x, y = row.x, row.y\n o0, o1, o2, o3, o4, o5 = row.iop\n delx, dely = row.ps_x, row.ps_y\n xx = o0 * delx * x + o3 * dely * y + sx\n yy = o1 * delx * x + o4 * dely * y + sy\n zz = o2 * delx * x + o5 * dely * y + sz\n return xx,yy,zz\n\ndef sag_to_ax(sub_coor, sub_meta): \n point = sub_coor[['ipp_x', 'ipp_y', 'ipp_z']].values #2d\n level_list = sub_coor.level.tolist()\n # here we project 2d to 3d\n center=[] \n for _, row in sub_coor.iterrows():\n xx,yy,zz = project_to_3d(row)\n center.append([xx,yy,zz])\n center = np.array(center) #3d\n\n # == 2. we get closest axial slices to the CSC points =================\n #df = valid_data[0].axial_t2[0].df\n\n orientation = np.array(sub_meta.iop.values.tolist())\n position= np.array(sub_meta[['ipp_x', 'ipp_y', 'ipp_z']].values.tolist())\n ox = orientation[:, :3]\n oy = orientation[:, 3:]\n oz = np.cross(ox,oy)\n t = center.reshape(-1,1,3) - position.reshape(1,-1,3)\n dis = (oz.reshape(1,-1,3) * t).sum(-1) # np.dot(point-s,oz)\n dis = np.fabs(dis)\n closest = dis.argmin(-1)\n closest_df = sub_meta.iloc[closest]\n closest_df['level'] = level_list#['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']\n closest_df['x'] = 0\n closest_df['y'] = 0\n #closest_df = pd.concat([closest_df, closest_df])\n #closest_df['condition'] = ['Left Subarticular Stenosis']*5 + ['Right Subarticular Stenosis']*5\n return closest_df[['study_id', 'series_id', 'instance_number', 'level']]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.338976Z","iopub.execute_input":"2025-06-07T01:13:47.339214Z","iopub.status.idle":"2025-06-07T01:13:47.350003Z","shell.execute_reply.started":"2025-06-07T01:13:47.339193Z","shell.execute_reply":"2025-06-07T01:13:47.349434Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# sagittal t2 => axial t2\nscs_coor = pred_coor_stage2.loc[pred_coor_stage2.condition=='Spinal Canal Stenosis'].copy()\nscs_coor = scs_coor.merge(meta_df, on=['study_id', 'series_id', 'instance_number'], how='left')\nstudy_id = scs_coor.study_id.unique()\nax_meta = meta_df.loc[(meta_df.series_description=='Axial T2')]\nclosest_ax_list = []\nfor s in tqdm(study_id, total=len(study_id)): \n sub_coor = scs_coor.loc[scs_coor.study_id==s]\n sub_meta = ax_meta.loc[ax_meta.study_id==s]\n closest_ax_list.append(sag_to_ax(sub_coor, sub_meta)) ","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.350781Z","iopub.execute_input":"2025-06-07T01:13:47.350988Z","iopub.status.idle":"2025-06-07T01:13:47.384257Z","shell.execute_reply.started":"2025-06-07T01:13:47.350973Z","shell.execute_reply":"2025-06-07T01:13:47.383545Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"closest_ax = pd.concat(closest_ax_list)\nclosest_ax.head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.384992Z","iopub.execute_input":"2025-06-07T01:13:47.385182Z","iopub.status.idle":"2025-06-07T01:13:47.391940Z","shell.execute_reply.started":"2025-06-07T01:13:47.385167Z","shell.execute_reply":"2025-06-07T01:13:47.391265Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Subarticular stenosis (ss) - hẹp ống dưới coordinate prediction dataset","metadata":{}},{"cell_type":"code","source":"class SSDetectDataset(Dataset):\n def __init__(self, ax, usage='train'):\n self.ax = ax\n self.id = ax.study_id.unique()\n self.usage = usage\n self.id = list(set(self.id) - set([3637444890]))\n #self.id = [2773343225]\n #self.id = [1782095928]\n\n self.resize = v2.Resize((384, 384))\n \n def __getitem__(self, index):\n study_id = self.id[index]\n volume = self.for_ss(study_id)\n return volume, torch.tensor(study_id)\n\n def for_ss(self, study_id):\n ax = self.ax.loc[self.ax.study_id==study_id]\n img_dict = {}\n for _, row in ax.iterrows():\n series_id, instance_number = row.series_id, row.instance_number\n img = self.load_dicom(IMAGE_PATH + f'{study_id}/{series_id}/{instance_number}.dcm').astype(np.float32)\n img = self.resize(torch.tensor(img)[None, ...])\n img = self.normalize(img)\n img_dict[row.level] = img\n img_list = []\n for k in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']: \n img_list.append(img_dict[k])\n volume = torch.stack(img_list).contiguous()\n return volume\n\n def normalize(self, x):\n upper = torch.quantile(x, torch.tensor([0.99]))\n lower = torch.quantile(x, torch.tensor([0.01]))\n x = torch.clip(x, lower, upper)\n x = x - torch.min(x)\n x = x / (torch.max(x)+1e-6)\n return x\n\n def __len__(self):\n return len(self.id)\n\n def load_dicom(self, path):\n dicom = dcm.dcmread(path)\n data = dicom.pixel_array\n return data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.392699Z","iopub.execute_input":"2025-06-07T01:13:47.392906Z","iopub.status.idle":"2025-06-07T01:13:47.405232Z","shell.execute_reply.started":"2025-06-07T01:13:47.392882Z","shell.execute_reply":"2025-06-07T01:13:47.404448Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## subarticular stenosis coordinate dectection model","metadata":{}},{"cell_type":"code","source":"class SSDetect(nn.Module):\n def __init__(self):\n super().__init__()\n self.encoder = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n self.in_features = self.encoder.num_features\n self.flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1),\n #nn.LayerNorm(self.in_features)\n )\n self.left = nn.Linear(self.in_features, 2)\n self.right = nn.Linear(self.in_features, 2)\n def forward(self, x, label=None):\n shape = x.shape\n x = x.reshape(shape[0]*shape[1], 1, shape[-2], shape[-1])\n x = self.encoder.forward_features(x)\n x = self.flatten(x)\n x = x.reshape(shape[0], shape[1], -1)\n x_left = x\n x_right = x\n left = self.left(x_left)\n right = self.right(x_right)\n return {'left_L1/L2': left[:, 0, :].sigmoid(),'left_L2/L3': left[:, 1, :].sigmoid(),'left_L3/L4': left[:, 2, :].sigmoid(), 'left_L4/L5': left[:, 3, :].sigmoid(), 'left_L5/S1': left[:, 4, :].sigmoid(),\n 'right_L1/L2': right[:, 0, :].sigmoid(), 'right_L2/L3': right[:, 1, :].sigmoid(), 'right_L3/L4': right[:, 2, :].sigmoid(), 'right_L4/L5': right[:, 3, :].sigmoid(), 'right_L5/S1': right[:, 4, :].sigmoid()}","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.406216Z","iopub.execute_input":"2025-06-07T01:13:47.406472Z","iopub.status.idle":"2025-06-07T01:13:47.420428Z","shell.execute_reply.started":"2025-06-07T01:13:47.406446Z","shell.execute_reply":"2025-06-07T01:13:47.419727Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Lightning module","metadata":{}},{"cell_type":"code","source":"class SSDetectModule(pl.LightningModule):\n def __init__(self):\n super().