![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3459992%2F2e4c48761299ec2b1d6ae69acac2793e%2FScreenshot%202025-09-24%20112600.png?generation=1759203766570990&alt=media\")
3D UNet Vessel Segmentation: Was anyone able to do vessel segmentation with good dice and use it to detect Aneurysms?
\nI have been working on this method, but have not had any success. I am not sure if I need to switch to 2D or use alternative approaches.
3D UNet Vessel Segmentation: Was anyone able to do vessel segmentation with good dice and use it to detect Aneurysms?
\nI have been working on this method, but have not had any success. I am not sure if I need to switch to 2D or use alternative approaches.
We trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!
", "rawMarkdown": "We trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!", "votes": 3, "replies": [ { "id": 3296219, "postDate": "2025-09-30T12:49:18.770Z", "content": "Thanks. I shaw approaches of many people and they are not using segmentation. I will try to train with nnunet, or I may need to use 2d CNN or dcm slices for aneurysm detection and classification.
", "rawMarkdown": "Thanks. I shaw approaches of many people and they are not using segmentation. I will try to train with nnunet, or I may need to use 2d CNN or dcm slices for aneurysm detection and classification." }, { "id": 3300503, "postDate": "2025-10-10T15:46:49.497Z", "content": "\n\nWe trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!
\n
Hi Evan, were you able to segment the left and right posterior communicating arteries using nnU-Net? I’m getting a Dice score of 0 for both of those classes.
", "rawMarkdown": "> We trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!\n\nHi Evan, were you able to segment the left and right posterior communicating arteries using nnU-Net? I’m getting a Dice score of 0 for both of those classes.\n", "replies": [ { "id": 3300514, "postDate": "2025-10-10T16:13:32.663Z", "content": "I had the same issue with nnunet. I trained for 1000 epochs, and around epoch 900 or so, those classes started increasing from 0. It was 0 until then
", "rawMarkdown": "I had the same issue with nnunet. I trained for 1000 epochs, and around epoch 900 or so, those classes started increasing from 0. It was 0 until then", "votes": 1 } ] } ] }, { "id": 3296044, "postDate": "2025-09-30T03:43:07.713Z", "content": "I managed to get decent enough segmentation model, but have not been able to effectively use it to increase my score.
Nice. Sounds counter intuitive. I thought having good segmentation model will reduce ROI and improve results much more.
", "rawMarkdown": "Nice. Sounds counter intuitive. I thought having good segmentation model will reduce ROI and improve results much more.", "replies": [ { "id": 3296220, "postDate": "2025-09-30T12:50:49.880Z", "content": "Indeed, I thought so too, spent quite some time on training a good segmentation model without any external GPU. Maybe I'm not using them properly. Either way, I have not gotten any improvement with segmentations.
", "rawMarkdown": "Indeed, I thought so too, spent quite some time on training a good segmentation model without any external GPU. Maybe I'm not using them properly. Either way, I have not gotten any improvement with segmentations." }, { "id": 3296869, "postDate": "2025-10-01T20:09:14.403Z", "content": "could you share how you trained for segmentation if you don't mind
", "rawMarkdown": "could you share how you trained for segmentation if you don't mind" }, { "id": 3296929, "postDate": "2025-10-02T01:10:08.237Z", "content": "Just a 3D Unet trained on 128x128x128 volumes. I did not use patch-based training , that did not work for me.
", "rawMarkdown": "Just a 3D Unet trained on 128x128x128 volumes. I did not use patch-based training , that did not work for me.", "votes": 1 }, { "id": 3296936, "postDate": "2025-10-02T01:41:42.513Z", "content": "Could you open source your segmentation model, want to give it a try?
", "rawMarkdown": "Could you open source your segmentation model, want to give it a try?", "votes": -1 }, { "id": 3296940, "postDate": "2025-10-02T01:49:18.960Z", "content": "I will after the competition ends. I spend 3 weeks of my GPU quota and time trying things to get this model, so I don't think I will share it now.
", "rawMarkdown": "I will after the competition ends. I spend 3 weeks of my GPU quota and time trying things to get this model, so I don't think I will share it now.", "votes": 1 }, { "id": 3296977, "postDate": "2025-10-02T04:32:09.833Z", "content": "One quick suggestion would be, to use this 3D UNet model to crop out sections for individual classes and then train a classification model on top of cropped sections.
", "rawMarkdown": "One quick suggestion would be, to use this 3D UNet model to crop out sections for individual classes and then train a classification model on top of cropped sections." }, { "id": 3296982, "postDate": "2025-10-02T04:37:36.383Z", "content": "@gowrishankarp I guess a lot of team have tried this.
