data/README.md

Scaling Up 3D Kernels with Bayesian Frequency Re-parameterization for Medical Image Segmentation

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

Scaling Up 3D Kernels with Bayesian Frequency Re-parameterization for Medical Image Segmentation. MICCAI 2023 (Provisional Accepted, top 14%)
Ho Hin Lee, Quan Liu, Shunxing Bao, Qi Yang, Xin Yu, Leon Y. Cai, Thomas Li, Yuankai Huo, Xenofon Koutsoukos, Bennet A. Landman
Vanderbilt University
[arXiv]

---

We propose 3D RepUX-Net, a pure volumetric convolutional network that effectively adapts current largest 3D kernel sizes (e.g., 21x21x21) with spatial frequency modeling as Bayesian prior for weight re-parameterization during training.

Installation

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

Training Tutorial

• FLARE 2021 Training Code TRAINING.md
• AMOS 2022 Finetuning Code TRAINING.md

Results

FLARE 2021 Train From Scratch Models (5-folds cross-validation)

Methods	resolution	#params	FLOPs	Mean Dice	Model
nn-UNet	96x96x96	31.2M	743.3G	0.926
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 (k=7)	96x96x96	53.0M	639.4G	0.934
3D UX-Net (k=21)	96x96x96	65.9M	757.6G	0.930
3D RepUX-Net	96x96x96	65.8M	757.4G	0.944

AMOS 2022 Models (T.F.S: Train From Scratch, F.T: Fine-Tuning)

Methods	resolution	#params	FLOPs	Mean Dice (T.F.S) with weights	Mean Dice (F.T)
nn-UNet	96x96x96	31.2M	743.3G	0.850	0.878
TransBTS	96x96x96	31.6M	110.4G	0.783	0.792
UNETR	96x96x96	92.8M	82.6G	0.740	0.762
nnFormer	96x96x96	149.3M	240.2G	0.785	0.790
SwinUNETR	96x96x96	62.2M	328.4G	0.871	0.880
3D UX-Net (k=7)	96x96x96	53.0M	639.4G	0.890	0.900
3D UX-Net (k=21)	96x96x96	65.9M	757.6G	0.891	0.898
3D RepUX-Net	96x96x96	65.8M	757.4G	0.902 (Weights)	0.911

External Testing of FLARE-trained Model with 4 Different Datasets

Methods	MSD Spleen	KiTS Kidney	LiTS Liver	TCIA Pancreas
nn-UNet	0.917	0.829	0.935	0.739
TransBTS	0.881	0.797	0.926	0.699
UNETR	0.857	0.801	0.920	0.679
nnFormer	0.880	0.774	0.927	0.690
SwinUNETR	0.901	0.815	0.933	0.736
3D UX-Net (k=7)	0.926	0.836	0.939	0.750
3D UX-Net (k=21)	0.908	0.808	0.929	0.720
3D RepUX-Net	0.932	0.847	0.949	0.779

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 REPUXNET --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{lee2023scaling,
  author  = {Lee, Ho Hin and Liu, Quan and Bao, Shunxing and Yang, Qi and Yu, Xin and Cai, Leon Y and Li, Thomas and Huo, Yuankai and Koutsoukos, Xenofon and Landman, Bennett A},
  title   = {Scaling Up 3D Kernels with Bayesian Frequency Re-parameterization for Medical Image Segmentation},
  journal = {arXiv preprint arXiv:2303.05785},
  year    = {2023}
}