__init__()\n self.model = SSDetect()\n def forward(self, batch):\n preds = self.model(batch)\n return preds","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.421311Z","iopub.execute_input":"2025-06-07T01:13:47.421627Z","iopub.status.idle":"2025-06-07T01:13:47.434280Z","shell.execute_reply.started":"2025-06-07T01:13:47.421609Z","shell.execute_reply":"2025-06-07T01:13:47.433669Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## SS Coordinate inference","metadata":{}},{"cell_type":"code","source":"%%time\nprefix = ''\nimport warnings\nwarnings.filterwarnings(\"ignore\")\ncoor_predict = {\n 'left_L1/L2': [], \n 'left_L2/L3': [], \n 'left_L3/L4': [], \n 'left_L4/L5': [], \n 'left_L5/S1': [], \n 'right_L1/L2': [], \n 'right_L2/L3': [], \n 'right_L3/L4': [], \n 'right_L4/L5': [], \n 'right_L5/S1': [], \n }\n\n##############COOR DETECT#########################\nfor i in [0, 1, 2, 3, 4]:\n _meta_df = meta_df.copy()\n _series = series.copy()\n _coor = pred_coor.copy()\n dataset_test = SSDetectDataset(closest_ax, 'sub')\n data_loader_test = DataLoader(\n dataset_test,\n batch_size=config[\"test_bs\"],\n shuffle=False,\n num_workers=4,\n pin_memory=False\n )\n\n model = SSDetectModule.load_from_checkpoint(f'/kaggle/input/rsna-spine-final-models/ss_detect_{i}.ckpt')\n model.eval()\n model.zero_grad()\n model.to(device)\n\n pred_temp = {}\n for k in coor_predict.keys(): \n pred_temp[k] = []\n study_id_list = []\n with torch.no_grad():\n for data in tqdm(data_loader_test, total=len(data_loader_test)):\n images, study_id = data\n images = images.to(device)\n preds = model.forward(images)\n #print(preds)\n for k, v in preds.items(): \n pred_temp[k].append(v.to('cpu').detach().numpy())\n study_id_list.append(study_id.to('cpu').reshape(-1).detach().numpy())\n for k, v in pred_temp.items(): \n coor_predict[k].append(np.concatenate(v))\n del pred_temp\n gc.collect()\n study_id = np.concatenate(study_id_list)\nfor k, v in coor_predict.items(): \n coor_predict[k] = np.mean(np.array(coor_predict[k]), axis=0)\ncoor_predict['study_id'] = study_id\ndel study_id\ngc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:13:47.434846Z","iopub.execute_input":"2025-06-07T01:13:47.435014Z","iopub.status.idle":"2025-06-07T01:14:39.950366Z","shell.execute_reply.started":"2025-06-07T01:13:47.435001Z","shell.execute_reply":"2025-06-07T01:14:39.949559Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"study_id = coor_predict['study_id']\ncoor_dict = {'study_id': [], 'x': [], 'y': [], 'condition': [], 'level': []}\nfor k, v in coor_predict.items(): \n if k == 'study_id': \n continue\n _lr, location = k.split('_')\n if _lr == 'left': \n lr = 'Left Subarticular Stenosis'\n else: \n lr = 'Right Subarticular Stenosis'\n coor_dict['study_id'].extend(list(coor_predict['study_id']))\n coor_dict['x'].extend(list(v[:, 0]))\n coor_dict['y'].extend(list(v[:, 1]))\n coor_dict['condition'].extend([lr]*len(v))\n coor_dict['level'].extend([location]*len(v))\nax_coor_pred = pd.DataFrame(coor_dict)\nax_coor_pred = pd.merge(closest_ax, ax_coor_pred, on=['study_id', 'level'], how='left')\nprint(ax_coor_pred.shape)\nax_coor_pred.head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:39.951364Z","iopub.execute_input":"2025-06-07T01:14:39.951665Z","iopub.status.idle":"2025-06-07T01:14:39.968402Z","shell.execute_reply.started":"2025-06-07T01:14:39.951638Z","shell.execute_reply":"2025-06-07T01:14:39.967565Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"ax_coor_pred = ax_coor_pred.merge(meta_df[['study_id', 'series_id', 'instance_number', 'height', 'width']], on=['study_id', 'series_id', 'instance_number'], how='left')\nax_coor_pred['x'] = ax_coor_pred['x']*ax_coor_pred['width']\nax_coor_pred['y'] = ax_coor_pred['y']*ax_coor_pred['height']\nax_coor_pred['x'] = ax_coor_pred['x'].apply(lambda x: round(x))\nax_coor_pred['y'] = ax_coor_pred['y'].apply(lambda x: round(x))\nax_coor_pred = ax_coor_pred.drop(['height', 'width'], axis=1)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:39.969373Z","iopub.execute_input":"2025-06-07T01:14:39.969650Z","iopub.status.idle":"2025-06-07T01:14:39.990216Z","shell.execute_reply.started":"2025-06-07T01:14:39.969626Z","shell.execute_reply":"2025-06-07T01:14:39.989517Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"pred_coor_stage3 = pd.concat([pred_coor_stage2, ax_coor_pred])\ndisplay(pred_coor_stage3)\npred_coor_stage3.to_csv('stage3_coor.csv', index=False)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:39.991008Z","iopub.execute_input":"2025-06-07T01:14:39.991276Z","iopub.status.idle":"2025-06-07T01:14:40.010464Z","shell.execute_reply.started":"2025-06-07T01:14:39.991251Z","shell.execute_reply":"2025-06-07T01:14:40.009801Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"del pred_coor_stage2, ax_coor_pred, pred_coor, closest_ax\ngc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.011284Z","iopub.execute_input":"2025-06-07T01:14:40.012097Z","iopub.status.idle":"2025-06-07T01:14:40.277555Z","shell.execute_reply.started":"2025-06-07T01:14:40.012074Z","shell.execute_reply":"2025-06-07T01:14:40.276673Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Stage 4: predict each severity","metadata":{}},{"cell_type":"markdown","source":"## Serverity prediction datasets","metadata":{}},{"cell_type":"code","source":"os.listdir('/kaggle/working/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.278394Z","iopub.execute_input":"2025-06-07T01:14:40.278918Z","iopub.status.idle":"2025-06-07T01:14:40.291692Z","shell.execute_reply.started":"2025-06-07T01:14:40.278893Z","shell.execute_reply":"2025-06-07T01:14:40.291066Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class ClassDataset(Dataset):\n def __init__(self, coor, meta, condition, channel, usage='sub'):\n self.coor = coor\n self.meta = meta\n self.condition = condition\n self.usage = usage\n self.sag_window = channel\n self.ax_window = channel\n self.wide_resize = v2.Resize((128, 224))\n #self.wide_resize = v2.Resize((224, 224))\n self.rec_resize = v2.Resize((256, 256))\n self.resize = v2.Resize((128, 128))\n self.resize_3d = v2.Resize((256, 256))\n self.pre_resize = v2.Resize((512, 512))\n self.id = list(meta.study_id.unique())\n if 3637444890 in self.id: \n self.id.remove(3637444890)\n def __getitem__(self, index):\n study_id = self.id[index]\n #print(study_id)\n res = {}\n #try:\n if self.condition == 'scs':\n sagt2_img, ax_img = self.for_scs(study_id)\n res['sagt2'] = sagt2_img.to(torch.float32)\n res['ax'] = ax_img.to(torch.float32)\n #res['sagt1'] = sagt1_img.to(torch.float32)\n elif self.condition == 'nfn':\n ax_img, sagt1_img = self.for_nfn(study_id)\n res['ax'] = ax_img.to(torch.float32)\n res['sagt1'] = sagt1_img.to(torch.float32)\n if self.condition == 'ss':\n ax_img = self.for_ss(study_id)\n #ax_img, sagt1_img, sagt2_img = self.for_ss(study_id)\n res['ax'] = ax_img.to(torch.float32)\n return res, torch.tensor(study_id)\n\n def crop(self, image, x, y, z, x_left, x_right, y_bottom, y_top, wide):\n size = [image[i].shape for i in z]\n #print([self.pre_resize(torch.tensor(image[i])[None, ...]).squeeze() for i, shape in zip(z, size)][0].shape)\n data = torch.stack([torch.tensor(self.pre_resize(torch.tensor(image[i])[None, ...]).squeeze()[max(int((y/shape[0])*512-y_top), 0):int((y/shape[0])*512+y_bottom), max(int((x/shape[1])*512-x_left), 0): int((x/shape[1])*512+x_right)]) for i, shape in zip(z, size)])\n\n if wide:\n data = self.wide_resize(data)\n else:\n data = self.rec_resize(data)\n\n return data\n def for_scs(self, study_id):\n sagt2_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T2/STIR')]\n #display(sagt2_meta)\n ax_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Axial T2')]\n sagt1_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T1')]\n #display(ax_meta)\n sagt2_meta = sagt2_meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n ax_meta = ax_meta.sort_values('ipp_z', ascending=False).reset_index(drop=True)\n sagt1_meta = sagt1_meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n sagt2_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in sagt2_meta.iterrows()]\n ax_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in ax_meta.iterrows()]\n sagt1_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in sagt1_meta.iterrows()]\n sagt1_img = [img if (img.shape[0]> 1 and img.shape[1] > 1) else np.zeros((512, 512)) for img in sagt1_img]\n sagt2_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Spinal Canal Stenosis')]\n ax_right_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Right Subarticular Stenosis')]\n ax_left_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Left Subarticular Stenosis')]\n sagt2_dict = {}\n for _, row in sagt2_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n #display(sagt2_meta)\n #display(row)\n mid = sagt2_meta.loc[(sagt2_meta.series_id==row.series_id)&(sagt2_meta.instance_number==row.instance_number)].index[0]\n if row.level == 'L5/S1':\n ushift = 20\n else:\n ushift = 0\n z = [min(max(mid+w+z_shift, 0), len(sagt2_meta)-1) for w in range(-(self.sag_window-1)//2, ((self.sag_window-1)//2)+1)]\n sagt2_dict[row.level] = self.crop(sagt2_img, row.x+x_shift, row.y+y_shift, z, 96, 32, 40+ushift, 40-ushift, wide=True)\n except: \n pass\n \n # AXIAL T2\n #in_list = ax_meta.instance_number.tolist()\n ax_dict = {}\n if np.random.choice([0, 1]) == 0:\n ax_sub_coor = ax_right_sub_coor\n lrshift = +20\n else:\n ax_sub_coor = ax_left_sub_coor\n lrshift = -20\n for _, row in ax_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n ax_meta_sub = ax_meta.loc[(ax_meta.series_id==row.series_id)]\n ax_meta_sub_original_idx = ax_meta_sub.index.tolist()\n ax_meta_sub = ax_meta_sub.reset_index(drop=True)\n mid = ax_meta_sub.loc[(ax_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(ax_meta_sub)-1) for w in range(-(self.ax_window-1)//2, ((self.ax_window-1)//2)+1)]\n ax_dict[row.level] = self.crop([ax_img[i] for i in range(len(ax_img)) if i in ax_meta_sub_original_idx], row.x+x_shift+lrshift, row.y+y_shift, z, 96, 96, 96, 96, wide=False)\n except: \n pass\n sagt2_img = [sagt2_dict.get(l, torch.zeros((self.sag_window, 128, 224))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n ax_img = [ax_dict.get(l, torch.zeros((self.ax_window, 256, 256))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n return torch.stack(sagt2_img).contiguous(), torch.stack(ax_img).contiguous()#, torch.stack(sagt1_img).contiguous()\n def for_ss(self, study_id):\n #display(sagt2_meta)\n ax_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Axial T2')]\n sagt1_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T1')]\n sagt2_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T2/STIR')]\n #display(ax_meta)\n ax_meta = ax_meta.sort_values('ipp_z', ascending=False).reset_index(drop=True)\n ax_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in ax_meta.iterrows()]\n ax_right_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Right Subarticular Stenosis')]\n ax_left_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Left Subarticular Stenosis')]\n\n # AXIAL T2\n #in_list = ax_meta.instance_number.tolist()\n ax_right_dict = {}\n for _, row in ax_right_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n ax_meta_sub = ax_meta.loc[(ax_meta.series_id==row.series_id)]\n ax_meta_sub_original_idx = ax_meta_sub.index.tolist()\n ax_meta_sub = ax_meta_sub.reset_index(drop=True)\n mid = ax_meta_sub.loc[(ax_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(ax_meta_sub)-1) for w in range(-(self.ax_window-1)//2, ((self.ax_window-1)//2)+1)]\n ax_right_dict[row.level] = self.crop([ax_img[i] for i in range(len(ax_img)) if i in ax_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 160-16, 32+16, 64+32, 64+32, wide=False)\n except: \n pass\n ax_left_dict = {}\n for _, row in ax_left_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n ax_meta_sub = ax_meta.loc[(ax_meta.series_id==row.series_id)]\n ax_meta_sub_original_idx = ax_meta_sub.index.tolist()\n ax_meta_sub = ax_meta_sub.reset_index(drop=True)\n mid = ax_meta_sub.loc[(ax_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(ax_meta_sub)-1) for w in