Hi,\nOne can consider full white tiles and pen marks as a kind of cutout method by analogy to data augmentation for object detectors. At least I tried to remove pen marks but it did not improve my score. \nBut what about tiles that have more than 30/40% white pixels but contain useful tissue portions ? I'm thinking on filling the blank area with some tissue of other tiles (same slide or not) like a kind of mixup or cutmix method. I'm wondering if some of you have already tried such approach ?
", "messages": [ { "id": 870598, "postDate": "2020-06-01T18:59:20.810Z", "content": "@iafoss dataset is not the best way, I did develop the best way (according to me) but takes days to process the whole dataset, it takes perfect segments from an image, now the reason why that dataset is not near perfect is because:-
\n\nThere are white tiles left up even when we only take 16 pieces as all of the data is not covered, a few examples of this can be due to the problem number 2...
The image size of every image is different which makes big part of the data not reliable for model to learn from after extracting patches from a particular size, however solving this one is a very hard one, I did develop some algos but did not get the success I wanted.
Solving these problems should do your work.
\n\nEDIT: I have got a new algorithm to completely solve problem number 2 however, it still takes days, I think I will have to optimize my algorithms AF.
", "rawMarkdown": "@iafoss dataset is not the best way, I did develop the best way (according to me) but takes days to process the whole dataset, it takes perfect segments from an image, now the reason why that dataset is not near perfect is because:-\n\n1. There are white tiles left up even when we only take 16 pieces as all of the data is not covered, a few examples of this can be due to the problem number 2...\n\n2. The image size of every image is different which makes big part of the data not reliable for model to learn from after extracting patches from a particular size, however solving this one is a very hard one, I did develop some algos but did not get the success I wanted.\n\nSolving these problems should do your work.\n\nEDIT: I have got a new algorithm to completely solve problem number 2 however, it still takes days, I think I will have to optimize my algorithms AF.", "votes": 1 }, { "id": 870292, "postDate": "2020-06-01T16:00:09.980Z", "content": "Hi,\nOne can consider full white tiles and pen marks as a kind of cutout method by analogy to data augmentation for object detectors. At least I tried to remove pen marks but it did not improve my score. \nBut what about tiles that have more than 30/40% white pixels but contain useful tissue portions ? I'm thinking on filling the blank area with some tissue of other tiles (same slide or not) like a kind of mixup or cutmix method. I'm wondering if some of you have already tried such approach ?
", "rawMarkdown": "Hi,\nOne can consider full white tiles and pen marks as a kind of cutout method by analogy to data augmentation for object detectors. At least I tried to remove pen marks but it did not improve my score. \nBut what about tiles that have more than 30/40% white pixels but contain useful tissue portions ? I'm thinking on filling the blank area with some tissue of other tiles (same slide or not) like a kind of mixup or cutmix method. I'm wondering if some of you have already tried such approach ?" } ], "comments": [ { "id": 870598, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-06-01T18:59:20.810000", "content": "@iafoss dataset is not the best way, I did develop the best way (according to me) but takes days to process the whole dataset, it takes perfect segments from an image, now the reason why that dataset is not near perfect is because:-
\n\nThere are white tiles left up even when we only take 16 pieces as all of the data is not covered, a few examples of this can be due to the problem number 2...
The image size of every image is different which makes big part of the data not reliable for model to learn from after extracting patches from a particular size, however solving this one is a very hard one, I did develop some algos but did not get the success I wanted.
Solving these problems should do your work.
\n\nEDIT: I have got a new algorithm to completely solve problem number 2 however, it still takes days, I think I will have to optimize my algorithms AF.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "870598": "@iafoss dataset is not the best way, I did develop the best way (according to me) but takes days to process the whole dataset, it takes perfect segments from an image, now the reason why that dataset is not near perfect is because:-\n\n1. There are white tiles left up even when we only take 16 pieces as all of the data is not covered, a few examples of this can be due to the problem number 2...\n\n2. The image size of every image is different which makes big part of the data not reliable for model to learn from after extracting patches from a particular size, however solving this one is a very hard one, I did develop some algos but did not get the success I wanted.\n\nSolving these problems should do your work.\n\nEDIT: I have got a new algorithm to completely solve problem number 2 however, it still takes days, I think I will have to optimize my algorithms AF.", "870292": "Hi,\nOne can consider full white tiles and pen marks as a kind of cutout method by analogy to data augmentation for object detectors. At least I tried to remove pen marks but it did not improve my score. \nBut what about tiles that have more than 30/40% white pixels but contain useful tissue portions ? I'm thinking on filling the blank area with some tissue of other tiles (same slide or not) like a kind of mixup or cutmix method. I'm wondering if some of you have already tried such approach ?" } }