", "rawMarkdown": "@gowrishankarp I guess a lot of team have tried this.", "votes": 1 }, { "id": 3296986, "postDate": "2025-10-02T04:49:12.890Z", "content": "That was the very first thing I tried. For some reason , does not work for me. Maybe a bug in my implementation , or maybe it just doesn't work.
", "rawMarkdown": "That was the very first thing I tried. For some reason , does not work for me. Maybe a bug in my implementation , or maybe it just doesn't work." }, { "id": 3298071, "postDate": "2025-10-04T12:35:17.840Z", "content": "@p4rallax What dice score did you manage to get on your segmentation model? Also you said that using it to increase ROI didn’t work right
", "rawMarkdown": "@p4rallax What dice score did you manage to get on your segmentation model? Also you said that using it to increase ROI didn’t work right" }, { "id": 3298076, "postDate": "2025-10-04T12:43:35.860Z", "content": "This model had a dice of about ~0.62. I have another model with dice ~0.71 that I haven't used yet. From my experiments so far, using that to crop out vessels, etc did not work for me. Could be because of poor quality of segmentations, might be bugs in my code. I don't have much experience with sampling crops , training on patches , etc so I am not entirely sure why it did not work for me. I will be giving it another go with the better segmentation model, within the next couple of days.
", "rawMarkdown": "This model had a dice of about ~0.62. I have another model with dice ~0.71 that I haven't used yet. From my experiments so far, using that to crop out vessels, etc did not work for me. Could be because of poor quality of segmentations, might be bugs in my code. I don't have much experience with sampling crops , training on patches , etc so I am not entirely sure why it did not work for me. I will be giving it another go with the better segmentation model, within the next couple of days.", "votes": 1 }, { "id": 3298077, "postDate": "2025-10-04T12:48:32.420Z", "content": "Sorry, one more question, even a simple bounding box crop around all vessels didn't work?
", "rawMarkdown": "Sorry, one more question, even a simple bounding box crop around all vessels didn't work?" }, { "id": 3298086, "postDate": "2025-10-04T13:13:33.703Z", "content": "Nope, not from what I tried. Don't take my word for it though - I spent about 3 weeks training segmentation models, so I wasted quite some time there and 2 of my teammates turned out ghosts, so I have not been giving too much attention to my experiments, since I had a lot of things to try out and I was basically solo. I just spent like a day implementing and running something, and if it did not seem to work, I just moved to the next idea. Hope this helps :)
", "rawMarkdown": "Nope, not from what I tried. Don't take my word for it though - I spent about 3 weeks training segmentation models, so I wasted quite some time there and 2 of my teammates turned out ghosts, so I have not been giving too much attention to my experiments, since I had a lot of things to try out and I was basically solo. I just spent like a day implementing and running something, and if it did not seem to work, I just moved to the next idea. Hope this helps :)", "votes": 1 } ] } ] }, { "id": 3295515, "postDate": "2025-09-29T02:21:41.410Z", "content": "The 3D U-Net is feasible, but transforms that may interfere with distinguishing left and right vessels should be removed. At a minimum, left-right flipping must be excluded.
", "rawMarkdown": "The 3D U-Net is feasible, but transforms that may interfere with distinguishing left and right vessels should be removed. At a minimum, left-right flipping must be excluded.", "votes": 1, "replies": [ { "id": 3296064, "postDate": "2025-09-30T05:09:34.997Z", "content": "May I ask — if both the image and label are flipped, why would left-right flipping still be an issue? Is the concern mainly that the model may struggle with different image orientations?
", "rawMarkdown": "May I ask — if both the image and label are flipped, why would left-right flipping still be an issue? Is the concern mainly that the model may struggle with different image orientations?", "replies": [ { "id": 3296077, "postDate": "2025-09-30T05:56:32.780Z", "content": "The human brain is somewhat symmetric, making the left and right sides highly similar in image features (e.g., the left and right eyes in CT scans). Current models rely heavily on image orientation to distinguish left from right, as they struggle to differentiate based on local features alone. Therefore, left-right flipping in data augmentation may cause confusion in laterality labels and is best avoided.
", "rawMarkdown": "The human brain is somewhat symmetric, making the left and right sides highly similar in image features (e.g., the left and right eyes in CT scans). Current models rely heavily on image orientation to distinguish left from right, as they struggle to differentiate based on local features alone. Therefore, left-right flipping in data augmentation may cause confusion in laterality labels and is best avoided.", "votes": 3 } ] } ] }, { "id": 3295516, "postDate": "2025-09-29T02:23:41.277Z", "content": "this: https://www.kaggle.com/code/tom99763/vessel-segmentation-tom
", "rawMarkdown": "this: https://www.kaggle.com/code/tom99763/vessel-segmentation-tom", "votes": 2, "replies": [ { "id": 3295911, "postDate": "2025-09-29T18:49:15.583Z", "content": "@tom99763 were you able to incorporate this into your submission, what is the score boost that was observed with vessel segmentation?
