Would you be willing to share which 3D models you have been using and which ones performed best, whether transformer based or convolutional? Also, are you working with patch based inputs or full resolution images?\nI have tested UNet, UNETR, SegFormer, and FPN. In my experiments, FPN with patch based training performed best.
", "messages": [ { "id": "3408889", "postDate": "02/21/2026 14:57:16", "content": "Would you be willing to share which 3D models you have been using and which ones performed best, whether transformer based or convolutional? Also, are you working with patch based inputs or full resolution images?\nI have tested UNet, UNETR, SegFormer, and FPN. In my experiments, FPN with patch based training performed best.
", "rawMarkdown": "Would you be willing to share which 3D models you have been using and which ones performed best, whether transformer based or convolutional? Also, are you working with patch based inputs or full resolution images?\nI have tested UNet, UNETR, SegFormer, and FPN. In my experiments, FPN with patch based training performed best.", "votes": null }, { "id": "3408945", "postDate": "02/21/2026 17:30:22", "content": "Any particular backbones you like with FPN? For TransUnet I like seresnext, but it doesn’t outperform nnUNet for me. MedNeXt in an almost medium config (training is very slow) is slightly worse for me than TransUnet. Custom UNets also seem to exceed these two in my experience, but not nnUNet. Note, I’m not currently using the skeleton loss trainer from the orgs.
", "rawMarkdown": "Any particular backbones you like with FPN? For TransUnet I like seresnext, but it doesn’t outperform nnUNet for me. MedNeXt in an almost medium config (training is very slow) is slightly worse for me than TransUnet. Custom UNets also seem to exceed these two in my experience, but not nnUNet. Note, I’m not currently using the skeleton loss trainer from the orgs.", "votes": null }, { "id": "3408958", "postDate": "02/21/2026 18:21:12", "content": "The same backbone with FPN :D, lol. All other backbones were too heavy. Also looking forward to testing MAnet. The hard thing to do is, on one hand, finding a good balance between the loss, model choice, batching, augmentation, and ensembling, and at the end all of these will be halfway. On the other hand, this weird score function, it seems most of the top performers exploit it with special post processing. Still could not figure it out. I have been trying morphology, image processing, graph, polynomial, but I will get a better score. Still experimenting to find the best combo to go all in with k-fold.
", "rawMarkdown": "The same backbone with FPN :D, lol. All other backbones were too heavy. Also looking forward to testing MAnet. The hard thing to do is, on one hand, finding a good balance between the loss, model choice, batching, augmentation, and ensembling, and at the end all of these will be halfway. On the other hand, this weird score function, it seems most of the top performers exploit it with special post processing. Still could not figure it out. I have been trying morphology, image processing, graph, polynomial, but I will get a better score. Still experimenting to find the best combo to go all in with k-fold.", "votes": null }, { "id": "3408966", "postDate": "02/21/2026 18:36:31", "content": "Well to control for some of those I’m plugging the other models into the nnUNet infra. Agreed on the post processing. I’ve been avoiding it because it feels less interesting to me but seems unavoidable.
", "rawMarkdown": "Well to control for some of those I’m plugging the other models into the nnUNet infra. Agreed on the post processing. I’ve been avoiding it because it feels less interesting to me but seems unavoidable.", "votes": null } ], "comments": [ { "id": 3408945, "author_name": "rob1080ti", "author_url": "", "post_date": "02/21/2026 17:30:22", "content": "Any particular backbones you like with FPN? For TransUnet I like seresnext, but it doesn’t outperform nnUNet for me. MedNeXt in an almost medium config (training is very slow) is slightly worse for me than TransUnet. Custom UNets also seem to exceed these two in my experience, but not nnUNet. Note, I’m not currently using the skeleton loss trainer from the orgs.
", "votes": null, "replies": [ { "id": 3408958, "author_name": "jamalsaeedi", "author_url": "", "post_date": "02/21/2026 18:21:12", "content": "The same backbone with FPN :D, lol. All other backbones were too heavy. Also looking forward to testing MAnet. The hard thing to do is, on one hand, finding a good balance between the loss, model choice, batching, augmentation, and ensembling, and at the end all of these will be halfway. On the other hand, this weird score function, it seems most of the top performers exploit it with special post processing. Still could not figure it out. I have been trying morphology, image processing, graph, polynomial, but I will get a better score. Still experimenting to find the best combo to go all in with k-fold.
", "votes": null, "replies": [ { "id": 3408966, "author_name": "rob1080ti", "author_url": "", "post_date": "02/21/2026 18:36:31", "content": "Well to control for some of those I’m plugging the other models into the nnUNet infra. Agreed on the post processing. I’ve been avoiding it because it feels less interesting to me but seems unavoidable.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3408889": "Would you be willing to share which 3D models you have been using and which ones performed best, whether transformer based or convolutional? Also, are you working with patch based inputs or full resolution images?\nI have tested UNet, UNETR, SegFormer, and FPN. In my experiments, FPN with patch based training performed best.", "3408945": "Any particular backbones you like with FPN? For TransUnet I like seresnext, but it doesn’t outperform nnUNet for me. MedNeXt in an almost medium config (training is very slow) is slightly worse for me than TransUnet. Custom UNets also seem to exceed these two in my experience, but not nnUNet. Note, I’m not currently using the skeleton loss trainer from the orgs.", "3408958": "The same backbone with FPN :D, lol. All other backbones were too heavy. Also looking forward to testing MAnet. The hard thing to do is, on one hand, finding a good balance between the loss, model choice, batching, augmentation, and ensembling, and at the end all of these will be halfway. On the other hand, this weird score function, it seems most of the top performers exploit it with special post processing. Still could not figure it out. I have been trying morphology, image processing, graph, polynomial, but I will get a better score. Still experimenting to find the best combo to go all in with k-fold.", "3408966": "Well to control for some of those I’m plugging the other models into the nnUNet infra. Agreed on the post processing. I’ve been avoiding it because it feels less interesting to me but seems unavoidable." }, "source": "meta" }