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. Arxiv 2022
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.

Training Tutorial

• FeTA 2021, FLARE 2021 Training Code (training.md)
• AMOS 2022 Finetuning Code (finetuning.md)

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

Methods	resolution	#params	FLOPs	Mean Dice (FeTA2021)	Mean Dice (FLARE2021)
TransBTS	96x96x96	31.6M	110.4G	0.868	0.902
UNETR	96x96x96	92.8M	82.6G	0.860	0.886
nnFormer	96x96x96	149.3M	240.2G	0.863	0.906
SwinUNETR	96x96x96	62.2M	328.4M	0.867	0.929
3D UX-Net	96x96x96	53.0M	639.4G	0.874 (kernel=7)	0.938 (kernel=13)

AMOS 2022 Fine-tuned Models

Methods	resolution	#params	FLOPs	Mean Dice (AMOS2022)
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	328.4M	0.880
3D UX-Net	96x96x96	53.0M	639.4G	0.900 (kernel=7)

Training

Training and fine-tuning instructions are in TRAINING.md

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{lee20223DUX-Net,
  author  = {Ho Hin Lee and Shunxing Bao and Yuankai Huo and Bennet A. Landman},
  title   = {3D UX-Net: A Large Kernel Volumetric ConvNet Modernizing Hierarchical Transformer for Medical Image Segmentation]},
  journal = {arXiv},
  year    = {2022},
}