", "rawMarkdown": "@tom99763 were you able to incorporate this into your submission, what is the score boost that was observed with vessel segmentation?", "replies": [ { "id": 3295983, "postDate": "2025-09-29T23:54:19.200Z", "content": "@gowrishankarp I can tell you yes, I have included in one of the submissions. But if you mean score boost, I cannot tell you because this is core method I cannot compare.
", "rawMarkdown": "@gowrishankarp I can tell you yes, I have included in one of the submissions. But if you mean score boost, I cannot tell you because this is core method I cannot compare." } ] }, { "id": 3296222, "postDate": "2025-09-30T12:53:07.260Z", "content": "Thanks. I'll look over it. I'm starting to think single stage classification model is better. Segmentations and localizers just add complexity rather than helping model
", "rawMarkdown": "Thanks. I'll look over it. I'm starting to think single stage classification model is better. Segmentations and localizers just add complexity rather than helping model" } ] } ], "comments": [ { "id": 3295502, "author_name": "Evan Calabrese", "author_url": "", "post_date": "2025-09-29T00:39:49.043000", "content": "We trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!
", "votes": 3, "replies": [ { "id": 3296219, "author_name": "Saroj Neupane", "author_url": "", "post_date": "2025-09-30T12:49:18.770000", "content": "Thanks. I shaw approaches of many people and they are not using segmentation. I will try to train with nnunet, or I may need to use 2d CNN or dcm slices for aneurysm detection and classification.
", "votes": 0, "replies": [] }, { "id": 3300503, "author_name": "Luca", "author_url": "", "post_date": "2025-10-10T15:46:49.497000", "content": "\n\nWe trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!
\n
Hi Evan, were you able to segment the left and right posterior communicating arteries using nnU-Net? I’m getting a Dice score of 0 for both of those classes.
", "votes": 0, "replies": [ { "id": 3300514, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-10T16:13:32.663000", "content": "I had the same issue with nnunet. I trained for 1000 epochs, and around epoch 900 or so, those classes started increasing from 0. It was 0 until then
", "votes": 1, "replies": [] } ] } ] }, { "id": 3296044, "author_name": "Satwik", "author_url": "", "post_date": "2025-09-30T03:43:07.713000", "content": "I managed to get decent enough segmentation model, but have not been able to effectively use it to increase my score.
Nice. Sounds counter intuitive. I thought having good segmentation model will reduce ROI and improve results much more.
", "votes": 0, "replies": [ { "id": 3296220, "author_name": "Satwik", "author_url": "", "post_date": "2025-09-30T12:50:49.880000", "content": "Indeed, I thought so too, spent quite some time on training a good segmentation model without any external GPU. Maybe I'm not using them properly. Either way, I have not gotten any improvement with segmentations.
", "votes": 0, "replies": [] }, { "id": 3296869, "author_name": "Taha_Alshatiri", "author_url": "", "post_date": "2025-10-01T20:09:14.403000", "content": "could you share how you trained for segmentation if you don't mind
", "votes": 0, "replies": [] }, { "id": 3296929, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-02T01:10:08.237000", "content": "Just a 3D Unet trained on 128x128x128 volumes. I did not use patch-based training , that did not work for me.
", "votes": 1, "replies": [] }, { "id": 3296936, "author_name": "Gowri Shankar Penugonda", "author_url": "", "post_date": "2025-10-02T01:41:42.513000", "content": "Could you open source your segmentation model, want to give it a try?
", "votes": -1, "replies": [] }, { "id": 3296940, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-02T01:49:18.960000", "content": "I will after the competition ends. I spend 3 weeks of my GPU quota and time trying things to get this model, so I don't think I will share it now.
", "votes": 1, "replies": [] }, { "id": 3296977, "author_name": "Gowri Shankar Penugonda", "author_url": "", "post_date": "2025-10-02T04:32:09.833000", "content": "One quick suggestion would be, to use this 3D UNet model to crop out sections for individual classes and then train a classification model on top of cropped sections.
", "votes": 0, "replies": [] }, { "id": 3296982, "author_name": "Tom", "author_url": "", "post_date": "2025-10-02T04:37:36.383000", "content": "@gowrishankarp I guess a lot of team have tried this.
", "votes": 1, "replies": [] }, { "id": 3296986, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-02T04:49:12.890000", "content": "That was the very first thing I tried. For some reason , does not work for me. Maybe a bug in my implementation , or maybe it just doesn't work.
