Skip to content

[IEEE TMI-2024] SASAN: Spectrum-Axial Spatial Approach Networks for Medical Image Segmentation

Notifications You must be signed in to change notification settings

IMOP-lab/SASAN-Pytorch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

62 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SASAN: Spectrum-Axial Spatial Approach Networks for Medical Image Segmentation

Our paper has been accepted by IEEE Transactions on Medical Imaging!

by Xingru Huang, Jian Huang, Kai Zhao, Tianyun Zhang, Zhi Li, Changpeng Yue, Wenhao Chen, Ruihao Wang, Xuanbin Chen, Qianni Zhang, Ying Fu, Yangyundou Wang, and Yihao Guo

Hangzhou Dianzi University IMOP-lab

Figure 1: Detailed network structure of the SASAN.

The proposed SASAN is an innovative 3D medical image segmentation network that integrates spectrum information. It achieves state-of-the-art performance over 13 previous methods on the CMED and OIMHS datasets.

We will first introduce our method and principles, then introduce the experimental environment and provide Github links to previous methods we have compared. Finally, we will present the experimental results.

Installation

We run SASAN and previous methods on a system running Ubuntu 22.04, with Python 3.9, PyTorch 2.0.0, and CUDA 11.8. For a full list of software packages and version numbers, see the experimental environment file 'environment.yaml'.

Experiment

Baselines

We provide GitHub links pointing to the PyTorch implementation code for all networks compared in this experiment here, so you can easily reproduce all these projects.

3D U-Net; RAUNet; UNETR; SwinUNETR; ResUNet; MultiResUNet; V-Net; 3D UX-Net; SegResNet; HighRes3DNet; TransBTS; nnFormer; SETR

Compare with others on the CMED dataset

Figure 2: Comparison experiments between our method and 13 previous segmentation methods on the CMED dataset.

Figure 3: The visual results of our method compared to the existing 13 segmentation methods on the CMED dataset.

Our method demonstrates the best performance across all categories and metrics. SASAN outperforms previous methods in detail segmentation of choroidal and macular edema categories, highlighting the role of low-frequency data in enhancing details and reducing noise. In addition, SASAN performs well in boundary segmentation of retinal categories, thanks to its BoundaryRea loss and self-updating mechanism, which improves sensitivity to boundary distance.

Ablation study

Key components of SASAN

Figure 4: Ablation experiments on key components of SASAN on the CMED dataset.

FINE introduces a wide range of low-frequency features, which has a good effect in reducing high-frequency noise and enhancing detail extraction. ASEM has a strong ability to enhance the network for the analysis of features that are difficult to distinguish objects.

Loss function strategy

Figure 5: Ablation experiments on Loss function strategy on the CMED dataset.

The self-updating mechanism and BoundaryRea Loss enhance the network's boundary segmentation ability. To a certain extent, it also improves the overall segmentation ability of the network.

Model Complexity

Figure 6: Comparative assessment of parameters, FLOPs, and inference time for our proposed method versus classical models under uniform evaluation settings and computer configuration.

Citation

If SASAN is useful for your research, please consider citing:

@article{huang2024sasan,
  title={SASAN: Spectrum-Axial Spatial Approach Networks for Medical Image Segmentation},
  author={Huang, Xingru and Huang, Jian and Zhao, Kai and Zhang, Tianyun and Li, Zhi and Yue, Changpeng and Chen, Wenhao and Wang, Ruihao and Chen, Xuanbin and Zhang, Qianni and others},
  journal={IEEE Transactions on Medical Imaging},
  year={2024},
  publisher={IEEE}
}

Question

if you have any questions, please contact 'j.huang@hdu.edu.cn'

About

[IEEE TMI-2024] SASAN: Spectrum-Axial Spatial Approach Networks for Medical Image Segmentation

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages