MNet is novel data-independent CNN segmentation architecture, which realizes adaptive 2D and 3D feature fusion to balance inter- and intra-slice representation, thus being robust to varying anisotropic degrees of medical datasets and helping get rid of manual architecture design.
For more information about MNet, please read the following paper （Accepted by IJCAI 2022）:
@misc{
doi = {10.48550/ARXIV.2205.04846},
url = { https://doi.org/10.48550/arxiv.2205.04846 },
author = {Dong, Zhangfu and He, Yuting and Qi, Xiaoming and Chen, Yang and Shu, Huazhong and Coatrieux, Jean-Louis and Yang, Guanyu and Li, Shuo},
title = {MNet: Rethinking 2D/3D Networks for Anisotropic Medical Image Segmentation},
publisher = {arXiv},
year = {2022},
}
Implementations of MNet with Pytorch and MindSpore (https://www.mindspore.cn/).
Although MNet and all the involved networks in our experiments are implemented and tested with Pytorch, we here provide the additional implementation with MindSpore to promote the widespread application of MNet.
The proposed MNet is trained with nnUNet framework, thus we provide the whole modified nnUNet project.
--Modifications we have done:
- Add MNet.py and basic_module.py to /nnUNet/nnunet/network_architecture
- Add myTrainer.py to /nnUNet/nnunet/training/network_training
--Training cmd:
nnUNet_train 3d_fullres myTrainer TaskXXX_MYTASK FOLD --npz (see https://github.com/MIC-DKFZ/nnUNet for details)
The public datasets used in our paper:
- LiTS: https://competitions.codalab.org/competitions/17094#learn_the_details-evaluation
- KiTS: https://kits19.grand-challenge.org/data/
- BraTS: https://www.kaggle.com/datasets/awsaf49/brats2020-training-data
- PROMISE: https://promise12.grand-challenge.org/
