{ "id": 341752, "title": "Build a MONAI based pipeline for this challenge", "url": "/competitions/mayo-clinic-strip-ai/discussion/341752", "author_name": "Yiheng Wang", "post_date": "2022-08-04T07:10:54.769000", "votes": 11, "comment_count": 1, "views": 0, "content": "

I'm working on building a MONAI based pipeline for this challenge, and any progresses will be updated in this topic.
\nI don't have any experience in pathology image, thus I expect to learn a lot during the following serval weeks : )

\n

update 1:
\nI created a notebook that shows how to use MONAI transforms to load whole slide images:
\nhttps://www.kaggle.com/code/yiheng/load-wsi-and-extract-patches-with-monai/notebook
\na whole slide image will be divided into multiple patches.
\nIt can be used as the input of a Multiple Instance Learning (MIL) model, which expected the shape of:
\n[Batch, Number of patches, Channel, H, W]

\n

Update 2:
\n
\nI met several issues, such as data loading and poor model performance.
\nNow, I resized the training set into n * 4096 (the longest dimension is 4096), and the dataset has been uploaded:
\nhttps://www.kaggle.com/datasets/yiheng/mayo-longest-4096

\n

For reference, I refered to: https://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline and use pyvips to load the image first, and then resized into n * 4096. The function I used is:

\n
def save_resized_img(image_id):\n    image_pth = os.path.join(test_img_dir, image_id + \".tif\")\n    img = pyvips.Image.thumbnail(image_pth, 2048).numpy()\n    img = img.transpose([2, 0, 1])\n    print(img.shape)\n    np.save(os.path.join(resize_dir, f\"{image_id}.npy\"), img)\n    del img\n    gc.collect()\n
\n

To be updated:

\n

an open sourced model which can at least have a reasonable performance (not sure if I'm about to do it, felt very hard so far).

", "messages": [ { "id": 1883900, "postDate": "2022-08-04T07:10:54.770Z", "content": "

I'm working on building a MONAI based pipeline for this challenge, and any progresses will be updated in this topic.
\nI don't have any experience in pathology image, thus I expect to learn a lot during the following serval weeks : )

\n

update 1:
\nI created a notebook that shows how to use MONAI transforms to load whole slide images:
\nhttps://www.kaggle.com/code/yiheng/load-wsi-and-extract-patches-with-monai/notebook
\na whole slide image will be divided into multiple patches.
\nIt can be used as the input of a Multiple Instance Learning (MIL) model, which expected the shape of:
\n[Batch, Number of patches, Channel, H, W]

\n

Update 2:
\n
\nI met several issues, such as data loading and poor model performance.
\nNow, I resized the training set into n * 4096 (the longest dimension is 4096), and the dataset has been uploaded:
\nhttps://www.kaggle.com/datasets/yiheng/mayo-longest-4096

\n

For reference, I refered to: https://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline and use pyvips to load the image first, and then resized into n * 4096. The function I used is:

\n
def save_resized_img(image_id):\n    image_pth = os.path.join(test_img_dir, image_id + \".tif\")\n    img = pyvips.Image.thumbnail(image_pth, 2048).numpy()\n    img = img.transpose([2, 0, 1])\n    print(img.shape)\n    np.save(os.path.join(resize_dir, f\"{image_id}.npy\"), img)\n    del img\n    gc.collect()\n
\n

To be updated:

\n

an open sourced model which can at least have a reasonable performance (not sure if I'm about to do it, felt very hard so far).

", "rawMarkdown": "I'm working on building a MONAI based pipeline for this challenge, and any progresses will be updated in this topic.\nI don't have any experience in pathology image, thus I expect to learn a lot during the following serval weeks : )\n\nupdate 1:\nI created a notebook that shows how to use MONAI transforms to load whole slide images:\nhttps://www.kaggle.com/code/yiheng/load-wsi-and-extract-patches-with-monai/notebook\na whole slide image will be divided into multiple patches.\nIt can be used as the input of a Multiple Instance Learning (MIL) model, which expected the shape of:\n`[Batch, Number of patches, Channel, H, W]`\n\nUpdate 2:\n~~I will build a MILModel based training pipeline.~~\nI met several issues, such as data loading and poor model performance.\nNow, I resized the training set into `n * 4096` (the longest dimension is 4096), and the dataset has been uploaded:\nhttps://www.kaggle.com/datasets/yiheng/mayo-longest-4096\n\nFor reference, I refered to: https://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline and use `pyvips` to load the image first, and then resized into `n * 4096`. The function I used is:\n```\ndef save_resized_img(image_id):\n image_pth = os.path.join(test_img_dir, image_id + \".tif\")\n img = pyvips.Image.thumbnail(image_pth, 2048).numpy()\n img = img.transpose([2, 0, 1])\n print(img.shape)\n np.save(os.path.join(resize_dir, f\"{image_id}.npy\"), img)\n del img\n gc.collect()\n```\n\nTo be updated:\n\nan open sourced model which can at least have a reasonable performance (not sure if I'm about to do it, felt very hard so far).", "votes": 11 }, { "id": 1884140, "postDate": "2022-08-04T10:16:44.977Z", "content": "

Kudos to your willingness to learn and develop your skills for the assignment! Good luck!

", "rawMarkdown": "Kudos to your willingness to learn and develop your skills for the assignment! Good luck!" } ], "comments": [ { "id": 1884140, "author_name": "Ravi Ramakrishnan", "author_url": "", "post_date": "2022-08-04T10:16:44.977000", "content": "

Kudos to your willingness to learn and develop your skills for the assignment! Good luck!

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1883900": "I'm working on building a MONAI based pipeline for this challenge, and any progresses will be updated in this topic.\nI don't have any experience in pathology image, thus I expect to learn a lot during the following serval weeks : )\n\nupdate 1:\nI created a notebook that shows how to use MONAI transforms to load whole slide images:\nhttps://www.kaggle.com/code/yiheng/load-wsi-and-extract-patches-with-monai/notebook\na whole slide image will be divided into multiple patches.\nIt can be used as the input of a Multiple Instance Learning (MIL) model, which expected the shape of:\n`[Batch, Number of patches, Channel, H, W]`\n\nUpdate 2:\n~~I will build a MILModel based training pipeline.~~\nI met several issues, such as data loading and poor model performance.\nNow, I resized the training set into `n * 4096` (the longest dimension is 4096), and the dataset has been uploaded:\nhttps://www.kaggle.com/datasets/yiheng/mayo-longest-4096\n\nFor reference, I refered to: https://www.kaggle.com/code/analokamus/how-to-use-pyvips-offline and use `pyvips` to load the image first, and then resized into `n * 4096`. The function I used is:\n```\ndef save_resized_img(image_id):\n image_pth = os.path.join(test_img_dir, image_id + \".tif\")\n img = pyvips.Image.thumbnail(image_pth, 2048).numpy()\n img = img.transpose([2, 0, 1])\n print(img.shape)\n np.save(os.path.join(resize_dir, f\"{image_id}.npy\"), img)\n del img\n gc.collect()\n```\n\nTo be updated:\n\nan open sourced model which can at least have a reasonable performance (not sure if I'm about to do it, felt very hard so far).", "1884140": "Kudos to your willingness to learn and develop your skills for the assignment! Good luck!" } }