Related Experiment Video
Updated: Jul 14, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
439
SCA-Former: transformer-like network based on stream-cross attention for medical image segmentation.
Chengrui Gao1,2, Junlong Cheng1,2, Ziyuan Yang1
1School of Computer Science, Sichuan University, Chengdu, People's Republic of China.
Physics in Medicine and Biology
|October 6, 2023
Summary
This study introduces SCA-Former, a novel network for medical image segmentation that balances local and global feature extraction. SCA-Former enhances convolutional neural networks (CNNs) performance by effectively capturing long-range interactions without increasing model size.
Area of Science:
- Medical Image Analysis
- Deep Learning
- Computer Vision
Background:
- Convolutional Neural Networks (CNNs) excel in medical image analysis but struggle with long-range dependencies due to limited receptive fields.
- Transformers offer solutions for long-range interactions but often lead to significantly larger model sizes, hindering practical application.
- There is a need for efficient models that combine strong feature representation with effective long-range interaction modeling.
Purpose of the Study:
- To propose a novel Transformer-like block-based U-shaped network, SCA-Former, for medical image segmentation.
- To address the limitations of CNNs in capturing long-range information while maintaining a small model size.
- To develop a method that balances local and global feature representations for improved segmentation accuracy.
Main Methods:
- Introduced SCA-Former, a U-shaped network incorporating Transformer-like blocks for medical image segmentation.
- Proposed a novel stream-cross attention (SCA) module to extract multi-scale and interactive features across spatial and channel dimensions.
- The SCA module focuses on balancing local and global representations by capturing channel, multi-scale spatial, and long-range information.
Main Results:
- SCA-Former demonstrated superior performance compared to current state-of-the-art (SOTA) methods.
- Experiments were conducted on three public datasets: GLAS, ISIC 2017, and LUNG.
- The proposed SCA module effectively enhanced feature representation for segmentation tasks.
Conclusions:
- SCA-Former presents a promising approach for enhancing CNN feature representation in medical image segmentation.
- The method effectively improves segmentation performance by integrating long-range interaction modeling with efficient network design.
- This work offers a valuable contribution to the field of medical image analysis, particularly in segmentation applications.
