{ "id": 269383, "title": "How to prepare the data for model", "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/269383", "author_name": "", "post_date": "2021-08-31T12:31:30.931647Z", "votes": 1, "comment_count": 3, "views": 0, "content": "

Hi All,
\nI am newbie to medical image analysis, I just started this competition today .
\nI want to through the some of the notebooks, but I didn't understand .
\nplease share me any blog or video for data preparation for model.
\nplease help me , I would like to learn this things.

", "messages": [ { "id": "1497776", "postDate": "08/31/2021 12:31:30", "content": "

Hi All,
\nI am newbie to medical image analysis, I just started this competition today .
\nI want to through the some of the notebooks, but I didn't understand .
\nplease share me any blog or video for data preparation for model.
\nplease help me , I would like to learn this things.

", "rawMarkdown": "Hi All,\nI am newbie to medical image analysis, I just started this competition today .\nI want to through the some of the notebooks, but I didn't understand .\nplease share me any blog or video for data preparation for model.\nplease help me , I would like to learn this things.", "votes": null }, { "id": "1498065", "postDate": "08/31/2021 16:17:44", "content": "

I'm unsure what others have been doing so this is definitely not the best method but you can consider adding zero-padding the scans to equalize the shape across the training dataset. I looped over the dataset and found the smallest sizes you could squeeze the images by removing zero values which were:
\ntarget_shapes = {\n 'FLAIR': (538, 477),\n 'T1w': (445, 394),\n 'T1wCE': (424, 388),\n 'T2w': (797, 616)\n }

\n

I know this isn't the best method, but it's a starting point if you're eager to get a benchmark for what a simple 2D CNN could do using a single scan (turns out not much better than guessing…).

\n

And the padding algorithm:

\n
def normalize_array(array, target_shape):\n    while np.all(array.shape != target_shape):\n        bool_array = array.shape < np.array(target_shape)\n        for i, small in enumerate(bool_array):\n            # if axis i is smaller than the target shape, padding should be applied across axis i\n            if small:\n                dx = target_shape[i] - array.shape[i]\n                # if adding zeros to dimension 0, then padding vectors should match along dimension 1\n                if i == 0:\n                    array = np.concatenate([np.zeros((dx // 2, array.shape[1])), array,\n                                            np.zeros((dx // 2 + int(dx % 2 != 0), array.shape[1]))], axis=i)\n                # and vice versa\n                elif i == 1:\n                    array = np.concatenate([np.zeros((array.shape[0], dx // 2)), array,\n                                            np.zeros((array.shape[0], dx // 2 + int(dx % 2 != 0)))], axis=i)\n\n            # if axis i is larger than the target shape, cut off the shape\n            elif not small:\n                dx = array.shape[i] - target_shape[i]\n                if i == 0:\n                    array = array[dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n                else:\n                    array = array[:, dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n    return array\n
\n

Other ideas such as factoring the orientation (coronal, transverse, sagittal) in as a feature are also ideas worth considering.

", "rawMarkdown": "I'm unsure what others have been doing so this is definitely not the best method but you can consider adding zero-padding the scans to equalize the shape across the training dataset. I looped over the dataset and found the smallest sizes you could squeeze the images by removing zero values which were:\n`target_shapes = {\n 'FLAIR': (538, 477),\n 'T1w': (445, 394),\n 'T1wCE': (424, 388),\n 'T2w': (797, 616)\n }`\n\nI know this isn't the best method, but it's a starting point if you're eager to get a benchmark for what a simple 2D CNN could do using a single scan (turns out not much better than guessing...).\n\nAnd the padding algorithm:\n```\ndef normalize_array(array, target_shape):\n while np.all(array.shape != target_shape):\n bool_array = array.shape < np.array(target_shape)\n for i, small in enumerate(bool_array):\n # if axis i is smaller than the target shape, padding should be applied across axis i\n if small:\n dx = target_shape[i] - array.shape[i]\n # if adding zeros to dimension 0, then padding vectors should match along dimension 1\n if i == 0:\n array = np.concatenate([np.zeros((dx // 2, array.shape[1])), array,\n np.zeros((dx // 2 + int(dx % 2 != 0), array.shape[1]))], axis=i)\n # and vice versa\n elif i == 1:\n array = np.concatenate([np.zeros((array.shape[0], dx // 2)), array,\n np.zeros((array.shape[0], dx // 2 + int(dx % 2 != 0)))], axis=i)\n\n # if axis i is larger than the target shape, cut off the shape\n elif not small:\n dx = array.shape[i] - target_shape[i]\n if i == 0:\n array = array[dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n else:\n array = array[:, dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n return array\n```\n\nOther ideas such as factoring the orientation (coronal, transverse, sagittal) in as a feature are also ideas worth considering.", "votes": null }, { "id": "1498723", "postDate": "09/01/2021 06:44:46", "content": "

Thanks Jasper

", "rawMarkdown": "Thanks Jasper", "votes": null }, { "id": "1500301", "postDate": "09/02/2021 09:10:15", "content": "

HI suri, this notebook: https://www.kaggle.com/experienceinai/aivo-vii-nopea includes step-by-step description about data preparation and modelling. Including pictures and also possibility to show additional information if you need.

