{"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":"gpu","dataSources":[{"sourceId":71549,"databundleVersionId":8561470,"sourceType":"competition"},{"sourceId":12249891,"sourceType":"datasetVersion","datasetId":7718540}],"dockerImageVersionId":31041,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"**CLASSIFER with test case**","metadata":{}},{"cell_type":"code","source":"import os\nimport pandas as pd\nfrom sklearn.model_selection import StratifiedShuffleSplit, StratifiedKFold\nfrom pathlib import Path\n\ndef split_train_val_test_5fold(\n    condition: str,\n    test_size: float = 0.2,\n    n_splits: int = 5,\n    random_state: int = 42\n):\n    \"\"\"\n    1) Reads /kaggle/input/csv-files/<condition>.csv, drops NaN scores\n    2) Splits off a stratified test set (test_size fraction)\n    3) Performs a stratified n_splits-fold CV on the remaining data\n    4) Writes out:\n       - ./<condition>/test/<condition>_test.csv\n       - ./<condition>/<condition>_trainval_<n_splits>folds.csv\n       - ./<condition>/fold_0/... fold_{n_splits-1}\n    \"\"\"\n    # 1) Load and clean\n    df = pd.read_csv(f'/kaggle/input/csv-files/{condition}.csv')\n    df = df.dropna(subset=['score']).reset_index(drop=True)\n\n    # 2) Stratified shuffle split → train_val vs test\n    sss = StratifiedShuffleSplit(n_splits=1, test_size=test_size, random_state=random_state)\n    train_val_idx, test_idx = next(sss.split(df, df['score']))\n    df_train_val = df.loc[train_val_idx].reset_index(drop=True)\n    df_test      = df.loc[test_idx].reset_index(drop=True)\n\n    # Save test set\n    condition_dir = Path(f'./{condition}')\n    (condition_dir / 'test').mkdir(parents=True, exist_ok=True)\n    df_test.to_csv(\n        condition_dir / 'test' / f'{condition}_test.csv', \n        index=False\n    )\n    print(f\"Saved test set ({len(df_test)} rows) to: {condition_dir/'test'/f'{condition}_test.csv'}\")\n\n    # 3) Stratified K-Fold on train_val\n    skf = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=random_state)\n    df_train_val['fold'] = -1\n    for fold, (_, val_idx) in enumerate(skf.split(df_train_val, df_train_val['score'])):\n        df_train_val.loc[val_idx, 'fold'] = fold\n\n    # Save combined train_val folds CSV\n    df_train_val.to_csv(\n        condition_dir / f'{condition}_trainval_{n_splits}folds.csv',\n        index=False\n    )\n    print(f\"Saved train/val folds CSV to: {condition_dir/f'{condition}_trainval_{n_splits}folds.csv'}\")\n\n    # 4) Write out per-fold train/val splits\n    for fold in range(n_splits):\n        fold_dir = condition_dir / f'fold_{fold}'\n        fold_dir.mkdir(parents=True, exist_ok=True)\n\n        train_df = df_train_val[df_train_val['fold'] != fold].reset_index(drop=True)\n        val_df   = df_train_val[df_train_val['fold'] == fold].reset_index(drop=True)\n\n        train_df.to_csv(fold_dir / f'{condition}_train.csv', index=False)\n        val_df.to_csv(  fold_dir / f'{condition}_val.csv',   index=False)\n        print(f\"  Fold {fold}: {len(train_df)} train rows, {len(val_df)} val rows\")\n\n# Run for each condition\nfor cond in ['Neural_Foraminal_Narrowing', 'Spinal_Canal_Stenosis', 'Subarticular_Stenosis']:\n    split_train_val_test_5fold(cond)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-07-06T07:45:47.687116Z","iopub.execute_input":"2025-07-06T07:45:47.687419Z","iopub.status.idle":"2025-07-06T07:45:49.445556Z","shell.execute_reply.started":"2025-07-06T07:45:47.687398Z","shell.execute_reply":"2025-07-06T07:45:49.444893Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport pandas as pd\nimport numpy as np\nimport random\nimport shutil\n\n# Define augmentation methods\naugmentations = ['rotate', 'horizontal_flip', 'vertical_flip', 'gaussian_noise']\n\ndef augment_data(df: pd.DataFrame, augmentations: list) -> pd.DataFrame:\n    \"\"\"\n    Oversample minority \"score\" classes to balance against the majority class.