A Novel Skip-Connection Strategy by Fusing Spatial and Channel Wise Features for Multi-Region Medical Image Segmentation

Dayu Tan , Rui Hao , Xiaoping Zhou

IEEE Journal of Biomedical and Health Informatics +

|

May 29, 2024

View abstract on PubMed

Summary

FSCA-Net introduces novel attention mechanisms to U-shaped networks for improved spatial and channel feature extraction in segmentation tasks. This efficient deep learning model significantly reduces computational costs and parameters while enhancing segmentation accuracy.

Area Of Science

Computer Vision
Deep Learning
Medical Image Analysis

Background

Existing U-shaped networks often use attention in skip connections but neglect spatial information and channel efficiency.
This leads to suboptimal feature extraction and inefficiency in capturing crucial spatial and channel details.

Purpose Of The Study

To propose the Fusing Spatial and Channel Attention Network (FSCA-Net), a novel deep learning architecture.
To enhance spatial and channel feature extraction within skip connections to compensate for downsampling losses.
To improve fine-grained segmentation of multiple organs and regions.

Main Methods

FSCA-Net employs a Parallel Attention Transformer (PAT) to boost spatial and channel feature extraction in skip connections.
A Cross-Attention Bridge Layer (CAB) is designed to minimize feature and resolution loss during downsampling.
A Dual-Path Channel Attention (DPCA) module refines Transformer features, ensuring semantic consistency with the U-Net decoder.

Main Results

FSCA-Net achieved over 48% reduction in FLOPs and over 32% reduction in parameters compared to state-of-the-art methods.
The model demonstrated superior performance across seven public datasets for fine-grained segmentation tasks.
Significant improvements in capturing spatial and channel information were observed.

Conclusions

FSCA-Net offers an efficient and effective solution for fine-grained medical image segmentation.
The proposed architecture successfully addresses limitations in spatial and channel information processing in U-shaped networks.
FSCA-Net represents a significant advancement in deep learning for complex segmentation challenges.