range(-(self.ax_window-1)//2, ((self.ax_window-1)//2)+1)]\n ax_left_dict[row.level] = self.crop([ax_img[i] for i in range(len(ax_img)) if i in ax_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 32+16, 160-16, 64+32, 64+32, wide=False)\n except: \n pass\n ax_right_img = [ax_right_dict.get(l, torch.zeros((self.ax_window, 256, 256))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n ax_left_img = [ax_left_dict.get(l, torch.zeros((self.ax_window, 256, 256))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n ax_img = ax_left_img + ax_right_img\n return torch.stack(ax_img).contiguous()#, torch.stack(sagt1_img).contiguous(), torch.stack(sagt2_img).contiguous()\n\n def for_nfn(self, study_id):\n ax_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Axial T2')]\n sagt1_meta = self.meta.loc[(self.meta.study_id==study_id) & (self.meta.series_description=='Sagittal T1')]\n\n ax_meta = ax_meta.sort_values('ipp_z', ascending=False).reset_index(drop=True)\n sagt1_meta = sagt1_meta.sort_values('ipp_x', ascending=True).reset_index(drop=True)\n ax_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in ax_meta.iterrows()]\n sagt1_img = [self.normalize(self.load_dicom(IMAGE_PATH + f'{row.study_id}/{row.series_id}/{row.instance_number}.dcm')) for _, row in sagt1_meta.iterrows()]\n ax_right_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Right Subarticular Stenosis')]\n ax_left_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Left Subarticular Stenosis')]\n sagt1_right_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Right Neural Foraminal Narrowing')]\n sagt1_left_sub_coor = self.coor.loc[(self.coor.study_id==study_id) & (self.coor.condition=='Left Neural Foraminal Narrowing')]\n\n # SAGITTAL T2\n # not implemented\n\n # AXIAL T2\n #in_list = ax_meta.instance_number.tolist()\n ax_right_dict = {}\n for _, row in ax_right_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n ax_meta_sub = ax_meta.loc[(ax_meta.series_id==row.series_id)]\n ax_meta_sub_original_idx = ax_meta_sub.index.tolist()\n ax_meta_sub = ax_meta_sub.reset_index(drop=True)\n mid = ax_meta_sub.loc[(ax_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(ax_meta_sub)-1) for w in range(-(self.ax_window-1)//2, ((self.ax_window-1)//2)+1)]\n ax_right_dict[row.level] = self.crop([ax_img[i] for i in range(len(ax_img)) if i in ax_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 160-16, 32+16, 64+32, 64+32, wide=False)\n except: \n pass\n ax_left_dict = {}\n for _, row in ax_left_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n ax_meta_sub = ax_meta.loc[(ax_meta.series_id==row.series_id)]\n ax_meta_sub_original_idx = ax_meta_sub.index.tolist()\n ax_meta_sub = ax_meta_sub.reset_index(drop=True)\n mid = ax_meta_sub.loc[(ax_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(ax_meta_sub)-1) for w in range(-(self.ax_window-1)//2, ((self.ax_window-1)//2)+1)]\n ax_left_dict[row.level] = self.crop([ax_img[i] for i in range(len(ax_img)) if i in ax_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 32+16, 160-16, 64+32, 64+32, wide=False)\n except: \n pass\n # SAGITTAL T1\n sagt1_right_dict = {}\n for _, row in sagt1_right_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n if row.level == 'L5/S1':\n ushift = 10\n else:\n ushift = 0\n sagt1_meta_sub = sagt1_meta.loc[(sagt1_meta.series_id==row.series_id)]\n sagt1_meta_sub_original_idx = sagt1_meta_sub.index.tolist()\n sagt1_meta_sub = sagt1_meta_sub.reset_index(drop=True)\n #display(sagt2_meta)\n #display(row)\n mid = sagt1_meta_sub.loc[(sagt1_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(sagt1_meta_sub)-1) for w in range(-(self.sag_window-1)//2, ((self.sag_window-1)//2)+1)]\n sagt1_right_dict[row.level] = self.crop([sagt1_img[i] for i in range(len(sagt1_img)) if i in sagt1_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 96, 64, 32+ushift, 32-ushift, wide=True)\n except: \n pass\n sagt1_left_dict = {}\n for _, row in sagt1_left_sub_coor.iterrows():\n try: \n y_shift = 0\n x_shift = 0\n z_shift = 0\n if row.level == 'L5/S1':\n ushift = 10\n else:\n ushift = 0\n sagt1_meta_sub = sagt1_meta.loc[(sagt1_meta.series_id==row.series_id)]\n sagt1_meta_sub_original_idx = sagt1_meta_sub.index.tolist()\n sagt1_meta_sub = sagt1_meta_sub.reset_index(drop=True)\n mid = sagt1_meta_sub.loc[(sagt1_meta_sub.instance_number==row.instance_number)].index[0]\n z = [min(max(mid+w+z_shift, 0), len(sagt1_meta_sub)-1) for w in range(-(self.sag_window-1)//2, ((self.sag_window-1)//2)+1)]\n sagt1_left_dict[row.level] = self.crop([sagt1_img[i] for i in range(len(sagt1_img)) if i in sagt1_meta_sub_original_idx], row.x+x_shift, row.y+y_shift, z, 96, 64, 32+ushift, 32-ushift, wide=True)\n except: \n pass\n ax_right_img = [ax_right_dict.get(l, torch.zeros((self.ax_window, 256, 256))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n ax_left_img = [ax_left_dict.get(l, torch.zeros((self.ax_window, 256, 256))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n sagt1_right_img = [sagt1_right_dict.get(l, torch.zeros((self.sag_window, 128, 224))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n sagt1_left_img = [sagt1_left_dict.get(l, torch.zeros((self.sag_window, 128, 224))) for l in ['L1/L2', 'L2/L3', 'L3/L4', 'L4/L5', 'L5/S1']]\n ax_img = ax_left_img + ax_right_img\n sagt1_img = sagt1_left_img + sagt1_right_img\n return torch.stack(ax_img).contiguous(), torch.stack(sagt1_img).contiguous()\n\n\n def normalize(self, x):\n lower, upper = np.percentile(x, (1, 99))\n x = np.clip(x, lower, upper)\n x = x - np.min(x)\n x = x / np.max(x)\n return x\n\n def __len__(self):\n return len(self.id)\n\n def load_dicom(self, path):\n dicom = dcm.dcmread(path)\n data = dicom.pixel_array\n return data","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.292602Z","iopub.execute_input":"2025-06-07T01:14:40.292805Z","iopub.status.idle":"2025-06-07T01:14:40.336133Z","shell.execute_reply.started":"2025-06-07T01:14:40.292790Z","shell.execute_reply":"2025-06-07T01:14:40.335542Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Severity predict models","metadata":{}},{"cell_type":"code","source":"class Flatten(nn.Sequential):\n def __init__(self):\n super().