![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4310004%2Fcde64c55cfc6d0ffe00dd91424e15149%2Fimage.png?generation=1765407430983923&alt=media\")
code：https://github.com/HuXiaoling/Warping\npaper：https://arxiv.org/abs/2112.07812\nI tried to transfer code from HuXiaoling/Warping to topo-pitfalls to support 3D, but result is not better than DiceLoss. The critical points number seems a lot when debugging, not as the paper.\nMaybe BettiMatch and ClDice is the correct way？
", "messages": [ { "id": "3362032", "postDate": "12/04/2025 05:25:12", "content": "code：https://github.com/HuXiaoling/Warping\npaper：https://arxiv.org/abs/2112.07812\nI tried to transfer code from HuXiaoling/Warping to topo-pitfalls to support 3D, but result is not better than DiceLoss. The critical points number seems a lot when debugging, not as the paper.\nMaybe BettiMatch and ClDice is the correct way？
", "rawMarkdown": "code：https://github.com/HuXiaoling/Warping\npaper：https://arxiv.org/abs/2112.07812\nI tried to transfer code from HuXiaoling/Warping to topo-pitfalls to support 3D, but result is not better than DiceLoss. The critical points number seems a lot when debugging, not as the paper.\nMaybe BettiMatch and ClDice is the correct way？", "votes": null }, { "id": "3362038", "postDate": "12/04/2025 05:35:04", "content": "Moreover, it does not appear in the reference papers of the competition.
", "rawMarkdown": "Moreover, it does not appear in the reference papers of the competition.", "votes": null }, { "id": "3362324", "postDate": "12/04/2025 13:56:22", "content": "I don't think we tried it. Did we, @seanjohnsonsp ?
", "rawMarkdown": "I don't think we tried it. Did we, @seanjohnsonsp ?", "votes": null }, { "id": "3362395", "postDate": "12/04/2025 16:29:13", "content": "I did not use this one unfortunately, it looks interesting though!
", "rawMarkdown": "I did not use this one unfortunately, it looks interesting though!", "votes": null }, { "id": "3362573", "postDate": "12/04/2025 23:55:54", "content": "@clwwlc basically my approach does similar thing.
", "rawMarkdown": "clwwlc basically my approach does similar thing.", "votes": null }, { "id": "3362926", "postDate": "12/05/2025 15:21:42", "content": "just a suggestion:
\nUse you ground truth to intersect your prediction. Those not intersected are fp bridges, etc. see if you can treat this a a new class. basically it is a iterative relabel process (for each detection, label fp and fn. that guarantees good betti numbers)
", "rawMarkdown": "just a suggestion:\n- i remember in early cell instance segmentation kaggle competition, the top solution detected the touching border pixel between cells.\n(so there are three class, bg,fg, touching)\n- hence instead of using update_simple_point() to weight the pixel loss, detect the critical point as an additional class.\n- use them as post processing later?\n\nUse you ground truth to intersect your prediction. Those not intersected are fp bridges, etc. see if you can treat this a a new class. basically it is a iterative relabel process (for each detection, label fp and fn. that guarantees good betti numbers)", "votes": null }, { "id": "3370435", "postDate": "12/10/2025 19:35:22", "content": "Skeleton Recall Loss seems interesting (mentioned in the overview page). Maybe MIC-DKFZ team used it and got 0.549 in one shot. -)
\nhttps://github.com/MIC-DKFZ/Skeleton-Recall
\nTo use it, one need to run tubed skeletonization on gt from dataloader. Or, how about gt are precomputed beforehand, same as the original gt.
", "rawMarkdown": "Skeleton Recall Loss seems interesting (mentioned in the overview page). Maybe MIC-DKFZ team used it and got 0.549 in one shot. -)\n\nhttps://github.com/MIC-DKFZ/Skeleton-Recall\n\nTo use it, one need to run tubed skeletonization on gt from dataloader. Or, how about gt are precomputed beforehand, same as the original gt.", "votes": null }, { "id": "3370442", "postDate": "12/10/2025 19:43:56", "content": "Great idea! The Vesuvius Challenge team tried to use in the past a modified version of this, which instead of doing a skeletonization in XY slices, does an aggregate skeletonization in all dimensions. This is a sort of approximation of the Medial Surface. It's definetely working quite well, but I don't think that alone will be enough to win the competition :)
", "rawMarkdown": "Great idea! The Vesuvius Challenge team tried to use in the past [a modified version of this](https://github.com/ScrollPrize/villa/blob/e85da047518542ea04b59d62ae7cbbd7691c838d/segmentation/models/arch/nnunet/nnunetv2/training/nnUNetTrainer/variants/loss/nnUNetTrainerSkeletonRecall.py#L513), which instead of doing a skeletonization in XY slices, does an aggregate skeletonization in all dimensions. This is a sort of approximation of the Medial Surface. It's definetely working quite well, but I don't think that alone will be enough to win the competition :)", "votes": null }, { "id": "3370551", "postDate": "12/10/2025 22:57:26", "content": "@ipythonx for us, skeleton recall give negative effect, here are the extracted skeleton in 2d view.
\n
why do you allow skeletion artifacts (junctions and small looping branches) in your ground truth?
", "rawMarkdown": "why do you allow skeletion artifacts (junctions and small looping branches) in your ground truth?", "votes": null }, { "id": "3370567", "postDate": "12/10/2025 23:43:15", "content": "maybe helpful for VOI loss if you do it instance wise (instance as class) for ground truth?
", "rawMarkdown": "maybe helpful for VOI loss if you do it instance wise (instance as class) for ground truth?", "votes": null }, { "id": "3385111", "postDate": "01/02/2026 14:56:01", "content": "I tried it, but it didn’t help.