\n    Adds an 'augmentation' column to each augmented sample.\n    \"\"\"\n    counts = df['score'].value_counts()\n    maj_cls = counts.idxmax()\n    target1 = counts[maj_cls] // 3\n    target2 = counts[maj_cls] // 2\n\n    minors = [cls for cls, c in counts.items() if c < counts[maj_cls]]\n    if len(minors) < 2:\n        minors = minors * 2\n\n    def sample_and_label(min_cls, target):\n        existing = df[df['score'] == min_cls]\n        needed = max(0, target - len(existing))\n        if needed <= 0:\n            return pd.DataFrame(columns=list(df.columns) + ['augmentation'])\n        sampled = existing.sample(needed, replace=True).copy()\n        sampled['augmentation'] = [random.choice(augmentations) for _ in range(needed)]\n        return sampled\n\n    aug1 = sample_and_label(minors[0], target1)\n    aug2 = sample_and_label(minors[1], target2)\n\n    df = df.copy()\n    df['augmentation'] = None\n    return pd.concat([df, aug1, aug2], ignore_index=True)\n\n# --- Adjusted for train/val/test splits ---\nconditions = ['Neural_Foraminal_Narrowing',\n              'Spinal_Canal_Stenosis',\n              'Subarticular_Stenosis']\nfolds = list(range(5))  # 5-fold CV\n\nfor cond in conditions:\n    # Augment each train fold and copy its val split\n    for fold in folds:\n        base_in  = f'./{cond}/fold_{fold}'\n        base_out = f'./augmented_output/{cond}/fold_{fold}'\n        os.makedirs(base_out, exist_ok=True)\n\n        # load train, augment, save\n        train_csv = os.path.join(base_in, f'{cond}_train.csv')\n        df_train  = pd.read_csv(train_csv)\n        df_aug    = augment_data(df_train, augmentations)\n        df_aug.to_csv(os.path.join(base_out, f'{cond}_augmented_train.csv'), index=False)\n        print(f'→ Augmented train saved: {base_out}/{cond}_augmented_train.csv')\n\n        # copy val unchanged\n        val_csv = os.path.join(base_in, f'{cond}_val.csv')\n        shutil.copy(val_csv, os.path.join(base_out, f'{cond}_val.csv'))\n        print(f'→ Val copied:           {base_out}/{cond}_val.csv')\n\n    # Copy test set unchanged\n    test_in  = f'./{cond}/test/{cond}_test.csv'\n    test_out = f'./augmented_output/{cond}/test'\n    os.makedirs(test_out, exist_ok=True)\n    shutil.copy(test_in, os.path.join(test_out, f'{cond}_test.csv'))\n    print(f'→ Test copied:          {test_out}/{cond}_test.csv')\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-07-06T07:45:49.446917Z","iopub.execute_input":"2025-07-06T07:45:49.447323Z","iopub.status.idle":"2025-07-06T07:45:51.284291Z","shell.execute_reply.started":"2025-07-06T07:45:49.447304Z","shell.execute_reply":"2025-07-06T07:45:51.283524Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ─────────────────────────────────────────────────────────────────────────────\n# Cell: Image Data Preparation for 5-Fold CV + Held-Out Test\n# ─────────────────────────────────────────────────────────────────────────────\n\nimport os\nimport pydicom\nimport pandas as pd\nimport numpy as np\nfrom PIL import Image, ImageOps\nfrom pathlib import Path\n\nclass DataPreparationImage:\n    def __init__(\n        self,\n        dataset_directory: str,\n        condition: str,\n        csv_directory: str,\n        num_folds: int = 