Many similar abnormality detection tasks in chest x-rays label their data by brushing the abnormalities to create segmentation maps. I was wondering why this competition uses a bounding box as the labels as segmentation maps seem more precise… Is it that there are too many classes to be a segmentation problem? Or is the peripheral information of the abnormality important somehow?
", "messages": [ { "id": 1151456, "postDate": "2021-01-13T10:36:55.030Z", "content": "Many similar abnormality detection tasks in chest x-rays label their data by brushing the abnormalities to create segmentation maps. I was wondering why this competition uses a bounding box as the labels as segmentation maps seem more precise… Is it that there are too many classes to be a segmentation problem? Or is the peripheral information of the abnormality important somehow?
", "rawMarkdown": "Many similar abnormality detection tasks in chest x-rays label their data by brushing the abnormalities to create segmentation maps. I was wondering why this competition uses a bounding box as the labels as segmentation maps seem more precise... Is it that there are too many classes to be a segmentation problem? Or is the peripheral information of the abnormality important somehow?", "votes": 10 }, { "id": 1155395, "postDate": "2021-01-16T11:26:51.383Z", "content": "It's a good question, but I think the answer is relatively straightforward. As someone who looks at CXRs most days, I can confidently say the peripheral information encoded in the bounding box is generally not useful. CXRs are a relatively blunt instrument for diagnosis, and complex findings will generally lead a radiologist to suggest progression to CT. So the bounding box just detects the problem and highlights it for the clinician to review. Segmentation maps might e.g. be useful in depicting growth of a tumour in a patient over time, but this is beyond the scope of this type of challenge.
", "rawMarkdown": "It's a good question, but I think the answer is relatively straightforward. As someone who looks at CXRs most days, I can confidently say the peripheral information encoded in the bounding box is generally not useful. CXRs are a relatively blunt instrument for diagnosis, and complex findings will generally lead a radiologist to suggest progression to CT. So the bounding box just detects the problem and highlights it for the clinician to review. Segmentation maps might e.g. be useful in depicting growth of a tumour in a patient over time, but this is beyond the scope of this type of challenge.", "votes": 8, "replies": [ { "id": 1155594, "postDate": "2021-01-16T13:44:20.070Z", "content": "Very interesting comment! So since you do a follow-up CT scan if there is an abnormality anyway, you don't need that much precision with x-rays. In terms of annotation, you can always convert from segmentation to bounding box but I do think it is way harder and time-consuming to label with segmentation brushes. So the reason could be that the cost of labelling segmentation is high and that 'precision' you get from the segmentation maps was not required for meeting the clinical objective of the challenge.
", "rawMarkdown": "Very interesting comment! So since you do a follow-up CT scan if there is an abnormality anyway, you don't need that much precision with x-rays. In terms of annotation, you can always convert from segmentation to bounding box but I do think it is way harder and time-consuming to label with segmentation brushes. So the reason could be that the cost of labelling segmentation is high and that 'precision' you get from the segmentation maps was not required for meeting the clinical objective of the challenge. ", "votes": 3 }, { "id": 1156035, "postDate": "2021-01-16T21:47:25.970Z", "content": "Yes, I think so. I had thought segmentation per pixel would be more compute expensive, so that's another reason why
", "rawMarkdown": "Yes, I think so. I had thought segmentation per pixel would be more compute expensive, so that's another reason why ", "votes": 2 } ] }, { "id": 1157601, "postDate": "2021-01-18T03:28:47.910Z", "content": "because segmentation annotation is expensive not really needed anyway
", "rawMarkdown": "because segmentation annotation is expensive not really needed anyway", "votes": 3 }, { "id": 1165351, "postDate": "2021-01-22T22:31:47.757Z", "content": "I think that's because chest x rays images are projection images. Many tissues are overlapped and so segmenting particular findings is very difficult.
", "rawMarkdown": "I think that's because chest x rays images are projection images. Many tissues are overlapped and so segmenting particular findings is very difficult. ", "votes": 1 } ], "comments": [ { "id": 1155395, "author_name": "Reuben Schmidt", "author_url": "", "post_date": "2021-01-16T11:26:51.383000", "content": "It's a good question, but I think the answer is relatively straightforward. As someone who looks at CXRs most days, I can confidently say the peripheral information encoded in the bounding box is generally not useful. CXRs are a relatively blunt instrument for diagnosis, and complex findings will generally lead a radiologist to suggest progression to CT. So the bounding box just detects the problem and highlights it for the clinician to review. Segmentation maps might e.g. be useful in depicting growth of a tumour in a patient over time, but this is beyond the scope of this type of challenge.
", "votes": 8, "replies": [ { "id": 1155594, "author_name": "ogureo", "author_url": "", "post_date": "2021-01-16T13:44:20.070000", "content": "Very interesting comment! So since you do a follow-up CT scan if there is an abnormality anyway, you don't need that much precision with x-rays. In terms of annotation, you can always convert from segmentation to bounding box but I do think it is way harder and time-consuming to label with segmentation brushes. So the reason could be that the cost of labelling segmentation is high and that 'precision' you get from the segmentation maps was not required for meeting the clinical objective of the challenge.
", "votes": 3, "replies": [] }, { "id": 1156035, "author_name": "Reuben Schmidt", "author_url": "", "post_date": "2021-01-16T21:47:25.970000", "content": "Yes, I think so. I had thought segmentation per pixel would be more compute expensive, so that's another reason why
", "votes": 2, "replies": [] } ] }, { "id": 1157601, "author_name": "DatNT", "author_url": "", "post_date": "2021-01-18T03:28:47.910000", "content": "because segmentation annotation is expensive not really needed anyway
", "votes": 3, "replies": [] }, { "id": 1165351, "author_name": "Massimiliano Porzio", "author_url": "", "post_date": "2021-01-22T22:31:47.757000", "content": "I think that's because chest x rays images are projection images. Many tissues are overlapped and so segmenting particular findings is very difficult.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1151456": "Many similar abnormality detection tasks in chest x-rays label their data by brushing the abnormalities to create segmentation maps. I was wondering why this competition uses a bounding box as the labels as segmentation maps seem more precise... Is it that there are too many classes to be a segmentation problem? Or is the peripheral information of the abnormality important somehow?", "1155395": "It's a good question, but I think the answer is relatively straightforward. As someone who looks at CXRs most days, I can confidently say the peripheral information encoded in the bounding box is generally not useful. CXRs are a relatively blunt instrument for diagnosis, and complex findings will generally lead a radiologist to suggest progression to CT. So the bounding box just detects the problem and highlights it for the clinician to review. Segmentation maps might e.g. be useful in depicting growth of a tumour in a patient over time, but this is beyond the scope of this type of challenge.", "1157601": "because segmentation annotation is expensive not really needed anyway", "1165351": "I think that's because chest x rays images are projection images. Many tissues are overlapped and so segmenting particular findings is very difficult. " } }