__init__(\n nn.AdaptiveAvgPool2d((1, 1)),\n nn.Flatten(1),\n #nn.LayerNorm(512)\n )\n \nclass ConConvnextSCS(nn.Module):\n def __init__(self, direction='sagt2'):\n super().__init__()\n self.direction = direction\n self.ax = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n self.sagt2 = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n self.num_features = self.ax.num_features\n self.flatten_ax = Flatten()\n self.flatten_sagt2 = Flatten()\n self.lin_ax = nn.Linear(self.num_features, 512)\n self.lin_sagt2 = nn.Linear(self.num_features, 512)\n self.aux_ax = nn.Linear(512, 3)\n self.aux_sagt2 = nn.Linear(512, 3)\n self.lin = nn.Linear(512*2, 512)\n self.out = nn.Linear(512, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax, sagt2, sagt1=None, label=None):\n shape = ax.shape\n ax = ax.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n shape=sagt2.shape\n sagt2 = sagt2.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n ax = nn.functional.leaky_relu(self.lin_ax(self.flatten_ax(self.ax.forward_features(ax))))\n sagt2 = nn.functional.leaky_relu(self.lin_sagt2(self.flatten_sagt2(self.sagt2.forward_features(sagt2))))\n x = torch.cat([ax, sagt2], dim=1)\n x = self.lin(x)\n x = nn.functional.leaky_relu(x)\n x = self.dropout(x)\n x = self.out(x)\n return x\n\nclass ConConvnextNFN(nn.Module):\n def __init__(self):\n super().__init__()\n self.ax = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n self.sagt1 = timm.create_model('convnext_base.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n self.num_features = self.ax.num_features\n self.flatten_ax = Flatten()\n self.flatten_sagt1 = Flatten()\n self.lin_ax = nn.Linear(self.num_features, 512)\n self.lin_sagt1 = nn.Linear(self.num_features, 512)\n self.aux_ax = nn.Linear(512, 3)\n self.aux_sagt1 = nn.Linear(512, 3)\n self.lin = nn.Linear(512*2, 512)\n self.out = nn.Linear(512, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax, sagt1, sagt2=None, label=None):\n shape = ax.shape\n ax = ax.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n shape=sagt1.shape\n sagt1 = sagt1.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n ax = nn.functional.leaky_relu(self.lin_ax(self.flatten_ax(self.ax.forward_features(ax))))\n sagt1 = nn.functional.leaky_relu(self.lin_sagt1(self.flatten_sagt1(self.sagt1.forward_features(sagt1))))\n x = torch.cat([ax, sagt1], dim=1)\n x = self.lin(x)\n x = nn.functional.leaky_relu(x)\n x = self.dropout(x)\n x = self.out(x)\n return x\n \nclass ConvnextSS(nn.Module):\n def __init__(self, direction='ax'):\n super().__init__()\n self.direction = direction\n self.encoder = timm.create_model('convnext_large.fb_in22k_ft_in1k_384', in_chans=3, pretrained=False, num_classes=0)\n self.in_features = self.encoder.num_features\n self.flatten = Flatten()\n self.lin = nn.Linear(self.in_features, 512)\n self.out = nn.Linear(512, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax, sagt1=None, sagt2=None, label=None):\n if self.direction == 'ax':\n shape = ax.shape\n x = ax.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n elif self.direction == 'sagt1':\n shape = sagt1.shape\n x = sagt1.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n elif self.direction == 'sagt2':\n shape = sagt2.shape\n x = sagt2.reshape(shape[0]*shape[1], 3, shape[-2], shape[-1])\n x = self.flatten(self.encoder.forward_features(x))\n x = self.lin(x)\n x = nn.functional.leaky_relu(x)\n x = self.dropout(x)\n x = self.out(x)\n return x#, ax, sagt1","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.336766Z","iopub.execute_input":"2025-06-07T01:14:40.336946Z","iopub.status.idle":"2025-06-07T01:14:40.352687Z","shell.execute_reply.started":"2025-06-07T01:14:40.336933Z","shell.execute_reply":"2025-06-07T01:14:40.351949Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class AttentionMIL(nn.Module):\n def __init__(self, input_dim, hidden_dim, num_classes, condition):\n super(AttentionMIL, self).__init__()\n self.condition = condition\n if condition == 'nfn': \n self.attention = nn.Sequential(\n nn.Linear(input_dim, hidden_dim),\n nn.Tanh(),\n nn.Linear(hidden_dim, 1)\n )\n else: \n self.lin = nn.Linear(input_dim, hidden_dim)\n self.attn_score = nn.Linear(hidden_dim, 1)\n self.act = nn.Tanh()\n def forward(self, bags):\n \"\"\"\n Args:\n bags: (batch_size, num_instances, input_dim)\n\n Returns:\n logits: (batch_size, num_classes)\n \"\"\"\n batch_size, num_instances, input_dim = bags.size()\n\n # Attention mechanism\n if self.condition=='nfn': \n attn_scores = self.attention(bags).squeeze(-1) # (batch_size, num_instances)\n else: \n x = self.lin(bags)\n attn_scores = self.attn_score(self.act(x)).squeeze(-1)\n attn_weights = torch.softmax(attn_scores, dim=-1) # (batch_size, num_instances)\n # Weighted sum of instances\n weighted_instances = torch.bmm(attn_weights.unsqueeze(1), bags).squeeze(1) # (batch_size, input_dim)\n\n # Classification\n #logits = self.classifier(weighted_instances)\n return weighted_instances, attn_scores\nclass SelfAttentionMIL(nn.Module):\n def __init__(self, input_dim, hidden_dim, num_classes, is_layer_norm=False):\n super(SelfAttentionMIL, self).__init__()\n self.is_layer_norm = is_layer_norm\n\n # Self-Attention層\n self.self_attn = nn.MultiheadAttention(input_dim, num_heads=8, batch_first=True)\n\n\n # バッグレベルの分類器\n #self.bag_classifier = nn.Sequential(\n self.layer_norm = nn.LayerNorm(input_dim)\n self.dropout = nn.Dropout(p=0.0)\n self.lin = nn.Linear(input_dim, hidden_dim)\n self.act = nn.Tanh()\n self.calc_attn_score = nn.Linear(hidden_dim, 1) # バッグレベルのスコア\n #)\n\n def forward(self, bags):\n # Self-Attention\n attn_output, _ = self.self_attn(bags, bags, bags)\n x = attn_output + bags\n if self.is_layer_norm: \n x = self.layer_norm(x)\n # バッグレベルのAttentionスコアを計算\n #bag_attn_scores = self.bag_classifier(attn_output).squeeze(-1)\n x = self.lin(x)\n bag_attn_scores = self.calc_attn_score(self.act(x)).squeeze(-1)\n bag_attn_weights = torch.softmax(bag_attn_scores, dim=-1)\n\n # Attention重み付き平均でバッグレベルの特徴量を計算\n bag_features = torch.bmm(bag_attn_weights.unsqueeze(1), attn_output).squeeze(1)\n\n return bag_features, bag_attn_scores\n \nclass LSTMMIL(nn.Module):\n def __init__(self, input_dim, hidden_dim, num_classes):\n super(LSTMMIL, self).