", "votes": 0, "replies": [] }, { "id": 3298071, "author_name": "Ravon", "author_url": "", "post_date": "2025-10-04T12:35:17.840000", "content": "@p4rallax What dice score did you manage to get on your segmentation model? Also you said that using it to increase ROI didn’t work right
", "votes": 0, "replies": [] }, { "id": 3298076, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-04T12:43:35.860000", "content": "This model had a dice of about ~0.62. I have another model with dice ~0.71 that I haven't used yet. From my experiments so far, using that to crop out vessels, etc did not work for me. Could be because of poor quality of segmentations, might be bugs in my code. I don't have much experience with sampling crops , training on patches , etc so I am not entirely sure why it did not work for me. I will be giving it another go with the better segmentation model, within the next couple of days.
", "votes": 1, "replies": [] }, { "id": 3298077, "author_name": "Ravon", "author_url": "", "post_date": "2025-10-04T12:48:32.420000", "content": "Sorry, one more question, even a simple bounding box crop around all vessels didn't work?
", "votes": 0, "replies": [] }, { "id": 3298086, "author_name": "Satwik", "author_url": "", "post_date": "2025-10-04T13:13:33.703000", "content": "Nope, not from what I tried. Don't take my word for it though - I spent about 3 weeks training segmentation models, so I wasted quite some time there and 2 of my teammates turned out ghosts, so I have not been giving too much attention to my experiments, since I had a lot of things to try out and I was basically solo. I just spent like a day implementing and running something, and if it did not seem to work, I just moved to the next idea. Hope this helps :)
", "votes": 1, "replies": [] } ] } ] }, { "id": 3295515, "author_name": "Shuolin Liu", "author_url": "", "post_date": "2025-09-29T02:21:41.410000", "content": "The 3D U-Net is feasible, but transforms that may interfere with distinguishing left and right vessels should be removed. At a minimum, left-right flipping must be excluded.
", "votes": 1, "replies": [ { "id": 3296064, "author_name": "Luca", "author_url": "", "post_date": "2025-09-30T05:09:34.997000", "content": "May I ask — if both the image and label are flipped, why would left-right flipping still be an issue? Is the concern mainly that the model may struggle with different image orientations?
", "votes": 0, "replies": [ { "id": 3296077, "author_name": "Shuolin Liu", "author_url": "", "post_date": "2025-09-30T05:56:32.780000", "content": "The human brain is somewhat symmetric, making the left and right sides highly similar in image features (e.g., the left and right eyes in CT scans). Current models rely heavily on image orientation to distinguish left from right, as they struggle to differentiate based on local features alone. Therefore, left-right flipping in data augmentation may cause confusion in laterality labels and is best avoided.
", "votes": 3, "replies": [] } ] } ] }, { "id": 3295516, "author_name": "Tom", "author_url": "", "post_date": "2025-09-29T02:23:41.277000", "content": "this: https://www.kaggle.com/code/tom99763/vessel-segmentation-tom
", "votes": 2, "replies": [ { "id": 3295911, "author_name": "Gowri Shankar Penugonda", "author_url": "", "post_date": "2025-09-29T18:49:15.583000", "content": "@tom99763 were you able to incorporate this into your submission, what is the score boost that was observed with vessel segmentation?
", "votes": 0, "replies": [ { "id": 3295983, "author_name": "Tom", "author_url": "", "post_date": "2025-09-29T23:54:19.200000", "content": "@gowrishankarp I can tell you yes, I have included in one of the submissions. But if you mean score boost, I cannot tell you because this is core method I cannot compare.
", "votes": 0, "replies": [] } ] }, { "id": 3296222, "author_name": "Saroj Neupane", "author_url": "", "post_date": "2025-09-30T12:53:07.260000", "content": "Thanks. I'll look over it. I'm starting to think single stage classification model is better. Segmentations and localizers just add complexity rather than helping model
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3295230": "3D UNet Vessel Segmentation: Was anyone able to do vessel segmentation with good dice and use it to detect Aneurysms?\nI have been working on this method, but have not had any success. I am not sure if I need to switch to 2D or use alternative approaches.", "3295502": "We trained an nnunet model when generating the vessel segmentations. Dice was >0.75. We used 3D fullres. Make sure you do not use mirroring for data augmentation!!!", "3296044": "I managed to get decent enough segmentation model, but have not been able to effectively use it to increase my score.", "3295515": "The 3D U-Net is feasible, but transforms that may interfere with distinguishing left and right vessels should be removed. At a minimum, left-right flipping must be excluded.", "3295516": "this: https://www.kaggle.com/code/tom99763/vessel-segmentation-tom" } }