![\"result\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3086083%2F725e1967c817f26ed36322189ae4d6a2%2Fkidney_result.png?generation=1701424272556296&alt=media\")
Sharing my PyTorch volumetric dataset. It provides small subvolumes, favoring sections with positive samples, and rotates them around one of the axes, permutes them, flips them, etc. It produces quite a range of samples to work with.
\nhttps://www.kaggle.com/code/limitz/pytorch-dataset-with-volumetric-augmentations
", "messages": [ { "id": "2544707", "postDate": "12/01/2023 03:26:27", "content": "Sharing my PyTorch volumetric dataset. It provides small subvolumes, favoring sections with positive samples, and rotates them around one of the axes, permutes them, flips them, etc. It produces quite a range of samples to work with.
\nhttps://www.kaggle.com/code/limitz/pytorch-dataset-with-volumetric-augmentations
", "rawMarkdown": "Sharing my PyTorch volumetric dataset. It provides small subvolumes, favoring sections with positive samples, and rotates them around one of the axes, permutes them, flips them, etc. It produces quite a range of samples to work with. \n\n[https://www.kaggle.com/code/limitz/pytorch-dataset-with-volumetric-augmentations](https://www.kaggle.com/code/limitz/pytorch-dataset-with-volumetric-augmentations)", "votes": null }, { "id": "2545025", "postDate": "12/01/2023 08:38:22", "content": "any comparsion results for 3d verus 2d?
\nin my experiments, 3d don't help much unless you have better and more data augmentation (more augmentation than 2d)
At this moment it seems 2d algorithms do better (probably due to the availability of good backbones ?). Still, for sport, I'm working on a volumetric (transformer) model to see how far I can take it.
\nAs for augmentation, the volumetric rotation provides more augmentation than doing standard 2d transforms on xy,zy,zx slices, as the volumetric transform slices on arbitrary axes. (unless of course you are actually doing something similar in 2d preprocessing).
", "rawMarkdown": "At this moment it seems 2d algorithms do better (probably due to the availability of good backbones ?). Still, for sport, I'm working on a volumetric (transformer) model to see how far I can take it. \n\nAs for augmentation, the volumetric rotation provides more augmentation than doing standard 2d transforms on xy,zy,zx slices, as the volumetric transform slices on arbitrary axes. (unless of course you are actually doing something similar in 2d preprocessing).", "votes": null }, { "id": "2545107", "postDate": "12/01/2023 09:53:19", "content": "It does seem to help with eliminating gaps, but it also misses the big arteries due to limited \"patch\" size
\n
source, gt, prediction, (gt-prediction):
\n![\"pred\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3086083%2Fecb6570ae384eaede60e246d2f51ead4%2Fkidney_diff.png?generation=1701424353179823&alt=media\")
any comparsion results for 3d verus 2d?
\nin my experiments, 3d don't help much unless you have better and more data augmentation (more augmentation than 2d)
At this moment it seems 2d algorithms do better (probably due to the availability of good backbones ?). Still, for sport, I'm working on a volumetric (transformer) model to see how far I can take it.
\nAs for augmentation, the volumetric rotation provides more augmentation than doing standard 2d transforms on xy,zy,zx slices, as the volumetric transform slices on arbitrary axes. (unless of course you are actually doing something similar in 2d preprocessing).
", "votes": null, "replies": [] }, { "id": 2545107, "author_name": "limitz", "author_url": "", "post_date": "12/01/2023 09:53:19", "content": "It does seem to help with eliminating gaps, but it also misses the big arteries due to limited \"patch\" size
\n
source, gt, prediction, (gt-prediction):
\n