__init__()\n #self.attention = nn.Sequential(\n # nn.Linear(input_dim, hidden_dim),\n # nn.Tanh(),\n # nn.Linear(hidden_dim, 1)\n #)\n #self.classifier = nn.Linear(input_dim, num_classes)\n self.lstm = nn.LSTM(input_dim, input_dim//2, num_layers=2, batch_first=True, dropout=0.1, bidirectional=True)\n #self.lin = nn.Linear(input_dim, hidden_dim)\n self.aux_attention = nn.Sequential(\n nn.Tanh(),\n nn.Linear(input_dim, 1)\n )\n self.attention = nn.Sequential(\n nn.Tanh(),\n nn.Linear(input_dim, 1)\n )\n def forward(self, bags):\n \"\"\"\n Args:\n bags: (batch_size, num_instances, input_dim)\n\n Returns:\n logits: (batch_size, num_classes)\n \"\"\"\n batch_size, num_instances, input_dim = bags.size()\n\n # Attention mechanism\n #attn_scores = self.attention(bags).squeeze(-1) # (batch_size, num_instances)\n bags_lstm, _ = self.lstm(bags)\n attn_scores = self.attention(bags_lstm).squeeze(-1)\n aux_attn_scores = self.aux_attention(bags_lstm).squeeze(-1)\n attn_weights = torch.softmax(attn_scores, dim=-1) # (batch_size, num_instances)\n #aux_attn_weights = torch.softmax(aux_attn_scores, dim=-1)\n # Weighted sum of instances\n weighted_instances = torch.bmm(attn_weights.unsqueeze(1), bags_lstm).squeeze(1) # (batch_size, input_dim)\n # Classification\n #logits = self.classifier(weighted_instances)\n return weighted_instances, aux_attn_scores\n \nclass SCSMIL(nn.Module):\n def __init__(self, model_name):\n super().__init__()\n if 'convnext' in model_name: \n self.sagt2_encoder = timm.create_model('convnext_small.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n self.ax_encoder = timm.create_model('convnext_small.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n elif 'effv2s' in model_name: \n self.sagt2_encoder = timm.create_model('tf_efficientnetv2_s.in21k_ft_in1k', in_chans=1, pretrained=False, num_classes=0)\n self.ax_encoder = timm.create_model('tf_efficientnetv2_s.in21k_ft_in1k', in_chans=1, pretrained=False, num_classes=0)\n self.sagt2_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1))\n #self.sagt1_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n # nn.Flatten(1))\n self.ax_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1))\n self.sagt2_num_features = self.sagt2_encoder.num_features\n #self.sagt1_num_features = self.sagt1_encoder.num_features\n self.ax_num_features = self.ax_encoder.num_features\n self.sagt2_head = LSTMMIL(self.sagt2_num_features, 512, 3)\n #self.sagt1_head = AttentionMIL(self.sagt1_num_features, 512, 3)\n self.ax_head = LSTMMIL(self.ax_num_features, 512, 3)\n \n self.out = nn.Linear(self.sagt2_num_features + self.ax_num_features, 3)\n self.aux_out = nn.Linear(self.sagt2_num_features, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax, sagt2, sagt1=None):\n ax_shape = ax.shape\n ax = ax.reshape(ax_shape[0]*ax_shape[1]*ax_shape[2], 1, ax_shape[-2], ax_shape[-1])\n ax = self.ax_encoder.forward_features(ax)\n ax = self.ax_flatten(ax)\n ax = ax.reshape(ax_shape[0]*ax_shape[1], ax_shape[2], -1)\n ax_weighted_sum, ax_attn = self.ax_head(ax)\n ax_attn = ax_attn.reshape(ax_shape[0], ax_shape[1], -1)\n\n sagt2_shape = sagt2.shape\n sagt2 = sagt2.reshape(sagt2_shape[0]*sagt2_shape[1]*sagt2_shape[2], 1, sagt2_shape[-2], sagt2_shape[-1])\n sagt2 = self.sagt2_encoder.forward_features(sagt2)\n sagt2 = self.sagt2_flatten(sagt2)\n sagt2 = sagt2.reshape(sagt2_shape[0]*sagt2_shape[1], sagt2_shape[2], -1)\n sagt2_weighted_sum, sagt2_attn = self.sagt2_head(sagt2)\n sagt2_attn = sagt2_attn.reshape(sagt2_shape[0], sagt2_shape[1], -1)\n\n out = torch.cat([ax_weighted_sum, sagt2_weighted_sum], dim=1)\n out = self.out(out)\n sagt2_out = self.aux_out(sagt2_weighted_sum)\n ax_out = self.aux_out(ax_weighted_sum)\n #print(sagt2_attn.shape, ax_attn.shape)\n ax_attn = {'L1/L2': ax_attn[:, 0, :], 'L2/L3': ax_attn[:, 1, :], 'L3/L4': ax_attn[:, 2, :], 'L4/L5': ax_attn[:, 3, :], 'L5/S1': ax_attn[:, 4, :]}\n sagt2_attn = {'L1/L2': sagt2_attn[:, 0, :], 'L2/L3': sagt2_attn[:, 1, :], 'L3/L4': sagt2_attn[:, 2, :], 'L4/L5': sagt2_attn[:, 3, :], 'L5/S1': sagt2_attn[:, 4, :]}\n #sagt1_attn = {'L1/L2': sagt1_attn[:, 0, :], 'L2/L3': sagt1_attn[:, 1, :], 'L3/L4': sagt1_attn[:, 2, :], 'L4/L5': sagt1_attn[:, 3, :], 'L5/S1': sagt1_attn[:, 4, :]}\n return out\n\n\nclass NFNMIL(nn.Module):\n def __init__(self, model_name):\n super().__init__()\n if 'convnext' in model_name: \n self.sagt1_encoder = timm.create_model('convnext_small.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n self.ax_encoder = timm.create_model('convnext_small.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n elif 'effv2s' in model_name: \n self.sagt1_encoder = timm.create_model('tf_efficientnetv2_s.in21k_ft_in1k', in_chans=1, pretrained=False, num_classes=0)\n self.ax_encoder = timm.create_model('tf_efficientnetv2_s.in21k_ft_in1k', in_chans=1, pretrained=False, num_classes=0)\n self.sagt1_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1))\n self.ax_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1))\n self.sagt1_num_features = self.sagt1_encoder.num_features\n self.ax_num_features = self.ax_encoder.num_features\n self.sagt1_head = LSTMMIL(self.sagt1_num_features, 512, 3)\n self.ax_head = LSTMMIL(self.ax_num_features, 512, 3)\n self.out = nn.Linear(self.sagt1_num_features+self.ax_num_features, 3)\n self.aux_out = nn.Linear(self.sagt1_num_features, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax, sagt1):\n ax_shape = ax.shape\n #print(ax.shape, sagt2.shape)\n ax = ax.reshape(ax_shape[0]*ax_shape[1]*ax_shape[2], 1, ax_shape[-2], ax_shape[-1])\n ax = self.ax_encoder.forward_features(ax)\n ax = self.ax_flatten(ax)\n ax = ax.reshape(ax_shape[0]*ax_shape[1], ax_shape[2], -1)\n ax_weighted_sum, ax_attn = self.ax_head(ax)\n ax_attn = ax_attn.reshape(ax_shape[0], ax_shape[1], -1)\n #ax_attn = ax_attn.transpose(1, 2)\n sagt1_shape = sagt1.shape\n sagt1 = sagt1.reshape(sagt1_shape[0]*sagt1_shape[1]*sagt1_shape[2], 1, sagt1_shape[-2], sagt1_shape[-1])\n sagt1 = self.sagt1_encoder.forward_features(sagt1)\n sagt1 = self.sagt1_flatten(sagt1)\n sagt1 = sagt1.reshape(sagt1_shape[0]*sagt1_shape[1], sagt1_shape[2], -1)\n sagt1_weighted_sum, sagt1_attn = self.sagt1_head(sagt1)\n sagt1_attn = sagt1_attn.reshape(sagt1_shape[0], sagt1_shape[1], -1)\n x = torch.cat([ax_weighted_sum, sagt1_weighted_sum], dim=1)\n out = self.out(x)\n return out\n\nclass SSMIL(nn.Module):\n def __init__(self, model_name):\n super().