![\"ChatGPT](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5918909%2F1fbbeb411e09646b6fdea608178f1f5a%2Fchatgpt2.PNG?generation=1671393767617156&alt=media\")
Several data labeled abnormalities as benign and malignant.
\nFor example in CBIS-DDSM,you have benign/malignant mass and benign/malignant clarification.
We have the following options to relabel external data using kaggle ground truth definition:
\nAnyone has experiences or results on these?
\nNote: there are other way to use external data that don't require ones to change external label. e.g if we use as external label as aux loss, or used feature from model trained via external data (instead of using external data itself)
", "messages": [ { "id": 2069132, "postDate": "2022-12-18T16:25:37.850Z", "content": "Several data labeled abnormalities as benign and malignant.
\nFor example in CBIS-DDSM,you have benign/malignant mass and benign/malignant clarification.
We have the following options to relabel external data using kaggle ground truth definition:
\nAnyone has experiences or results on these?
\nNote: there are other way to use external data that don't require ones to change external label. e.g if we use as external label as aux loss, or used feature from model trained via external data (instead of using external data itself)
", "rawMarkdown": "Several data labeled abnormalities as benign and malignant. \nFor example in CBIS-DDSM,you have benign/malignant mass and benign/malignant clarification.\n\nWe have the following options to relabel external data using kaggle ground truth definition:\n1. benign = non cancer, malignant= cancer\n2. benign = cancer, malignant= cancer (we are willing to accept false positive if this improves recall)\n3. benign = 0.5, malignant= 1.0 (use soft label instead of discrete label)\n4. don't use benign images at all (don't care)\n5. benign = pseudo label, malignant= 1.0\n6. benign = pseudo label, malignant= pseudo label (ignore samples if pseudo label are weak)\n\nAnyone has experiences or results on these? \n\nNote: there are other way to use external data that don't require ones to change external label. e.g if we use as external label as aux loss, or used feature from model trained via external data (instead of using external data itself)", "votes": 14 }, { "id": 2070205, "postDate": "2022-12-19T17:40:37.727Z", "content": "I believe #1 is correct, but I'll reach out to the host team for confirmation.
", "rawMarkdown": "I believe #1 is correct, but I'll reach out to the host team for confirmation.", "votes": 4, "replies": [ { "id": 2070226, "postDate": "2022-12-19T18:01:53.460Z", "content": "The host team confirmed that benign masses were labeled as negative for cancer.
", "rawMarkdown": "The host team confirmed that benign masses were labeled as negative for cancer.", "votes": 7, "replies": [ { "id": 2070375, "postDate": "2022-12-19T23:57:43.537Z", "content": "Thanks a lot!
", "rawMarkdown": "Thanks a lot!" }, { "id": 2070799, "postDate": "2022-12-20T10:54:16.363Z", "content": "The term \"cancer\" is defined as a malignant neoplasm, right? In other words, \"cancer\" always refers to malignancy, not benignity. The term \"tumor\" can be benign or malignant.
\nHowever, as for labels, benign tumors can sometimes resemble malignant tumors in imaging findings, so it may be effective to use a soft label for benign tumors to train models.
", "rawMarkdown": "The term \"cancer\" is defined as a malignant neoplasm, right? In other words, \"cancer\" always refers to malignancy, not benignity. The term \"tumor\" can be benign or malignant.\n\nHowever, as for labels, benign tumors can sometimes resemble malignant tumors in imaging findings, so it may be effective to use a soft label for benign tumors to train models." }, { "id": 2071134, "postDate": "2022-12-20T17:19:45.160Z", "content": "Some questions are -
\nI would guess that the first option would work best. Benign masses and calcifications are a separate category in BIRADS and we have the BIRADS scores in this competition. But in addition, we also have to consider normal mammograms without masses.
\nLet's see how ChatGPT responds:
\n
(I love your posts and especially ones with ChatGPT so I had to try it out)
", "rawMarkdown": "I would guess that the first option would work best. Benign masses and calcifications are a separate category in BIRADS and we have the BIRADS scores in this competition. But in addition, we also have to consider normal mammograms without masses. \n\nLet's see how ChatGPT responds:\n\n\n\n(I love your posts and especially ones with ChatGPT so I had to try it out)", "votes": 2, "replies": [ { "id": 2069423, "postDate": "2022-12-19T01:02:26.697Z", "content": "conclusion: ChatGPT is the winner
", "rawMarkdown": "conclusion: ChatGPT is the winner", "votes": 3 } ] }, { "id": 2069419, "postDate": "2022-12-19T00:45:09.297Z", "content": "Great question. Would really like to know what benign/cancer/invasive/malignant mean in the various different contexts.