5,\n        augmentation_list=None,\n    ):\n        self.dataset_directory = Path(dataset_directory)\n        self.condition       = condition\n        self.csv_directory   = Path(csv_directory)\n        self.num_folds       = num_folds\n        self.augmentation_list = augmentation_list or [\n            'rotate', 'horizontal_flip', 'vertical_flip', 'gaussian_noise'\n        ]\n\n        print(f\"Starting image data prep for {self.condition} ({self.num_folds} folds + test)\")\n        self._create_folders()\n        self._process_all_folds()\n        self._process_test()\n\n    def _create_folders(self):\n        base = Path(f\"./{self.condition}\")\n        base.mkdir(exist_ok=True)\n        # train / val\n        for fold in range(self.num_folds):\n            (base / f\"fold_{fold}\" / \"train\").mkdir(parents=True, exist_ok=True)\n            (base / f\"fold_{fold}\" / \"val\").mkdir(parents=True, exist_ok=True)\n        # held-out test\n        (base / \"test\").mkdir(parents=True, exist_ok=True)\n\n    def _read_csv(self, split: str, fold: int = None) -> pd.DataFrame:\n        \"\"\"\n        split: 'train', 'val', or 'test'\n        fold: required for train/val, ignored for test\n        \"\"\"\n        if split in ('train', 'val'):\n            if fold is None:\n                raise ValueError(\"Must provide fold for train/val\")\n            suffix = 'augmented_train' if split == 'train' else 'val'\n            filename = f\"{self.condition}_{suffix}.csv\"\n            path = self.csv_directory / self.condition / f\"fold_{fold}\" / filename\n        elif split == 'test':\n            filename = f\"{self.condition}_test.csv\"\n            path = self.csv_directory / self.condition / \"test\" / filename\n        else:\n            raise ValueError(f\"Unknown split: {split}\")\n        return pd.read_csv(path)\n\n    def _read_dicom(self, path: Path) -> np.ndarray:\n        ds = pydicom.dcmread(str(path))\n        img = ds.pixel_array.astype(float)\n        img = (img - img.min()) / (img.max() - img.min() + 1e-6) * 255.0\n        rgb = np.stack([img] * 3, axis=-1)\n        return rgb.astype('uint8')\n\n    def _crop(self, image: np.ndarray, x: float, y: float, box: int = 16) -> Image.Image:\n        img = Image.fromarray(image)\n        left, top   = int(x - box), int(y - box)\n        right, bottom = int(x + box), int(y + box)\n        return img.crop((left, top, right, bottom))\n\n    def _apply_augmentation(self, img: Image.Image, aug: str) -> Image.Image:\n        if aug == 'rotate':\n            return img.rotate(np.random.uniform(-20,20), expand=True)\n        if aug == 'horizontal_flip':\n            return ImageOps.mirror(img)\n        if aug == 'vertical_flip':\n            return ImageOps.flip(img)\n        if aug == 'gaussian_noise':\n            arr   = np.array(img)\n            noise = np.random.normal(0,25,arr.shape)\n            return Image.fromarray(np.clip(arr+noise,0,255).astype('uint8'))\n        return img\n\n    def _process_all_folds(self):\n        for fold in range(self.num_folds):\n            print(f\" Processing fold {fold}…\")\n            # TRAIN (with augmentation)\n            df_train = self._read_csv('train', fold)\n            for _, row in df_train.iterrows():\n                sid, seid, inst = row['study_id'], row['series_id'], row['instance_number']\n                x, y    = row['x'], row['y']\n                aug_op  = row.get('augmentation')\n                dcm     = self.dataset_directory / str(sid) / str(seid) / f\"{inst}.dcm\"\n                img     = self._read_dicom(dcm)\n                patch   = self._crop(img, x, y)\n                if aug_op in self.augmentation_list:\n                    out_img = self._apply_augmentation(patch, aug_op)\n                    suffix  = f\"_{aug_op}\"\n                else:\n                    out_img = patch\n                    suffix  = \"\"\n                fname = f\"{sid}_{seid}_{inst}_{int(x)}_{int(y)}{suffix}.png\"\n                out_path = Path(self.condition) / f\"fold_{fold}\" / \"train\" / fname\n                out_img.save(out_path)\n\n            # VAL (no augmentation)\n            df_val = self._read_csv('val', fold)\n            for _, row in df_val.iterrows():\n                sid, seid, inst = row['study_id'], row['series_id'], row['instance_number']\n                x, y    = row['x'], row['y']\n                dcm     = self.dataset_directory / str(sid) / str(seid) / f\"{inst}.dcm\"\n                img     = self._read_dicom(dcm)\n                patch   = self._crop(img, x, y)\n                fname   = f\"{sid}_{seid}_{inst}_{int(x)}_{int(y)}.png\"\n                out_path = Path(self.condition) / f\"fold_{fold}\" / \"val\" / fname\n                patch.save(out_path)\n\n    def _process_test(self):\n        print(\" Processing held-out test set…\")\n        df_test = self._read_csv('test')\n        for _, row in df_test.iterrows():\n            sid, seid, inst = row['study_id'], row['series_id'], row['instance_number']\n            x, y    = row['x'], row['y']\n            dcm     = self.dataset_directory / str(sid) / str(seid) / f\"{inst}.dcm\"\n            img     = self._read_dicom(dcm)\n            patch   = self._crop(img, x, y)\n            fname   = f\"{sid}_{seid}_{inst}_{int(x)}_{int(y)}.png\"\n            out_path = Path(self.condition) / \"test\" / fname\n            patch.save(out_path)\n        print(f\" Saved {len(df_test)} test images to ./{self.condition}/test\")\n\n\n# ─────────────────────────────────────────────────────────────────────────────\n# Run for your condition (example: Neural Foraminal Narrowing)\n# ─────────────────────────────────────────────────────────────────────────────\n\nDATASET_DIR = \"/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/train_images\"\nCSV_DIR     = \"/kaggle/working/augmented_output\"\n\nDataPreparationImage(\n    dataset_directory=DATASET_DIR,\n    condition=\"Neural_Foraminal_Narrowing\",\n    csv_directory=CSV_DIR,\n    num_folds=5\n)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-07-06T07:45:51.285058Z","iopub.execute_input":"2025-07-06T07:45:51.285240Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import os\nimport pandas as pd\nfrom pathlib import Path\n\n# 1) Parameters\nCSV_DIR    = Path('/kaggle/working/augmented_output')   # where your augmented CSVs live\nCONDITIONS = [\n    'Spinal_Canal_Stenosis',\n    'Neural_Foraminal_Narrowing',\n    'Subarticular_Stenosis'\n]\nFOLDS      = list(range(5))  # five folds: 0–4\nOUT_ROOT   = Path('/kaggle/working')    # base for label output folders\n\n# 2) Helper: build 'subject' and 'label' columns\ndef working_on_csv(csv_path: Path, split: str) -> pd.DataFrame:\n    df = pd.read_csv(csv_path)\n\n    # ensure we always have an 'augmentation' column\n    if 'augmentation' not in df.columns:\n        df['augmentation'] = None\n\n    def make_subject(r):\n        name = f\"{r.study_id}_{r.series_id}_{r.instance_number}_{int(r.x)}_{int(r.y)}.png\"\n        # only append \"_augmented\" on training samples that were actually augmented\n        if split == 'train' and pd.notna(r.augmentation) and