I tried to feed pairs of images of patients to a MIL model. Since some of the patients have more than 2 images per breast, I created combinations of image pairs from them. Regardless of fusing type; avg, max, concat, attention, I couldn't get any good result using this approach. I got 0.75 AUC at best. Does anyone able make this approach work?
", "messages": [ { "id": 2168560, "postDate": "2023-03-04T10:20:46.737Z", "content": "Hi @gunesevitan , we applied MIL and got good score on both CV, public LB, and private LB.
\nPlease check
\nhttps://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/391779
Did you take into account view and laterality? And if there are multiple L-CC, R-CC, L-MLO and R-MLO images you could choose the latest (if there is a timestamp). I guess you have read already that L-CC and L-MLO match each other only partially and same for R-CC and R-MLO even though they depict the same breast. It is common that the lesions are not visible in both of the projections--perhaps neither or any of them.
", "rawMarkdown": "Did you take into account [view and laterality](https://www.kaggle.com/code/anttiisosalo/solt-based-image-augs-rsna-bc-detection?scriptVersionId=118634331&cellId=13)? And if there are multiple L-CC, R-CC, L-MLO and R-MLO images you could choose the latest (if there is a timestamp). I guess you have read already that L-CC and L-MLO match each other only partially and same for R-CC and R-MLO even though they depict the same breast. It is common that the lesions are not visible in both of the projections--perhaps neither or any of them.", "votes": 1, "replies": [ { "id": 2139730, "postDate": "2023-02-11T05:07:14.983Z", "content": "I wanted to model single breast images as they have a single label. For example; if there are 4 left breast images for a patient then I was creating 6 combinations of 2 from them and all of their labels were same. I didn't expect my score would get worse than per image approach.
", "rawMarkdown": "I wanted to model single breast images as they have a single label. For example; if there are 4 left breast images for a patient then I was creating 6 combinations of 2 from them and all of their labels were same. I didn't expect my score would get worse than per image approach.", "votes": 1 } ] }, { "id": 2139664, "postDate": "2023-02-11T01:45:40.010Z", "content": "I tried feeding per patient images in channels but AUC wasn't better than 80 ish.
", "rawMarkdown": "I tried feeding per patient images in channels but AUC wasn't better than 80 ish. ", "votes": 1 }, { "id": 2138198, "postDate": "2023-02-10T16:23:46.327Z", "content": "Did you do EDA? Otherwise, imbalanced dataset affects your results.
", "rawMarkdown": "Did you do EDA? Otherwise, imbalanced dataset affects your results.", "votes": 1, "replies": [ { "id": 2139727, "postDate": "2023-02-11T05:01:23.930Z", "content": "I was aggregating labels on patient, laterality groups so imbalance reduced but score got worse.
", "rawMarkdown": "I was aggregating labels on patient, laterality groups so imbalance reduced but score got worse." } ] }, { "id": 2138151, "postDate": "2023-02-10T15:45:29.423Z", "content": "I tried to feed pairs of images of patients to a MIL model. Since some of the patients have more than 2 images per breast, I created combinations of image pairs from them. Regardless of fusing type; avg, max, concat, attention, I couldn't get any good result using this approach. I got 0.75 AUC at best. Does anyone able make this approach work?
", "rawMarkdown": "I tried to feed pairs of images of patients to a MIL model. Since some of the patients have more than 2 images per breast, I created combinations of image pairs from them. Regardless of fusing type; avg, max, concat, attention, I couldn't get any good result using this approach. I got 0.75 AUC at best. Does anyone able make this approach work?", "votes": 2 } ], "comments": [ { "id": 2168560, "author_name": "ForcewithMe", "author_url": "", "post_date": "2023-03-04T10:20:46.737000", "content": "Hi @gunesevitan , we applied MIL and got good score on both CV, public LB, and private LB.
\nPlease check
\nhttps://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/391779
Did you take into account view and laterality? And if there are multiple L-CC, R-CC, L-MLO and R-MLO images you could choose the latest (if there is a timestamp). I guess you have read already that L-CC and L-MLO match each other only partially and same for R-CC and R-MLO even though they depict the same breast. It is common that the lesions are not visible in both of the projections--perhaps neither or any of them.
", "votes": 1, "replies": [ { "id": 2139730, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-02-11T05:07:14.983000", "content": "I wanted to model single breast images as they have a single label. For example; if there are 4 left breast images for a patient then I was creating 6 combinations of 2 from them and all of their labels were same. I didn't expect my score would get worse than per image approach.
", "votes": 1, "replies": [] } ] }, { "id": 2139664, "author_name": "Phaedrus", "author_url": "", "post_date": "2023-02-11T01:45:40.010000", "content": "I tried feeding per patient images in channels but AUC wasn't better than 80 ish.
", "votes": 1, "replies": [] }, { "id": 2138198, "author_name": "Hyunsoo Lee 1010", "author_url": "", "post_date": "2023-02-10T16:23:46.327000", "content": "Did you do EDA? Otherwise, imbalanced dataset affects your results.
", "votes": 1, "replies": [ { "id": 2139727, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-02-11T05:01:23.930000", "content": "I was aggregating labels on patient, laterality groups so imbalance reduced but score got worse.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2168560": "Hi @gunesevitan , we applied MIL and got good score on both CV, public LB, and private LB.\nPlease check\nhttps://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/391779", "2139679": "Did you take into account [view and laterality](https://www.kaggle.com/code/anttiisosalo/solt-based-image-augs-rsna-bc-detection?scriptVersionId=118634331&cellId=13)? And if there are multiple L-CC, R-CC, L-MLO and R-MLO images you could choose the latest (if there is a timestamp). I guess you have read already that L-CC and L-MLO match each other only partially and same for R-CC and R-MLO even though they depict the same breast. It is common that the lesions are not visible in both of the projections--perhaps neither or any of them.", "2139664": "I tried feeding per patient images in channels but AUC wasn't better than 80 ish. ", "2138198": "Did you do EDA? Otherwise, imbalanced dataset affects your results.", "2138151": "I tried to feed pairs of images of patients to a MIL model. Since some of the patients have more than 2 images per breast, I created combinations of image pairs from them. Regardless of fusing type; avg, max, concat, attention, I couldn't get any good result using this approach. I got 0.75 AUC at best. Does anyone able make this approach work?" } }