", "rawMarkdown": "Great question. Would really like to know what benign/cancer/invasive/malignant mean in the various different contexts." }, { "id": 2069213, "postDate": "2022-12-18T18:15:29.773Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 2070205, "author_name": "Sohier Dane", "author_url": "", "post_date": "2022-12-19T17:40:37.727000", "content": "I believe #1 is correct, but I'll reach out to the host team for confirmation.
", "votes": 4, "replies": [ { "id": 2070226, "author_name": "Sohier Dane", "author_url": "", "post_date": "2022-12-19T18:01:53.460000", "content": "The host team confirmed that benign masses were labeled as negative for cancer.
", "votes": 7, "replies": [ { "id": 2070375, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-19T23:57:43.537000", "content": "Thanks a lot!
", "votes": 0, "replies": [] }, { "id": 2070799, "author_name": "YYama", "author_url": "", "post_date": "2022-12-20T10:54:16.363000", "content": "The term \"cancer\" is defined as a malignant neoplasm, right? In other words, \"cancer\" always refers to malignancy, not benignity. The term \"tumor\" can be benign or malignant.
\nHowever, as for labels, benign tumors can sometimes resemble malignant tumors in imaging findings, so it may be effective to use a soft label for benign tumors to train models.
", "votes": 0, "replies": [] }, { "id": 2071134, "author_name": "@kaggleqrdl", "author_url": "", "post_date": "2022-12-20T17:19:45.160000", "content": "Some questions are -
\nI would guess that the first option would work best. Benign masses and calcifications are a separate category in BIRADS and we have the BIRADS scores in this competition. But in addition, we also have to consider normal mammograms without masses.
\nLet's see how ChatGPT responds:
\n
(I love your posts and especially ones with ChatGPT so I had to try it out)
", "votes": 2, "replies": [ { "id": 2069423, "author_name": "hengck23", "author_url": "", "post_date": "2022-12-19T01:02:26.697000", "content": "conclusion: ChatGPT is the winner
", "votes": 3, "replies": [] } ] }, { "id": 2069419, "author_name": "@kaggleqrdl", "author_url": "", "post_date": "2022-12-19T00:45:09.297000", "content": "Great question. Would really like to know what benign/cancer/invasive/malignant mean in the various different contexts.
", "votes": 0, "replies": [] }, { "id": 2069213, "author_name": "", "author_url": "", "post_date": "2022-12-18T18:15:29.773000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2069132": "Several data labeled abnormalities as benign and malignant. \nFor example in CBIS-DDSM,you have benign/malignant mass and benign/malignant clarification.\n\nWe have the following options to relabel external data using kaggle ground truth definition:\n1. benign = non cancer, malignant= cancer\n2. benign = cancer, malignant= cancer (we are willing to accept false positive if this improves recall)\n3. benign = 0.5, malignant= 1.0 (use soft label instead of discrete label)\n4. don't use benign images at all (don't care)\n5. benign = pseudo label, malignant= 1.0\n6. benign = pseudo label, malignant= pseudo label (ignore samples if pseudo label are weak)\n\nAnyone has experiences or results on these? \n\nNote: there are other way to use external data that don't require ones to change external label. e.g if we use as external label as aux loss, or used feature from model trained via external data (instead of using external data itself)", "2070205": "I believe #1 is correct, but I'll reach out to the host team for confirmation.", "2069276": "I would guess that the first option would work best. Benign masses and calcifications are a separate category in BIRADS and we have the BIRADS scores in this competition. But in addition, we also have to consider normal mammograms without masses. \n\nLet's see how ChatGPT responds:\n\n\n\n(I love your posts and especially ones with ChatGPT so I had to try it out)", "2069419": "Great question. Would really like to know what benign/cancer/invasive/malignant mean in the various different contexts.", "2069213": "" } }