r.augmentation:\n            name = name.replace('.png', '_augmented.png')\n        return name\n\n    def make_label(score):\n        # if your original CSV already used 1,2,3 as numeric scores, just cast\n        if pd.api.types.is_numeric_dtype(type(score)):\n            return int(score)\n        # otherwise map from the text labels you used\n        return {\n            'Normal/Mild': 1,\n            'Moderate':    2,\n            'Severe':      3\n        }.get(score, None)\n\n    df['subject'] = df.apply(make_subject, axis=1)\n    df['label']   = df['score'].apply(make_label)\n\n    return df[['subject', 'label']]\n\n# 3) Create output folder structure (including test)\nfor cond in CONDITIONS:\n    base = OUT_ROOT / f\"{cond}_label\"\n    for fold in FOLDS:\n        (base / f\"fold_{fold}\").mkdir(parents=True, exist_ok=True)\n    (base / \"test\").mkdir(parents=True, exist_ok=True)\n\n# 4) Process each condition: train/val and then test\nfor cond in CONDITIONS:\n    label_root = OUT_ROOT / f\"{cond}_label\"\n\n    # train & val\n    for fold in FOLDS:\n        train_csv = CSV_DIR / cond / f\"fold_{fold}\" / f\"{cond}_augmented_train.csv\"\n        val_csv   = CSV_DIR / cond / f\"fold_{fold}\" / f\"{cond}_val.csv\"\n\n        train_labels = working_on_csv(train_csv, 'train')\n        val_labels   = working_on_csv(val_csv,   'val')\n\n        out_base = label_root / f\"fold_{fold}\"\n        train_labels.to_csv(out_base / f\"{cond}_augmented_labels.csv\", index=False)\n        val_labels.to_csv(  out_base / f\"{cond}_val_labels.csv\",       index=False)\n\n        print(f\"→ {cond} fold {fold}: train={len(train_labels)}, val={len(val_labels)}\")\n\n    # test\n    test_csv    = CSV_DIR / cond / \"test\" / f\"{cond}_test.csv\"\n    test_labels = working_on_csv(test_csv, 'test')\n    out_test    = label_root / \"test\"\n    test_labels.to_csv(out_test / f\"{cond}_test_labels.csv\", index=False)\n\n    print(f\"→ {cond} test: {len(test_labels)} rows → {out_test}/{cond}_test_labels.csv\")\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ─────────────────────────────────────────────────────────────────────────────\n# Cell: 5-Fold CV + Held-Out Test Evaluation for Neural Foraminal Narrowing\n# ─────────────────────────────────────────────────────────────────────────────\n\nimport os\nimport gc\nimport pandas as pd\nimport numpy as np\nfrom pathlib import Path\nimport tensorflow as tf\nfrom tensorflow.keras.applications import ResNet50\nfrom tensorflow.keras.models import Model, load_model\nfrom tensorflow.keras.layers import Dense, Dropout, GlobalAveragePooling2D\nfrom tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom sklearn.metrics import (\n    accuracy_score, precision_score, recall_score, f1_score,\n    confusion_matrix, roc_curve, auc\n)\nfrom sklearn.preprocessing import label_binarize\nimport matplotlib.pyplot as plt\nimport json\n\n# ── Configuration ────────────────────────────────────────────────────────────\nCONDITION     = 'Neural_Foraminal_Narrowing'\nFOLDS         = list(range(5))\nIMAGE_ROOT_TMPL = '/kaggle/working/Neural_Foraminal_Narrowing/fold_{fold}'\nLABEL_ROOT_TMPL = '/kaggle/working/Neural_Foraminal_Narrowing_label/fold_{fold}'\nTEST_IMAGE_DIR  = f'/kaggle/working/{CONDITION}/test'\nTEST_LABEL_CSV  = f'/kaggle/working/{CONDITION}_label/test/{CONDITION}_test_labels.csv'\nRESULTS_DIR     = Path(f'./results/{CONDITION}')\nBATCH_SIZE   = 16\nPATIENCE     = 5\nEPOCHS       = 10\nLEARNING_RATE= 1e-4\nIMG_SIZE     = (224, 224)\n\n# ── Focal Loss ────────────────────────────────────────────────────────────────\ndef