![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F337700%2Fffaf06c0f3482dbfd9cca5d1e1317c76%2Fprostate_4658.png?generation=1656924870176628&alt=media\")
In preprocessing the data I noticed that there are often large empty spaces labeled as foreground. They happen most often in prostate, lung images and sometimes in intestine as well.
\nIn my own preprocessing I am trying to discard those labels at empty pixel locations (colored blue) since they probably will confuse the model. But I am not sure if this would be the best strategy for doing inference on test data.
\nHas anyone tried any strategy for these \"empty\" labels?
\n
In preprocessing the data I noticed that there are often large empty spaces labeled as foreground. They happen most often in prostate, lung images and sometimes in intestine as well.
\nIn my own preprocessing I am trying to discard those labels at empty pixel locations (colored blue) since they probably will confuse the model. But I am not sure if this would be the best strategy for doing inference on test data.
\nHas anyone tried any strategy for these \"empty\" labels?
\n
In the functional tissue units, these empty spaces are the glands or the actual alveoli in the lung. Their secrete is washed away when staining.
\nWhile it is a good idea to remove these pixels from the annotation, you will probably see a decrease in validation accuracy, since they will probably be labeled as positive in the validation/test data as well.
You might get decent/better results when training on the actual glandular epithelial cells (this is what you are proposing) and filling the masks again in postprocessing?
\nIt could also introduce more errors.
Thank you for the great insights! It looks like it is worthwhile to experiment with different strategies, including treating the empty spaces as a separate label.
", "rawMarkdown": "Thank you for the great insights! It looks like it is worthwhile to experiment with different strategies, including treating the empty spaces as a separate label.", "votes": null }, { "id": "1851389", "postDate": "07/11/2022 08:00:54", "content": "Preds on the prostrate class appear to be generating a lot of false positives - two examples are shown in the image below. On these two images, could it be that the model preds are in fact correct and the problem lies with the labels? Thanks.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1086574%2Fbb71b83d921ccd37cb818d206cf7884f%2Fprostate-false-positives.png?generation=1657525338163741&alt=media\")
I do believe, that in some cases the annotations are only done, if the gland is completely surrounded by tissue.
\nThis would also be the case for your example.
\nThere have been similar issues in the last year's hubmap challenge.
Thank you for answering @theudas. Much appreciated.
", "rawMarkdown": "Thank you for answering @theudas. Much appreciated.", "votes": null } ], "comments": [ { "id": 1842909, "author_name": "theudas", "author_url": "", "post_date": "07/04/2022 11:41:33", "content": "In the functional tissue units, these empty spaces are the glands or the actual alveoli in the lung. Their secrete is washed away when staining.
\nWhile it is a good idea to remove these pixels from the annotation, you will probably see a decrease in validation accuracy, since they will probably be labeled as positive in the validation/test data as well.
You might get decent/better results when training on the actual glandular epithelial cells (this is what you are proposing) and filling the masks again in postprocessing?
\nIt could also introduce more errors.
Thank you for the great insights! It looks like it is worthwhile to experiment with different strategies, including treating the empty spaces as a separate label.
", "votes": null, "replies": [] }, { "id": 1851389, "author_name": "vbookshelf", "author_url": "", "post_date": "07/11/2022 08:00:54", "content": "Preds on the prostrate class appear to be generating a lot of false positives - two examples are shown in the image below. On these two images, could it be that the model preds are in fact correct and the problem lies with the labels? Thanks.
\n
I do believe, that in some cases the annotations are only done, if the gland is completely surrounded by tissue.
\nThis would also be the case for your example.
\nThere have been similar issues in the last year's hubmap challenge.
Thank you for answering @theudas. Much appreciated.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1842760": "In preprocessing the data I noticed that there are often large empty spaces labeled as foreground. They happen most often in prostate, lung images and sometimes in intestine as well. \n\nIn my own preprocessing I am trying to discard those labels at empty pixel locations (colored blue) since they probably will confuse the model. But I am not sure if this would be the best strategy for doing inference on test data. \n\nHas anyone tried any strategy for these \"empty\" labels? \n\n", "1842909": "In the functional tissue units, these empty spaces are the glands or the actual alveoli in the lung. Their secrete is washed away when staining.\nWhile it is a good idea to remove these pixels from the annotation, you will probably see a decrease in validation accuracy, since they will probably be labeled as positive in the validation/test data as well.\n\nYou might get decent/better results when training on the actual glandular epithelial cells (this is what you are proposing) and filling the masks again in postprocessing?\nIt could also introduce more errors.", "1844232": "Thank you for the great insights! It looks like it is worthwhile to experiment with different strategies, including treating the empty spaces as a separate label.", "1851389": "Preds on the prostrate class appear to be generating a lot of false positives - two examples are shown in the image below. On these two images, could it be that the model preds are in fact correct and the problem lies with the labels? Thanks.\n\n", "1851536": "I do believe, that in some cases the annotations are only done, if the gland is completely surrounded by tissue.\nThis would also be the case for your example.\nThere have been similar issues in the last year's hubmap challenge.", "1851690": "Thank you for answering @theudas. Much appreciated." }, "source": "meta" }