/3DUX-Net

Primary LanguagePython

3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation

Official Pytorch implementation of 3D UX-Net, from the following paper:

3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation. ICLR 2023 (Accepted, Poster)
Ho Hin Lee, Shunxing Bao, Yuankai Huo, Bennet A. Landman
Vanderbilt University
[arXiv]

We propose 3D UX-Net, a pure volumetric convolutional network to adapt hierarchical transformers behaviour (e.g. Swin Transformer) for Medical Image Segmentation with less model parameters.

Installation

Please look into the INSTALL.md for creating conda environment and package installation procedures.

Training Tutorial

(Feel free to post suggestions in issues of recommending latest proposed transformer network for comparison. Currently, the network folder is to put the current SOTA transformer. We can further add the recommended network in it for training.)

Results

FeTA 2021 Trained Models (5-folds cross-validation)

Methods resolution #params FLOPs Mean Dice Model
TransBTS 96x96x96 31.6M 110.4G 0.868
UNETR 96x96x96 92.8M 82.6G 0.860
nnFormer 96x96x96 149.3M 240.2G 0.863
SwinUNETR 96x96x96 62.2M 328.4G 0.867
3D UX-Net 96x96x96 53.0M 639.4G 0.874 Weights

FLARE 2021 Trained Models (5-folds cross-validation)

Methods resolution #params FLOPs Mean Dice Model
TransBTS 96x96x96 31.6M 110.4G 0.902
UNETR 96x96x96 92.8M 82.6G 0.886
nnFormer 96x96x96 149.3M 240.2G 0.906
SwinUNETR 96x96x96 62.2M 328.4G 0.929
3D UX-Net 96x96x96 53.0M 639.4G 0.936 (latest) Weights

AMOS 2022 Fine-tuned Models

Methods resolution #params FLOPs Mean Dice (AMOS2022) Model
TransBTS 96x96x96 31.6M 110.4G 0.792
UNETR 96x96x96 92.8M 82.6G 0.762
nnFormer 96x96x96 149.3M 240.2G 0.790
SwinUNETR 96x96x96 62.2M 328.4G 0.880
3D UX-Net 96x96x96 53.0M 639.4G 0.900 (kernel=7) Weights

Training

Training and fine-tuning instructions are in TRAINING.md. Pretrained model weights will be uploaded for public usage later on.

Evaluation

Efficient evaulation can be performed for the above three public datasets as follows:

python test_seg.py --root path_to_image_folder --output path_to_output \
--dataset flare --network 3DUXNET --trained_weights path_to_trained_weights \
--mode test --sw_batch_size 4 --overlap 0.7 --gpu 0 --cache_rate 0.2 \

Acknowledgement

This repository is built using the timm library.

License

This project is released under the MIT license. Please see the LICENSE file for more information.

Citation

If you find this repository helpful, please consider citing:

@article{lee20223d,
  title={3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation},
  author={Lee, Ho Hin and Bao, Shunxing and Huo, Yuankai and Landman, Bennett A},
  journal={arXiv preprint arXiv:2209.15076},
  year={2022}
}