focal_loss(gamma=2., alpha=0.25):\n    def focal_loss_fixed(y_true, y_pred):\n        eps = tf.keras.backend.epsilon()\n        y_pred = tf.clip_by_value(y_pred, eps, 1. - eps)\n        ce     = -y_true * tf.math.log(y_pred)\n        loss   = alpha * tf.math.pow(1 - y_pred, gamma) * ce\n        return tf.reduce_sum(loss, axis=1)\n    return focal_loss_fixed\n\n# ── Keep track of each fold’s validation F1 ─────────────────────────────────\nfold_results = []\n\nfor fold in FOLDS:\n    print(f\"\\n▶ Training fold {fold}\")\n    IMAGE_ROOT = IMAGE_ROOT_TMPL.format(fold=fold)\n    LABEL_ROOT = LABEL_ROOT_TMPL.format(fold=fold)\n\n    # — Load labels\n    train_df = pd.read_csv(Path(LABEL_ROOT) / f'{CONDITION}_augmented_labels.csv')\n    val_df   = pd.read_csv(Path(LABEL_ROOT) / f'{CONDITION}_val_labels.csv')\n\n    # — Strip “_augmented” and re-index to 0/1/2\n    for df in (train_df, val_df):\n        df['subject'] = df['subject'].str.replace('_augmented', '', regex=False)\n        df['label']   = (df['label'].astype(int) - 1).astype(str)\n\n    # — Generators (only rescale)\n    datagen = ImageDataGenerator(rescale=1./255)\n    train_gen = datagen.flow_from_dataframe(\n        train_df, directory=os.path.join(IMAGE_ROOT,'train'),\n        x_col='subject', y_col='label',\n        target_size=IMG_SIZE, batch_size=BATCH_SIZE,\n        class_mode='categorical', shuffle=True\n    )\n    val_gen = datagen.flow_from_dataframe(\n        val_df, directory=os.path.join(IMAGE_ROOT,'val'),\n        x_col='subject', y_col='label',\n        target_size=IMG_SIZE, batch_size=BATCH_SIZE,\n        class_mode='categorical', shuffle=False\n    )\n\n    # — Build & compile model\n    base = ResNet50(include_top=False, weights='imagenet', input_shape=IMG_SIZE+(3,))\n    x = GlobalAveragePooling2D()(base.output)\n    x = Dense(256, activation='relu')(x)\n    x = Dropout(0.5)(x)\n    preds = Dense(3, activation='softmax')(x)\n    model = Model(base.input, preds)\n    model.compile(\n        optimizer=tf.keras.optimizers.Adam(LEARNING_RATE),\n        loss=focal_loss(gamma=2.0, alpha=0.5),\n        metrics=['accuracy']\n    )\n\n    # — Callbacks & checkpoints\n    out_dir = RESULTS_DIR / f'fold_{fold}'\n    out_dir.mkdir(parents=True, exist_ok=True)\n    ckpt = ModelCheckpoint(out_dir/'best_model.keras', monitor='val_loss',\n                           save_best_only=True, verbose=1)\n    es   = EarlyStopping(monitor='val_loss', patience=PATIENCE,\n                         restore_best_weights=True, verbose=1)\n\n    # — Train\n    history = model.fit(\n        train_gen, validation_data=val_gen,\n        epochs=EPOCHS, callbacks=[es,ckpt], verbose=2\n    )\n    \n    hist_path = out_dir / 'history.json'\n    with open(hist_path, 'w') as f:\n        json.dump(history.history, f)\n\n    # — Evaluate on validation\n    val_gen.reset()\n    y_pred = np.argmax(model.predict(val_gen), axis=1)\n    y_true = val_gen.classes\n    fold_metrics = {\n        'fold': fold,\n        'Accuracy':  accuracy_score(y_true, y_pred),\n        'Precision': precision_score(y_true, y_pred, average='weighted'),\n        'Recall':    recall_score(y_true, y_pred, average='weighted'),\n        'F1 Score':  f1_score(y_true, y_pred, average='weighted')\n    }\n    fold_results.append(fold_metrics)\n\n    # — Save fold metrics\n    with open(out_dir/'metrics.txt','w') as f:\n        for k,v in fold_metrics.items():\n            if k!