__init__()\n if 'convnext' in model_name: \n self.ax_encoder = timm.create_model('convnext_small.fb_in22k_ft_in1k_384', in_chans=1, pretrained=False, num_classes=0)\n elif 'effv2s' in model_name: \n self.ax_encoder = timm.create_model('tf_efficientnetv2_s.in21k_ft_in1k', in_chans=1, pretrained=False, num_classes=0)\n \n self.ax_flatten = nn.Sequential(nn.AdaptiveAvgPool2d((1,1)),\n nn.Flatten(1))\n self.ax_num_features = self.ax_encoder.num_features\n self.ax_head = LSTMMIL(self.ax_num_features, 512, 3)\n self.out = nn.Linear(self.ax_num_features, 3)\n self.dropout = nn.Dropout(0.0)\n def forward(self, ax):\n ax_shape = ax.shape\n #print(ax.shape, sagt2.shape)\n ax = ax.reshape(ax_shape[0]*ax_shape[1]*ax_shape[2], 1, ax_shape[-2], ax_shape[-1])\n ax = self.ax_encoder.forward_features(ax)\n ax = self.ax_flatten(ax)\n ax = ax.reshape(ax_shape[0]*ax_shape[1], ax_shape[2], -1)\n ax_weighted_sum, ax_attn = self.ax_head(ax)\n ax_attn = ax_attn.reshape(ax_shape[0], ax_shape[1], -1)\n\n #out = torch.cat([ax_weighted_sum, sagt2_weighted_sum], dim=1)\n out = ax_weighted_sum\n out = self.out(out)\n #print(sagt2_attn.shape, ax_attn.shape)\n ax_attn = {'left_L1/L2': ax_attn[:, 0, :],'left_L2/L3': ax_attn[:, 1, :],'left_L3/L4': ax_attn[:, 2, :], 'left_L4/L5': ax_attn[:, 3, :], 'left_L5/S1': ax_attn[:, 4, :],\n 'right_L1/L2': ax_attn[:, 5, :], 'right_L2/L3': ax_attn[:, 6, :], 'right_L3/L4': ax_attn[:, 7, :], 'right_L4/L5': ax_attn[:, 8, :], 'right_L5/S1': ax_attn[:, 9, :]}\n return out","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.353486Z","iopub.execute_input":"2025-06-07T01:14:40.353755Z","iopub.status.idle":"2025-06-07T01:14:40.382841Z","shell.execute_reply.started":"2025-06-07T01:14:40.353731Z","shell.execute_reply":"2025-06-07T01:14:40.382249Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Severity prediction lightning modules","metadata":{}},{"cell_type":"code","source":"class ClassModule(pl.LightningModule):\n def __init__(self, condition, model_name):\n super().__init__()\n self.condition = condition\n if condition == 'scs':\n if 'mil' in model_name: \n self.model = SCSMIL(model_name)\n else: \n self.model = ConConvnextSCS('sagt2')\n elif condition == 'nfn':\n if 'mil' in model_name: \n self.model = NFNMIL(model_name)\n else: \n self.model = ConConvnextNFN()\n elif condition == 'ss':\n if 'mil' in model_name: \n self.model = SSMIL(model_name)\n else: \n self.model = ConvnextSS('ax')\n\n def forward(self, batch):\n preds = self.model(**batch)\n return preds","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.383552Z","iopub.execute_input":"2025-06-07T01:14:40.383778Z","iopub.status.idle":"2025-06-07T01:14:40.396966Z","shell.execute_reply.started":"2025-06-07T01:14:40.383763Z","shell.execute_reply":"2025-06-07T01:14:40.396439Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Severity inference","metadata":{}},{"cell_type":"code","source":"%%time\nprefix = ''\nimport warnings\nwarnings.filterwarnings(\"ignore\")\nseverity_predict = {'scs': {\n 'L1/L2':[], \n 'L2/L3': [], \n 'L3/L4': [], \n 'L4/L5': [], \n 'L5/S1': []\n }, \n 'nfn': {\n 'left_L1/L2': [], \n 'left_L2/L3': [], \n 'left_L3/L4': [], \n 'left_L4/L5': [], \n 'left_L5/S1': [], \n 'right_L1/L2': [], \n 'right_L2/L3': [], \n 'right_L3/L4': [], \n 'right_L4/L5': [], \n 'right_L5/S1': [], \n }, \n 'ss': {\n 'left_L1/L2': [], \n 'left_L2/L3': [], \n 'left_L3/L4': [], \n 'left_L4/L5': [], \n 'left_L5/S1': [], \n 'right_L1/L2': [], \n 'right_L2/L3': [], \n 'right_L3/L4': [], \n 'right_L4/L5': [], \n 'right_L5/S1': [], \n }\n }\nmodel_path_dict = {\n 'scs': [\n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_convnext-s_for_exp0.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_convnext-s_for_exp1.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_convnext-s_for_exp2.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_convnext-s_for_exp3.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_convnext-s_for_exp4.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_effv2s_for_exp0.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_effv2s_for_exp1.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_effv2s_for_exp2.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_effv2s_for_exp3.ckpt', \n '/kaggle/input/rsna-spine-final-models/_scs_classify_5ch_axsagt2-lstm-mil_auxloss_auxdepth_effv2s_for_exp4.ckpt', \n ], \n 'nfn': [\n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_convnext-s_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_convnext-s_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_convnext-s_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_convnext-s_3.ckpt',\n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_convnext-s_4.ckpt',\n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_effv2s_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_effv2s_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_effv2s_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_effv2s_3.ckpt',\n '/kaggle/input/rsna-spine-final-models/nfn_classify_5ch_axsagt1-lstm-mil_auxloss_auxdepth_2shift_effv2s_4.ckpt',\n ], \n 'ss': [\n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_effv2s_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_effv2s_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_effv2s_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_effv2s_3.ckpt',\n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_effv2s_4.ckpt',\n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_convnext-s_0.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_convnext-s_1.