![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13477216%2F6955423c6ddae6e37d75a22e8dd236f1%2F7002.png?generation=1677379377311828&alt=media\")
👋 Greetings,
\nI've been attempting to make and train a multi-view model to directly predict labels for L and R from 16-bit mammograms.
\nThe model often (at least in 20% of the cases) outputs nan's as predictions for both or one of the lateralities. (I've concatenated the output from the two branches and the positive predictions are at places 0 and 2.)
\n
In this particular case the default windowing / VOI LUT does not provide the best outcome.
\nAnother thing might be that I have standardized each image individually to have zero mean and unit standard deviation. This has been the safest way with mammograms in my earlier studies. Perhaps I should reconsider?
\nWhat I could have also done would have been to train separate models for the two TransferSyntaxUID's, but it might be too late for that.
\nIt seems that multi-view models in general are challenging to train even though here when I flip the right hand side images to face right and use flip augmentation it is rather multi-projection model, where the differences between MLO and CC might do some harm for the training.
\nEDIT: nan's were due to improper use of amp. 😅 Let's see if I can get a better Late Submission score.
Best,
\nAntti
", "messages": [ { "id": 2159766, "postDate": "2023-02-26T03:17:15.550Z", "content": "👋 Greetings,
\nI've been attempting to make and train a multi-view model to directly predict labels for L and R from 16-bit mammograms.
\nThe model often (at least in 20% of the cases) outputs nan's as predictions for both or one of the lateralities. (I've concatenated the output from the two branches and the positive predictions are at places 0 and 2.)
\n
In this particular case the default windowing / VOI LUT does not provide the best outcome.
\nAnother thing might be that I have standardized each image individually to have zero mean and unit standard deviation. This has been the safest way with mammograms in my earlier studies. Perhaps I should reconsider?
\nWhat I could have also done would have been to train separate models for the two TransferSyntaxUID's, but it might be too late for that.
\nIt seems that multi-view models in general are challenging to train even though here when I flip the right hand side images to face right and use flip augmentation it is rather multi-projection model, where the differences between MLO and CC might do some harm for the training.
\nEDIT: nan's were due to improper use of amp. 😅 Let's see if I can get a better Late Submission score.
Best,
\nAntti
", "rawMarkdown": "👋 Greetings,\n\nI've been attempting to make and train a multi-view model to directly predict labels for L and R from 16-bit mammograms.\n\nThe model often (at least in 20% of the cases) outputs nan's as predictions for both or one of the lateralities. (I've concatenated the output from the two branches and the positive predictions are at places 0 and 2.)\n\n\n\nIn this particular case the default windowing / VOI LUT does not provide the best outcome.\n\nAnother thing might be that I have standardized each image individually to have zero mean and unit standard deviation. This has been the safest way with mammograms in my earlier studies. Perhaps I should reconsider?\n\nWhat I could have also done would have been to train separate models for the two TransferSyntaxUID's, but it might be too late for that.\n\nIt seems that multi-view models in general are challenging to train even though here when I flip the right hand side images to face right and use flip augmentation it is rather multi-projection model, where the differences between MLO and CC might do some harm for the training.\n\nEDIT: nan's were due to improper use of `amp`. 😅 Let's see if I can get a better Late Submission score.\n\nBest,\n\nAntti " } ], "comments": [], "raw_markdown_by_id": { "2159766": "👋 Greetings,\n\nI've been attempting to make and train a multi-view model to directly predict labels for L and R from 16-bit mammograms.\n\nThe model often (at least in 20% of the cases) outputs nan's as predictions for both or one of the lateralities. (I've concatenated the output from the two branches and the positive predictions are at places 0 and 2.)\n\n\n\nIn this particular case the default windowing / VOI LUT does not provide the best outcome.\n\nAnother thing might be that I have standardized each image individually to have zero mean and unit standard deviation. This has been the safest way with mammograms in my earlier studies. Perhaps I should reconsider?\n\nWhat I could have also done would have been to train separate models for the two TransferSyntaxUID's, but it might be too late for that.\n\nIt seems that multi-view models in general are challenging to train even though here when I flip the right hand side images to face right and use flip augmentation it is rather multi-projection model, where the differences between MLO and CC might do some harm for the training.\n\nEDIT: nan's were due to improper use of `amp`. 😅 Let's see if I can get a better Late Submission score.\n\nBest,\n\nAntti " } }