gatsby2016 / Scale-Slice-awareNet

The offical implementation of the network architecture: Scale- and Slice- aware Network for 3D segmentation of organs and musculoskeletal structures in pelvic MRI

Scale‐ and Slice‐aware Net for 3D segmentation of organs and musculoskeletal structures in pelvic MRI

The offical implementation of the network architecture in Paper: Scale- and Slice- aware Net for 3D segmentation of organs and musculoskeletal structures in pelvic MRI

Research Type：Machine Learning/Deep Learning，Image processing/Image analysis， Technical Research

Research Focus：Anatomy， Muscle，Musculoskeletal

Abstract

S^2aNet is presented for 3D dense segmentation of 54 organs and musculoskeletal structures in female pelvic MR images. A Scale- aware module is designed to capture the spatial and semantic information of different-scale structures. A Slice-aware module is introduced to model similar spatial relationships of consecutive slices in 3D data. Moreover, S^2aNet leverages a weight-adaptive loss optimization strategy to reinforce the supervision with more discriminative capability on hard samples and categories.

Highlights

• a weight-adaptive loss optimization strategy is introduced to alleviate difficult samples and the problems of class imbalance.

• a multislice-aware feature fusion module is proposed to encode and fuse features from different slices by a parameter-sharing mechanism.

• a parallel multiscale-aware module is designed to extract both spatial information of large-scale categories and semantic information of small-scale categories without losing spatial resolution.

• To our knowledge, this is the first report to achieve a 3D dense segmentation for pelvic 54 structures on MRI.

Results

Experiments have been performed on a pelvic MRI cohort of 27 MR images from 27 patient cases. Across the cohort and 54 categories of organs and musculoskeletal structures manually delineated, S^2aNet was shown to outperform the UNet framework and other state-of-the-art fully convolutional networks in terms of sensitivity, Dice similarity coefficient and relative volume difference.

The segmentation results are given below:

Installation and Usage

You need to config the environment firstly, install python and corresponding packages, including torch, opencv, SimpleITK, and so on.

For independent evaluation, run

python A5_test.py

and if you want to generate the predicted results, run

python A6_inference25D.py 

or if you only want to infer in 2D mode:

python A6_inference2D.py 

Our network architecture files are Scale_slice_awareNet.py and net_msacunet.py, and both of them are the same.

You can also train the network on your data from scratch.

python A4_trainNet_macunet.py

!!! remember to adjust the data path or others privately to yours.

Citation

If the project helps your research, please cite the following paper:

@article{yan2022scale,
  title={Scale-and Slice-aware Net (S2aNet) for 3D segmentation of organs and musculoskeletal structures in pelvic MRI},
  author={Yan, Chaoyang and Lu, Jing-Jing and Chen, Kang and Wang, Lei and Lu, Haoda and Yu, Li and Sun, Mengyan and Xu, Jun},
  journal={Magnetic Resonance in Medicine},
  volume={87},
  number={1},
  pages={431--445},
  year={2022},
  publisher={Wiley Online Library}
}

Acknowledgement

I cherished the memories in AIMLab and thank you all for the wonderful time.

Contact

If you have any problems, just raise an issue in this repo.