I have read somewhere that a 3D approach would not require segmentation.
\nI do not understand why wouldn't a segmentator, such as TotalSegmentator would be helpful alongside a 3D based CNN.
It's always quite challenging to train pure 3d classification. So segmentation can be good first stage and than each segmented organ of interest cropped separately and classified separately in 3d or 2.5d manner.
", "rawMarkdown": "It's always quite challenging to train pure 3d classification. So segmentation can be good first stage and than each segmented organ of interest cropped separately and classified separately in 3d or 2.5d manner.", "votes": 1 }, { "id": 2430805, "postDate": "2023-09-09T15:37:13.337Z", "content": "Are we even able to run TotalSegmentator on the test set in 9h runtime? Someone said there are 1.1k labels, so probably about 1.5k, scans. This gives about 22 seconds per scan to be preprocessed + inference. TotalSegmentator works up to a minute on the biggest scans.
", "rawMarkdown": "Are we even able to run TotalSegmentator on the test set in 9h runtime? Someone said there are 1.1k labels, so probably about 1.5k, scans. This gives about 22 seconds per scan to be preprocessed + inference. TotalSegmentator works up to a minute on the biggest scans. ", "votes": 1, "replies": [ { "id": 2430813, "postDate": "2023-09-09T15:46:33.243Z", "content": "what do you mean by 1.1k labels?
", "rawMarkdown": "what do you mean by 1.1k labels?", "replies": [ { "id": 2430818, "postDate": "2023-09-09T15:53:23.250Z", "content": "sorry, I meant subjects
1.1k patient. some patient has more than one scan.
\nexpected number of scan is around 1.5k
but do note that there is a low res 3mm version. the high res version is 1.5 mm.
", "rawMarkdown": "1.1k patient. some patient has more than one scan.\nexpected number of scan is around~~ 2.5k~~ 1.5k\n\nbut do note that there is a low res 3mm version. the high res version is 1.5 mm.", "votes": 1 }, { "id": 2430972, "postDate": "2023-09-09T17:33:39.657Z", "content": "Data description says there is only 1 or 2 scans per patient, no more, so I would say 1.5k.
\nI have tried 1.1k training patients on low ress and p100 gpu and it sometimes failed due to gpu memory limitations and took more than 9 hours.
You mean you have tried segmenting all images and that took more than 9 hours and failed?
", "rawMarkdown": "You mean you have tried segmenting all images and that took more than 9 hours and failed?" }, { "id": 2431487, "postDate": "2023-09-10T06:37:40.547Z", "content": "I mean after removing subjects with too big scans, which leads to gpu memory issues and reset notebook, it took more than 9 hours (it wasnt submission so the cap was 12 hours) to segment 1.1k patients from the training set. Additionally, TotalSegmentator was logging some problems with processing some data but it haven't stopped notebook and I didn't inspected those logs.
", "rawMarkdown": "I mean after removing subjects with too big scans, which leads to gpu memory issues and reset notebook, it took more than 9 hours (it wasnt submission so the cap was 12 hours) to segment 1.1k patients from the training set. Additionally, TotalSegmentator was logging some problems with processing some data but it haven't stopped notebook and I didn't inspected those logs. ", "votes": 1 }, { "id": 2431501, "postDate": "2023-09-10T06:49:59.657Z", "content": "Hm I see. Thanks for the answer. Hope we will find a way to speed it up
", "rawMarkdown": "Hm I see. Thanks for the answer. Hope we will find a way to speed it up" } ] }, { "id": 2443216, "postDate": "2023-09-17T15:34:06.090Z", "content": "Have you used Quantization?
", "rawMarkdown": "Have you used Quantization?" }, { "id": 2444071, "postDate": "2023-09-18T06:34:13.307Z", "content": "How did you run it so quickly? For me one image is 1.5 minutes
", "rawMarkdown": "How did you run it so quickly? For me one image is 1.5 minutes" } ] }, { "id": 2430752, "postDate": "2023-09-09T14:57:53.073Z", "content": "I have read somewhere that a 3D approach would not require segmentation.
\nI do not understand why wouldn't a segmentator, such as TotalSegmentator would be helpful alongside a 3D based CNN.
It's always quite challenging to train pure 3d classification. So segmentation can be good first stage and than each segmented organ of interest cropped separately and classified separately in 3d or 2.5d manner.
", "votes": 1, "replies": [] }, { "id": 2430805, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-09T15:37:13.337000", "content": "Are we even able to run TotalSegmentator on the test set in 9h runtime? Someone said there are 1.1k labels, so probably about 1.5k, scans. This gives about 22 seconds per scan to be preprocessed + inference. TotalSegmentator works up to a minute on the biggest scans.
", "votes": 1, "replies": [ { "id": 2430813, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-09T15:46:33.243000", "content": "what do you mean by 1.1k labels?
", "votes": 0, "replies": [ { "id": 2430818, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-09T15:53:23.250000", "content": "sorry, I meant subjects
1.1k patient. some patient has more than one scan.
\nexpected number of scan is around 1.5k
but do note that there is a low res 3mm version. the high res version is 1.5 mm.
", "votes": 1, "replies": [] }, { "id": 2430972, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-09T17:33:39.657000", "content": "Data description says there is only 1 or 2 scans per patient, no more, so I would say 1.5k.
\nI have tried 1.1k training patients on low ress and p100 gpu and it sometimes failed due to gpu memory limitations and took more than 9 hours.
You mean you have tried segmenting all images and that took more than 9 hours and failed?
", "votes": 0, "replies": [] }, { "id": 2431487, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-10T06:37:40.547000", "content": "I mean after removing subjects with too big scans, which leads to gpu memory issues and reset notebook, it took more than 9 hours (it wasnt submission so the cap was 12 hours) to segment 1.1k patients from the training set. Additionally, TotalSegmentator was logging some problems with processing some data but it haven't stopped notebook and I didn't inspected those logs.
", "votes": 1, "replies": [] }, { "id": 2431501, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-10T06:49:59.657000", "content": "Hm I see. Thanks for the answer. Hope we will find a way to speed it up
", "votes": 0, "replies": [] } ] }, { "id": 2443216, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-17T15:34:06.090000", "content": "Have you used Quantization?
", "votes": 0, "replies": [] }, { "id": 2444071, "author_name": "Gyula Maloveczky4", "author_url": "", "post_date": "2023-09-18T06:34:13.307000", "content": "How did you run it so quickly? For me one image is 1.5 minutes
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2431593": "It's always quite challenging to train pure 3d classification. So segmentation can be good first stage and than each segmented organ of interest cropped separately and classified separately in 3d or 2.5d manner.", "2430805": "Are we even able to run TotalSegmentator on the test set in 9h runtime? Someone said there are 1.1k labels, so probably about 1.5k, scans. This gives about 22 seconds per scan to be preprocessed + inference. TotalSegmentator works up to a minute on the biggest scans. ", "2430752": "I have read somewhere that a 3D approach would not require segmentation.\nI do not understand why wouldn't a segmentator, such as TotalSegmentator would be helpful alongside a 3D based CNN." } }