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_convnext-s_2.ckpt', \n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_convnext-s_3.ckpt',\n '/kaggle/input/rsna-spine-final-models/ss_classify_5ch_ax-lstm-mil_auxloss_auxdepth_convnext-s_4.ckpt',\n ]\n}\n##############SEVERITY PREDICT#########################\nfor condition in ['nfn', 'scs', 'ss']:\n print(condition)\n for path in model_path_dict[condition]:\n _meta_df = meta_df.copy()\n _coor_df = pred_coor_stage3.copy()\n model_name = path.split('/')[-1]\n if '5ch' in model_name: \n dataset_channel = 5\n else: \n dataset_channel = 3\n dataset_test = ClassDataset(_coor_df, _meta_df, condition, dataset_channel, 'sub')\n data_loader_test = DataLoader(\n dataset_test,\n batch_size=4,\n shuffle=False,\n num_workers=4,\n pin_memory=False\n )\n\n model = ClassModule.load_from_checkpoint(path, condition=condition, model_name=model_name, strict=False)\n model.eval()\n model.zero_grad()\n model.to(device)\n pred_temp = {}\n for k in severity_predict[condition].keys(): \n pred_temp[k] = []\n study_id_list = []\n with torch.no_grad():\n for data in tqdm(data_loader_test, total=len(data_loader_test)):\n images, study_id = data\n for k, v in images.items(): \n images[k] = v.to(device)\n bs = v.shape[0]\n preds = model.forward(images)\n preds = nn.functional.softmax(preds, dim=1)\n preds = preds.reshape((bs, -1, 3))\n preds = preds.to('cpu').detach().numpy()\n #print(preds)\n if condition == 'scs': \n pred_temp['L1/L2'].append(preds[:, 0, :])\n pred_temp['L2/L3'].append(preds[:, 1, :])\n pred_temp['L3/L4'].append(preds[:, 2, :])\n pred_temp['L4/L5'].append(preds[:, 3, :])\n pred_temp['L5/S1'].append(preds[:, 4, :])\n else: \n pred_temp['left_L1/L2'].append(preds[:, 0, :])\n pred_temp['left_L2/L3'].append(preds[:, 1, :])\n pred_temp['left_L3/L4'].append(preds[:, 2, :])\n pred_temp['left_L4/L5'].append(preds[:, 3, :])\n pred_temp['left_L5/S1'].append(preds[:, 4, :])\n pred_temp['right_L1/L2'].append(preds[:, 5, :])\n pred_temp['right_L2/L3'].append(preds[:, 6, :])\n pred_temp['right_L3/L4'].append(preds[:, 7, :])\n pred_temp['right_L4/L5'].append(preds[:, 8, :])\n pred_temp['right_L5/S1'].append(preds[:, 9, :])\n study_id_list.append(study_id.to('cpu').reshape(-1).detach().numpy())\n del images, preds\n gc.collect()\n for k, v in pred_temp.items(): \n severity_predict[condition][k].append(np.concatenate(v))\n all_study_id = np.concatenate(study_id_list)\n del pred_temp, study_id_list\n gc.collect()\n \n for k, v in severity_predict[condition].items(): \n severity_predict[condition][k] = np.mean(np.array(severity_predict[condition][k]), axis=0)\n severity_predict[condition]['study_id'] = all_study_id\n gc.collect()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:14:40.397741Z","iopub.execute_input":"2025-06-07T01:14:40.398013Z","iopub.status.idle":"2025-06-07T01:20:35.988609Z","shell.execute_reply.started":"2025-06-07T01:14:40.397987Z","shell.execute_reply":"2025-06-07T01:20:35.986295Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Stage Final: Submission","metadata":{}},{"cell_type":"code","source":"condition_mapper = {'scs': 'spinal_canal_stenosis', 'nfn': 'neural_foraminal_narrowing', 'ss': 'subarticular_stenosis'}\npredict_list = []\nfor k, v in severity_predict.items():\n condition = condition_mapper[k]\n study_id = severity_predict[k]['study_id']\n for kk, vv in v.items(): \n if kk == 'study_id': \n continue\n level = kk.split('_')[-1].lower().replace('/', '_')\n loc = kk.split('_')[0]\n if loc == 'left':\n loc = 'left_'\n elif loc == 'right':\n loc = 'right_'\n else: \n loc = ''\n row_id = [f'{str(si)}_' + loc + condition + '_' + level for si in study_id]\n df = pd.DataFrame({'row_id': row_id, 'normal_mild': vv[:, 0], 'moderate': vv[:, 1], 'severe': vv[:, 2]})\n predict_list.append(df)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:32:50.878934Z","iopub.execute_input":"2025-06-07T01:32:50.879685Z","iopub.status.idle":"2025-06-07T01:32:50.890969Z","shell.execute_reply.started":"2025-06-07T01:32:50.879658Z","shell.execute_reply":"2025-06-07T01:32:50.890312Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predict_df = pd.concat(predict_list)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:32:56.061027Z","iopub.execute_input":"2025-06-07T01:32:56.061583Z","iopub.status.idle":"2025-06-07T01:32:56.066783Z","shell.execute_reply.started":"2025-06-07T01:32:56.061560Z","shell.execute_reply":"2025-06-07T01:32:56.066233Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"predict_df","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:33:52.197518Z","iopub.execute_input":"2025-06-07T01:33:52.197982Z","iopub.status.idle":"2025-06-07T01:33:52.209840Z","shell.execute_reply.started":"2025-06-07T01:33:52.197960Z","shell.execute_reply":"2025-06-07T01:33:52.209052Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub = pd.read_csv('/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/sample_submission.csv')\nsub = sub.drop(['normal_mild', 'moderate', 'severe'], axis=1)\nsub = sub.merge(predict_df, on='row_id', how='left')\nsub = sub.fillna(1/3)\nsub[['normal_mild', 'moderate', 'severe']] = sub[['normal_mild', 'moderate', 'severe']]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:33:02.152772Z","iopub.execute_input":"2025-06-07T01:33:02.153321Z","iopub.status.idle":"2025-06-07T01:33:02.169100Z","shell.execute_reply.started":"2025-06-07T01:33:02.153299Z","shell.execute_reply":"2025-06-07T01:33:02.168426Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"sub.to_csv('submission.csv', index=False)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:33:28.309979Z","iopub.execute_input":"2025-06-07T01:33:28.310516Z","iopub.status.idle":"2025-06-07T01:33:28.315660Z","shell.execute_reply.started":"2025-06-07T01:33:28.310474Z","shell.execute_reply":"2025-06-07T01:33:28.314814Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"os.listdir('/kaggle/working/')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:54:36.178338Z","iopub.execute_input":"2025-06-07T01:54:36.179021Z","iopub.status.idle":"2025-06-07T01:54:36.184182Z","shell.execute_reply.started":"2025-06-07T01:54:36.178996Z","shell.execute_reply":"2025-06-07T01:54:36.183551Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"data_submiss = pd.read_csv('submission.csv')\ndata_submiss","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-06-07T01:57:38.800206Z","iopub.execute_input":"2025-06-07T01:57:38.800479Z","iopub.status.idle":"2025-06-07T01:57:38.812985Z","shell.execute_reply.started":"2025-06-07T01:57:38.800459Z","shell.execute_reply":"2025-06-07T01:57:38.812324Z"}},"outputs":[],"execution_count":null}]}