='fold': f.write(f\"{k}: {v:.4f}\\n\")\n    print(f\"✅ Fold {fold} metrics:\", fold_metrics)\n\n    # — Cleanup\n    tf.keras.backend.clear_session()\n    gc.collect()\n\n# ── Pick best fold by validation F1 ─────────────────────────────────────────\nbest = max(fold_results, key=lambda x: x['F1 Score'])\nbest_fold = best['fold']\nprint(f\"\\n▶ Best fold selected: {best_fold} (F1 = {best['F1 Score']:.4f})\")\n\n# ── Held-Out Test Evaluation ────────────────────────────────────────────────\nprint(\"▶ Evaluating on held-out test set…\")\n\n# — Load best model\nbest_model_path = RESULTS_DIR / f'fold_{best_fold}' / 'best_model.keras'\nmodel = load_model(\n    best_model_path,\n    custom_objects={'focal_loss_fixed': focal_loss(gamma=2.0, alpha=0.5)}\n)\n\n# — Prepare test DataFrame\ntest_df = pd.read_csv(TEST_LABEL_CSV)\ntest_df['label'] = (test_df['label'].astype(int) - 1).astype(str)\n\ntest_datagen = ImageDataGenerator(rescale=1./255)\ntest_gen = test_datagen.flow_from_dataframe(\n    test_df, directory=TEST_IMAGE_DIR,\n    x_col='subject', y_col='label',\n    target_size=IMG_SIZE, batch_size=BATCH_SIZE,\n    class_mode='categorical', shuffle=False\n)\n\n# — Predict\nprobs = model.predict(test_gen, verbose=1)\ny_pred = np.argmax(probs, axis=1)\ny_true = test_gen.classes\n\n# — Compute metrics\nacc_test      = accuracy_score(y_true, y_pred)\nerr_test      = 1 - acc_test\nprec_test     = precision_score(y_true, y_pred, average='weighted')\nrec_test      = recall_score(y_true, y_pred, average='weighted')\nf1_test       = f1_score(y_true, y_pred, average='weighted')\ncm            = confusion_matrix(y_true, y_pred)\ny_true_bin    = label_binarize(y_true, classes=[0,1,2])\n\n# — ROC / AUC per class\nfpr, tpr, roc_auc = {}, {}, {}\nfor i in range(3):\n    fpr[i], tpr[i], _ = roc_curve(y_true_bin[:,i], probs[:,i])\n    roc_auc[i] = auc(fpr[i], tpr[i])\n\n# — Save test results\ntest_out = RESULTS_DIR / 'test'\ntest_out.mkdir(parents=True, exist_ok=True)\n\nwith open(test_out/'metrics_test.txt','w') as f:\n    f.write(f\"Accuracy:  {acc_test:.4f}\\n\")\n    f.write(f\"Error:     {err_test:.4f}\\n\")\n    f.write(f\"Precision: {prec_test:.4f}\\n\")\n    f.write(f\"Recall:    {rec_test:.4f}\\n\")\n    f.write(f\"F1 Score:  {f1_test:.4f}\\n\")\n\nnp.savetxt(test_out/'confusion_matrix.csv', cm, delimiter=',', fmt='%d')\nwith open(test_out/'roc_auc.txt','w') as f:\n    for i,aucv in roc_auc.items():\n        f.write(f\"Class {i} AUC: {aucv:.4f}\\n\")\n\n# — Plot & save ROC curves\nplt.figure()\nfor i in range(3):\n    plt.plot(fpr[i], tpr[i], label=f'Class {i} (AUC={roc_auc[i]:.2f})')\nplt.plot([0,1],[0,1],'k--')\nplt.xlabel('False Positive Rate')\nplt.ylabel('True Positive Rate')\nplt.title('ROC Curves on Held-Out Test Set')\nplt.legend(loc='lower right')\nplt.savefig(test_out/'roc_curves.png')\nplt.close()\n\nprint(f\"✅ Saved test evaluation to: {test_out}\")\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ─────────────────────────────────────────────────────────────────────────────\n# Cell: Summary for Best Fold – Loss Curve, Confusion Matrix & ROC\n# ─────────────────────────────────────────────────────────────────────────────\n\nimport json\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\n\nfrom pathlib import Path\nfrom sklearn.metrics import (\n    confusion_matrix, ConfusionMatrixDisplay,\n    roc_curve, auc\n)\nfrom sklearn.preprocessing import label_binarize\nfrom tensorflow.keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.keras.models import load_model\n\n# ── Config ────────────────────────────────────────────────────────────────────\nCONDITION   = 'Neural_Foraminal_Narrowing'\nRESULTS_DIR = Path(f'./results/{CONDITION}')\nLABEL_DIR   = Path(f'/kaggle/working/{CONDITION}_label')\nIMG_DIR     = Path(f'/kaggle/working/{CONDITION}')\nBATCH_SIZE  = 16\nIMG_SIZE    = (224, 224)\n\n# ── 1) Pick best fold by saved metrics ────────────────────────────────────────\nfold_dirs = sorted(RESULTS_DIR.glob('fold_*'), key=lambda p: int(p.name.split('_')[1]))\nbest_fold = None\nbest_f1   = -1.0\n\nfor fd in fold_dirs:\n    # read the F1 from metrics.txt\n    txt = (fd / 'metrics.txt').read_text().splitlines()\n    f1_line = [l for l in txt if l.startswith('F1 Score')][0]\n    f1_val  = float(f1_line.split(':')[-1])\n    fold_idx = int(fd.name.split('_')[1])\n    if f1_val > best_f1:\n        best_f1   = f1_val\n        best_fold = fold_idx\n\nprint(f\"✨ Best fold is {best_fold} with val F1 = {best_f1:.4f}\")\n\nbest_dir = RESULTS_DIR / f'fold_{best_fold}'\n\n# ── 2) Plot Train vs Val Loss ────────────────────────────────────────────────\nhist = json.load(open(best_dir / 'history.json'))\nepochs = range(1, len(hist['loss'])+1)\n\nplt.figure()\nplt.plot(epochs, hist['loss'],    label='Train Loss')\nplt.plot(epochs, hist['val_loss'],label='Val Loss')\nplt.xlabel('Epoch'); plt.ylabel('Loss')\nplt.title(f'Fold {best_fold} Loss Curve')\nplt.legend()\nplt.show()\n\n# ── 3) Build validation generator for best fold ─────────────────────────────\n# load labels\nval_csv = LABEL_DIR / 'fold_{}'.format(best_fold) / f'{CONDITION}_val_labels.csv'\nval_df  = pd.read_csv(val_csv)\nval_df['subject'] = val_df['subject'].str.replace('_augmented', '', regex=False)\nval_df['label']   = (val_df['label'].astype(int) - 1).astype(str)\n\ntest_datagen = ImageDataGenerator(rescale=1./255)\nval_gen = test_datagen.flow_from_dataframe(\n    dataframe=val_df,\n    directory=IMG_DIR / f'fold_{best_fold}' / 'val',\n    x_col='subject', y_col='label',\n    target_size=IMG_SIZE, batch_size=BATCH_SIZE,\n    class_mode='categorical', shuffle=False\n)\n\n# class names\ninv_map = {v:k for k,v in val_gen.class_indices.items()}\nclasses = [inv_map[i] for i in range(len(inv_map))]\n\n# ── 4) Load best model & predict ────────────────────────────────────────────\nmodel = load_model(\n    best_dir / 'best_model.keras',\n    custom_objects={'focal_loss_fixed': focal_loss(gamma=2.0, alpha=0.5)}\n)\n\nprint(\"Val folder listing:\", sorted((IMG_DIR/f'fold_{best_fold}'/'val').iterdir())[:5])\nprint(\"Val subjects:\", val_df['subject'].unique()[:5])\n\nval_gen.reset()\nprobs   = model.predict(val_gen, verbose=0)\ny_true  = val_gen.classes\ny_pred  = np.argmax(probs, axis=1)\n\n# ── 5) Plot Confusion Matrix ────────────────────────────────────────────────\ncm = confusion_matrix(y_true, y_pred)\ndisp = ConfusionMatrixDisplay(cm, display_labels=classes)\n\nplt.figure(figsize=(5,5))\ndisp.plot(ax=plt.gca(), cmap='Blues', colorbar=False)\nplt.title(f'Fold {best_fold} Validation Confusion Matrix')\nplt.show()\n\n# ── 6) Plot ROC Curve ───────────────────────────────────────────────────────\n# binarize for multiclass\ny_bin   = label_binarize(y_true, classes=list(range(len(classes))))\nfpr, tpr, roc_auc = {}, {}, {}\n\nfor i in range(len(classes)):\n    fpr[i], tpr[i], _   = roc_curve(y_bin[:,i], probs[:,i])\n    roc_auc[i]          = auc(fpr[i], tpr[i])\n\nplt.figure()\nfor i, cls in enumerate(classes):\n    plt.plot(fpr[i], tpr[i], label=f'{cls} (AUC={roc_auc[i]:.2f})')\nplt.plot([0,1],[0,1], 'k--', alpha=0.3)\nplt.xlabel('False Positive Rate'); plt.ylabel('True Positive Rate')\nplt.title(f'Fold {best_fold} ROC Curve')\nplt.legend(loc='lower right')\nplt.show()\n","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}