{ "id": 338869, "title": "H&E normalization with torchstain", "url": "/competitions/hubmap-organ-segmentation/discussion/338869", "author_name": "Carlo Alberto", "post_date": "2022-07-22T11:05:07.506000", "votes": 11, "comment_count": 3, "views": 0, "content": "

Hi all,

\n

Given this competition is using histological images, you might find useful the torchstain package for normalizing H&E-stained images https://github.com/EIDOSLAB/torchstain.

\n

It supports the Mackenko [1] normalization algorithm and it is implemented in pytorch, so you can use it directly as image transform with torchvision, e.g.:

\n

transforms.Lambda(lambda x: normalizer.normalize(I=x, stains=False)[0] .

\n

Support for tensorflow is still in beta (you can test the library by installing it directly from github)

\n

Good luck to everyone!

\n

[1] Macenko, Marc, et al. \"A method for normalizing histology slides for quantitative analysis.\" 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE, 2009.

", "messages": [ { "id": 1866241, "postDate": "2022-07-22T11:05:07.507Z", "content": "

Hi all,

\n

Given this competition is using histological images, you might find useful the torchstain package for normalizing H&E-stained images https://github.com/EIDOSLAB/torchstain.

\n

It supports the Mackenko [1] normalization algorithm and it is implemented in pytorch, so you can use it directly as image transform with torchvision, e.g.:

\n

transforms.Lambda(lambda x: normalizer.normalize(I=x, stains=False)[0] .

\n

Support for tensorflow is still in beta (you can test the library by installing it directly from github)

\n

Good luck to everyone!

\n

[1] Macenko, Marc, et al. \"A method for normalizing histology slides for quantitative analysis.\" 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE, 2009.

", "rawMarkdown": "Hi all, \n\nGiven this competition is using histological images, you might find useful the torchstain package for normalizing H&E-stained images https://github.com/EIDOSLAB/torchstain.\n\nIt supports the Mackenko [1] normalization algorithm and it is implemented in pytorch, so you can use it directly as image transform with torchvision, e.g.:\n\n` transforms.Lambda(lambda x: normalizer.normalize(I=x, stains=False)[0]` .\n\nSupport for tensorflow is still in beta (you can test the library by installing it directly from github)\n\nGood luck to everyone!\n\n[1] Macenko, Marc, et al. \"A method for normalizing histology slides for quantitative analysis.\" *2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro*. IEEE, 2009.", "votes": 11 }, { "id": 1872806, "postDate": "2022-07-27T08:51:20.910Z", "content": "

I did add this normalization method to my data pipeline, but my CV did not improve.
\nDid you achieve a performance improvement when adding this normalization method?

", "rawMarkdown": "I did add this normalization method to my data pipeline, but my CV did not improve.\nDid you achieve a performance improvement when adding this normalization method?", "votes": 1, "replies": [ { "id": 1872809, "postDate": "2022-07-27T08:54:23.793Z", "content": "

I still haven't tested it, but the results might depend also on the target image you choose. Did you try with different images?
\nA useful case is if you pretrain a model on another dataset, then it might be beetter to normalize the images to look similar to the pretrain data

", "rawMarkdown": "I still haven't tested it, but the results might depend also on the target image you choose. Did you try with different images? \nA useful case is if you pretrain a model on another dataset, then it might be beetter to normalize the images to look similar to the pretrain data" } ] }, { "id": 1874765, "postDate": "2022-07-28T13:33:37.907Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 1872806, "author_name": "Mark Wijkhuizen", "author_url": "", "post_date": "2022-07-27T08:51:20.910000", "content": "

I did add this normalization method to my data pipeline, but my CV did not improve.
\nDid you achieve a performance improvement when adding this normalization method?

", "votes": 1, "replies": [ { "id": 1872809, "author_name": "Carlo Alberto", "author_url": "", "post_date": "2022-07-27T08:54:23.793000", "content": "

I still haven't tested it, but the results might depend also on the target image you choose. Did you try with different images?
\nA useful case is if you pretrain a model on another dataset, then it might be beetter to normalize the images to look similar to the pretrain data

", "votes": 0, "replies": [] } ] }, { "id": 1874765, "author_name": "", "author_url": "", "post_date": "2022-07-28T13:33:37.907000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1866241": "Hi all, \n\nGiven this competition is using histological images, you might find useful the torchstain package for normalizing H&E-stained images https://github.com/EIDOSLAB/torchstain.\n\nIt supports the Mackenko [1] normalization algorithm and it is implemented in pytorch, so you can use it directly as image transform with torchvision, e.g.:\n\n` transforms.Lambda(lambda x: normalizer.normalize(I=x, stains=False)[0]` .\n\nSupport for tensorflow is still in beta (you can test the library by installing it directly from github)\n\nGood luck to everyone!\n\n[1] Macenko, Marc, et al. \"A method for normalizing histology slides for quantitative analysis.\" *2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro*. IEEE, 2009.", "1872806": "I did add this normalization method to my data pipeline, but my CV did not improve.\nDid you achieve a performance improvement when adding this normalization method?", "1874765": "" } }