Related Experiment Video
Updated: Jun 18, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Dual-scale shifted window attention network for medical image segmentation
De-Wei Han1, Xiao-Lei Yin2, Jian Xu1
1School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, 1088 Xueyuan Boulevard, Nanshan District, Shenzhen, 518055, China.
Scientific Reports
|July 31, 2024
Summary
This study introduces a Dual-Scale Transformer using double-sized shifted windows to improve image segmentation. This novel approach enhances information flow between image patches, achieving state-of-the-art results.
Area of Science:
- Computer Vision
- Deep Learning
- Image Segmentation
Background:
- Transformers offer excellent performance in computer vision but can be computationally complex.
- Window-based and shifted window-based self-attention mechanisms in Swin Transformers aim to reduce this complexity.
- Optimizing patch information communication is key to improving Transformer efficiency in vision tasks.
Purpose of the Study:
- To propose a Dual-Scale Transformer with a double-sized shifted window attention method.
- To investigate the impact of different shifted window sizes on patch information communication efficiency.
- To enhance image segmentation performance by improving information flow within Transformer networks.
Main Methods:
- Developed a Dual-Scale Transformer architecture incorporating double-sized shifted window attention.
- Evaluated the proposed method against CNN-based (U-Net, AttenU-Net, ResU-Net, CE-Net) and single-scale Swin Transformer (SwinT) models.
- Conducted an ablation study to analyze the effect of shifted window size on information flow and segmentation performance.
Main Results:
- The Dual-Scale Transformer significantly outperformed CNN-based methods by 3%-6% and single-scale Swin Transformer by approximately 1%.
- Achieved state-of-the-art segmentation results on Kvasir-SEG, ISIC2017, MICCAI EndoVisSub-Instrument, and CadVesSet datasets.
- Confirmed that dual-scale shifted window attention enhances patch information communication and segmentation accuracy.
Conclusions:
- The proposed dual-scale shifted window attention is an optimized network design for image segmentation.
- This approach effectively improves information flow, leading to enhanced segmentation performance.
- Network architecture design significantly impacts visual performance in Transformer-based models.