", "rawMarkdown": "I tried it, but it didn’t help.", "votes": null } ], "comments": [ { "id": 3362038, "author_name": "clwwlc", "author_url": "", "post_date": "12/04/2025 05:35:04", "content": "Moreover, it does not appear in the reference papers of the competition.
", "votes": null, "replies": [] }, { "id": 3362324, "author_name": "giorgioangelotti", "author_url": "", "post_date": "12/04/2025 13:56:22", "content": "I don't think we tried it. Did we, @seanjohnsonsp ?
", "votes": null, "replies": [ { "id": 3362395, "author_name": "seanjohnsonsp", "author_url": "", "post_date": "12/04/2025 16:29:13", "content": "I did not use this one unfortunately, it looks interesting though!
", "votes": null, "replies": [] } ] }, { "id": 3362573, "author_name": "tom99763", "author_url": "", "post_date": "12/04/2025 23:55:54", "content": "@clwwlc basically my approach does similar thing.
", "votes": null, "replies": [] }, { "id": 3362926, "author_name": "hengck23", "author_url": "", "post_date": "12/05/2025 15:21:42", "content": "just a suggestion:
\nUse you ground truth to intersect your prediction. Those not intersected are fp bridges, etc. see if you can treat this a a new class. basically it is a iterative relabel process (for each detection, label fp and fn. that guarantees good betti numbers)
", "votes": null, "replies": [] }, { "id": 3370435, "author_name": "ipythonx", "author_url": "", "post_date": "12/10/2025 19:35:22", "content": "Skeleton Recall Loss seems interesting (mentioned in the overview page). Maybe MIC-DKFZ team used it and got 0.549 in one shot. -)
\nhttps://github.com/MIC-DKFZ/Skeleton-Recall
\nTo use it, one need to run tubed skeletonization on gt from dataloader. Or, how about gt are precomputed beforehand, same as the original gt.
", "votes": null, "replies": [ { "id": 3370442, "author_name": "giorgioangelotti", "author_url": "", "post_date": "12/10/2025 19:43:56", "content": "Great idea! The Vesuvius Challenge team tried to use in the past a modified version of this, which instead of doing a skeletonization in XY slices, does an aggregate skeletonization in all dimensions. This is a sort of approximation of the Medial Surface. It's definetely working quite well, but I don't think that alone will be enough to win the competition :)
", "votes": null, "replies": [] }, { "id": 3370551, "author_name": "tom99763", "author_url": "", "post_date": "12/10/2025 22:57:26", "content": "@ipythonx for us, skeleton recall give negative effect, here are the extracted skeleton in 2d view.
\n
why do you allow skeletion artifacts (junctions and small looping branches) in your ground truth?
", "votes": null, "replies": [] } ] }, { "id": 3370567, "author_name": "hengck23", "author_url": "", "post_date": "12/10/2025 23:43:15", "content": "maybe helpful for VOI loss if you do it instance wise (instance as class) for ground truth?
", "votes": null, "replies": [] } ] }, { "id": 3385111, "author_name": "shigengtian", "author_url": "", "post_date": "01/02/2026 14:56:01", "content": "I tried it, but it didn’t help.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3362032": "code：https://github.com/HuXiaoling/Warping\npaper：https://arxiv.org/abs/2112.07812\nI tried to transfer code from HuXiaoling/Warping to topo-pitfalls to support 3D, but result is not better than DiceLoss. The critical points number seems a lot when debugging, not as the paper.\nMaybe BettiMatch and ClDice is the correct way？", "3362038": "Moreover, it does not appear in the reference papers of the competition.", "3362324": "I don't think we tried it. Did we, @seanjohnsonsp ?", "3362395": "I did not use this one unfortunately, it looks interesting though!", "3362573": "clwwlc basically my approach does similar thing.", "3362926": "just a suggestion:\n- i remember in early cell instance segmentation kaggle competition, the top solution detected the touching border pixel between cells.\n(so there are three class, bg,fg, touching)\n- hence instead of using update_simple_point() to weight the pixel loss, detect the critical point as an additional class.\n- use them as post processing later?\n\nUse you ground truth to intersect your prediction. Those not intersected are fp bridges, etc. see if you can treat this a a new class. basically it is a iterative relabel process (for each detection, label fp and fn. that guarantees good betti numbers)", "3370435": "Skeleton Recall Loss seems interesting (mentioned in the overview page). Maybe MIC-DKFZ team used it and got 0.549 in one shot. -)\n\nhttps://github.com/MIC-DKFZ/Skeleton-Recall\n\nTo use it, one need to run tubed skeletonization on gt from dataloader. Or, how about gt are precomputed beforehand, same as the original gt.", "3370442": "Great idea! The Vesuvius Challenge team tried to use in the past [a modified version of this](https://github.com/ScrollPrize/villa/blob/e85da047518542ea04b59d62ae7cbbd7691c838d/segmentation/models/arch/nnunet/nnunetv2/training/nnUNetTrainer/variants/loss/nnUNetTrainerSkeletonRecall.py#L513), which instead of doing a skeletonization in XY slices, does an aggregate skeletonization in all dimensions. This is a sort of approximation of the Medial Surface. It's definetely working quite well, but I don't think that alone will be enough to win the competition :)", "3370551": "ipythonx for us, skeleton recall give negative effect, here are the extracted skeleton in 2d view.\n\n", "3370566": "why do you allow skeletion artifacts (junctions and small looping branches) in your ground truth?", "3370567": "maybe helpful for VOI loss if you do it instance wise (instance as class) for ground truth?", "3385111": "I tried it, but it didn’t help." }, "source": "meta" }