", "rawMarkdown": "HI suri, this notebook: https://www.kaggle.com/experienceinai/aivo-vii-nopea includes step-by-step description about data preparation and modelling. Including pictures and also possibility to show additional information if you need.", "votes": null } ], "comments": [ { "id": 1498065, "author_name": "jasperbutcher", "author_url": "", "post_date": "08/31/2021 16:17:44", "content": "

I'm unsure what others have been doing so this is definitely not the best method but you can consider adding zero-padding the scans to equalize the shape across the training dataset. I looped over the dataset and found the smallest sizes you could squeeze the images by removing zero values which were:
\ntarget_shapes = {\n 'FLAIR': (538, 477),\n 'T1w': (445, 394),\n 'T1wCE': (424, 388),\n 'T2w': (797, 616)\n }

\n

I know this isn't the best method, but it's a starting point if you're eager to get a benchmark for what a simple 2D CNN could do using a single scan (turns out not much better than guessing…).

\n

And the padding algorithm:

\n
def normalize_array(array, target_shape):\n    while np.all(array.shape != target_shape):\n        bool_array = array.shape < np.array(target_shape)\n        for i, small in enumerate(bool_array):\n            # if axis i is smaller than the target shape, padding should be applied across axis i\n            if small:\n                dx = target_shape[i] - array.shape[i]\n                # if adding zeros to dimension 0, then padding vectors should match along dimension 1\n                if i == 0:\n                    array = np.concatenate([np.zeros((dx // 2, array.shape[1])), array,\n                                            np.zeros((dx // 2 + int(dx % 2 != 0), array.shape[1]))], axis=i)\n                # and vice versa\n                elif i == 1:\n                    array = np.concatenate([np.zeros((array.shape[0], dx // 2)), array,\n                                            np.zeros((array.shape[0], dx // 2 + int(dx % 2 != 0)))], axis=i)\n\n            # if axis i is larger than the target shape, cut off the shape\n            elif not small:\n                dx = array.shape[i] - target_shape[i]\n                if i == 0:\n                    array = array[dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n                else:\n                    array = array[:, dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n    return array\n
\n

Other ideas such as factoring the orientation (coronal, transverse, sagittal) in as a feature are also ideas worth considering.

", "votes": null, "replies": [ { "id": 1498723, "author_name": "sureshsuri", "author_url": "", "post_date": "09/01/2021 06:44:46", "content": "

Thanks Jasper

", "votes": null, "replies": [] } ] }, { "id": 1500301, "author_name": "experienceinai", "author_url": "", "post_date": "09/02/2021 09:10:15", "content": "

HI suri, this notebook: https://www.kaggle.com/experienceinai/aivo-vii-nopea includes step-by-step description about data preparation and modelling. Including pictures and also possibility to show additional information if you need.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1497776": "Hi All,\nI am newbie to medical image analysis, I just started this competition today .\nI want to through the some of the notebooks, but I didn't understand .\nplease share me any blog or video for data preparation for model.\nplease help me , I would like to learn this things.", "1498065": "I'm unsure what others have been doing so this is definitely not the best method but you can consider adding zero-padding the scans to equalize the shape across the training dataset. I looped over the dataset and found the smallest sizes you could squeeze the images by removing zero values which were:\n`target_shapes = {\n 'FLAIR': (538, 477),\n 'T1w': (445, 394),\n 'T1wCE': (424, 388),\n 'T2w': (797, 616)\n }`\n\nI know this isn't the best method, but it's a starting point if you're eager to get a benchmark for what a simple 2D CNN could do using a single scan (turns out not much better than guessing...).\n\nAnd the padding algorithm:\n```\ndef normalize_array(array, target_shape):\n while np.all(array.shape != target_shape):\n bool_array = array.shape < np.array(target_shape)\n for i, small in enumerate(bool_array):\n # if axis i is smaller than the target shape, padding should be applied across axis i\n if small:\n dx = target_shape[i] - array.shape[i]\n # if adding zeros to dimension 0, then padding vectors should match along dimension 1\n if i == 0:\n array = np.concatenate([np.zeros((dx // 2, array.shape[1])), array,\n np.zeros((dx // 2 + int(dx % 2 != 0), array.shape[1]))], axis=i)\n # and vice versa\n elif i == 1:\n array = np.concatenate([np.zeros((array.shape[0], dx // 2)), array,\n np.zeros((array.shape[0], dx // 2 + int(dx % 2 != 0)))], axis=i)\n\n # if axis i is larger than the target shape, cut off the shape\n elif not small:\n dx = array.shape[i] - target_shape[i]\n if i == 0:\n array = array[dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n else:\n array = array[:, dx // 2:-(dx // 2 + int(dx % 2 != 0))]\n return array\n```\n\nOther ideas such as factoring the orientation (coronal, transverse, sagittal) in as a feature are also ideas worth considering.", "1498723": "Thanks Jasper", "1500301": "HI suri, this notebook: https://www.kaggle.com/experienceinai/aivo-vii-nopea includes step-by-step description about data preparation and modelling. Including pictures and also possibility to show additional information if you need." }, "source": "meta" }