I just found this paper and it seems to be helpful.
\n\nImage Thresholding Improves 3-Dimensional Convolutional Neural Network Diagnosis of Different Acute Brain Hemorrhages on Computed Tomography Scans\nJustin Ker, Satya P. Singh, [...], and Lipo Wang
\n\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6539746/
", "messages": [ { "id": 657995, "postDate": "2019-10-25T15:14:10.470Z", "content": "I just found this paper and it seems to be helpful.
\n\nImage Thresholding Improves 3-Dimensional Convolutional Neural Network Diagnosis of Different Acute Brain Hemorrhages on Computed Tomography Scans\nJustin Ker, Satya P. Singh, [...], and Lipo Wang
\n\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6539746/
", "rawMarkdown": "I just found this paper and it seems to be helpful.\n\nImage Thresholding Improves 3-Dimensional Convolutional Neural Network Diagnosis of Different Acute Brain Hemorrhages on Computed Tomography Scans\nJustin Ker, Satya P. Singh, [...], and Lipo Wang\n\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6539746/\n\n", "votes": 6 }, { "id": 662124, "postDate": "2019-10-31T05:48:51.923Z", "content": "Thanks for sharing @felipekitamura \nJust to add on a simple article which i could understand atleast :)\nhttps://www.slideshare.net/ganesahyogananthem/ct-numbers-window-width-and-window-level \nIs it something similar to below example\nA window with a level of 0 HU and a width of 400 HU will have a \nrange of −200 HU to +200 HU. Any tissue with a density of −200 HU or less will be black, and any tissue with a \ndensity of +200 HU or more will be white.
", "rawMarkdown": "Thanks for sharing @felipekitamura \nJust to add on a simple article which i could understand atleast :)\nhttps://www.slideshare.net/ganesahyogananthem/ct-numbers-window-width-and-window-level \nIs it something similar to below example\nA window with a level of 0 HU and a width of 400 HU will have a \nrange of −200 HU to +200 HU. Any tissue with a density of −200 HU or less will be black, and any tissue with a \ndensity of +200 HU or more will be white.\n", "votes": 1 }, { "id": 660686, "postDate": "2019-10-29T13:23:13.347Z", "content": "Damn, this article is fun and I want to try so many things but resources are so limited as an individual... God I wish kaggle didn't put such a low limit on gpu quota. 😿
", "rawMarkdown": "Damn, this article is fun and I want to try so many things but resources are so limited as an individual... God I wish kaggle didn't put such a low limit on gpu quota. 😿 ", "votes": 2, "replies": [ { "id": 660768, "postDate": "2019-10-29T15:15:58.827Z", "content": "Haha, yeah I feel ya Max. I lost 39 hours of gpu quota due to some bugs, and another time due to the kernel taking more than 9 hours to complete training. 😄 Try out GCP!
", "rawMarkdown": "Haha, yeah I feel ya Max. I lost 39 hours of gpu quota due to some bugs, and another time due to the kernel taking more than 9 hours to complete training. 😄 Try out GCP!", "votes": 2 } ] }, { "id": 658327, "postDate": "2019-10-25T22:28:50.197Z", "content": "Thanks for sharing:\nThresholding seems interesting, not entirely clear how it was implemented. Though the improvement in training time is substantial!!
", "rawMarkdown": "Thanks for sharing:\nThresholding seems interesting, not entirely clear how it was implemented. Though the improvement in training time is substantial!!", "votes": 2, "replies": [ { "id": 658353, "postDate": "2019-10-25T23:09:32.740Z", "content": "I would say, zero out everything below ~45HU and everything above ~110HU. But this is just a guess from a radiologist.
", "rawMarkdown": "I would say, zero out everything below ~45HU and everything above ~110HU. But this is just a guess from a radiologist. ", "votes": 1 }, { "id": 660998, "postDate": "2019-10-29T21:09:30.287Z", "content": "Tried this recently, a simpler version though. I really creates sharp edges, different individual parts become more visible, it improved score on a simple, low res model, but gave no substantial benefit on full resolution.
", "rawMarkdown": "Tried this recently, a simpler version though. I really creates sharp edges, different individual parts become more visible, it improved score on a simple, low res model, but gave no substantial benefit on full resolution. ", "votes": 4 } ] } ], "comments": [ { "id": 662124, "author_name": "NeerajSharma", "author_url": "", "post_date": "2019-10-31T05:48:51.923000", "content": "Thanks for sharing @felipekitamura \nJust to add on a simple article which i could understand atleast :)\nhttps://www.slideshare.net/ganesahyogananthem/ct-numbers-window-width-and-window-level \nIs it something similar to below example\nA window with a level of 0 HU and a width of 400 HU will have a \nrange of −200 HU to +200 HU. Any tissue with a density of −200 HU or less will be black, and any tissue with a \ndensity of +200 HU or more will be white.
", "votes": 1, "replies": [] }, { "id": 660686, "author_name": "madmax0404", "author_url": "", "post_date": "2019-10-29T13:23:13.347000", "content": "Damn, this article is fun and I want to try so many things but resources are so limited as an individual... God I wish kaggle didn't put such a low limit on gpu quota. 😿
", "votes": 2, "replies": [ { "id": 660768, "author_name": "David Tang", "author_url": "", "post_date": "2019-10-29T15:15:58.827000", "content": "Haha, yeah I feel ya Max. I lost 39 hours of gpu quota due to some bugs, and another time due to the kernel taking more than 9 hours to complete training. 😄 Try out GCP!
", "votes": 2, "replies": [] } ] }, { "id": 658327, "author_name": "David Tang", "author_url": "", "post_date": "2019-10-25T22:28:50.197000", "content": "Thanks for sharing:\nThresholding seems interesting, not entirely clear how it was implemented. Though the improvement in training time is substantial!!
", "votes": 2, "replies": [ { "id": 658353, "author_name": "FelipeKitamura, MD, PhD", "author_url": "", "post_date": "2019-10-25T23:09:32.740000", "content": "I would say, zero out everything below ~45HU and everything above ~110HU. But this is just a guess from a radiologist.
", "votes": 1, "replies": [] }, { "id": 660998, "author_name": "Eek The Cat", "author_url": "", "post_date": "2019-10-29T21:09:30.287000", "content": "Tried this recently, a simpler version though. I really creates sharp edges, different individual parts become more visible, it improved score on a simple, low res model, but gave no substantial benefit on full resolution.
", "votes": 4, "replies": [] } ] } ], "raw_markdown_by_id": { "657995": "I just found this paper and it seems to be helpful.\n\nImage Thresholding Improves 3-Dimensional Convolutional Neural Network Diagnosis of Different Acute Brain Hemorrhages on Computed Tomography Scans\nJustin Ker, Satya P. Singh, [...], and Lipo Wang\n\nhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC6539746/\n\n", "662124": "Thanks for sharing @felipekitamura \nJust to add on a simple article which i could understand atleast :)\nhttps://www.slideshare.net/ganesahyogananthem/ct-numbers-window-width-and-window-level \nIs it something similar to below example\nA window with a level of 0 HU and a width of 400 HU will have a \nrange of −200 HU to +200 HU. Any tissue with a density of −200 HU or less will be black, and any tissue with a \ndensity of +200 HU or more will be white.\n", "660686": "Damn, this article is fun and I want to try so many things but resources are so limited as an individual... God I wish kaggle didn't put such a low limit on gpu quota. 😿 ", "658327": "Thanks for sharing:\nThresholding seems interesting, not entirely clear how it was implemented. Though the improvement in training time is substantial!!" } }