Related Experiment Video
Updated: Jul 13, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Graph- and transformer-guided boundary aware network for medical image segmentation.
Shanshan Xu1, Lianhong Duan2, Yang Zhang3
1School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China; Beijing Key Laboratory of Traffic Data Analysis and Mining, Beijing Jiaotong University, Beijing 100044, China.
Computer Methods and Programs in Biomedicine
|October 14, 2023
Summary
A new medical image segmentation model, GTBA-Net, effectively handles complex backgrounds and noise. It accurately delineates ambiguous boundaries, outperforming existing methods across diverse imaging modalities.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Artificial Intelligence
Background:
- Medical image segmentation is crucial but challenging due to complex backgrounds, noise, and ambiguous boundaries.
- Existing U-Net-based models face limitations in addressing these segmentation difficulties.
Purpose of the Study:
- To introduce a novel U-shaped Graph- and Transformer-guided Boundary Aware Network (GTBA-Net) for improved medical image segmentation.
- To address challenges of complex backgrounds, irrelevant noises, and ambiguous boundaries in medical imaging.
Main Methods:
- GTBA-Net utilizes a ResNet34 backbone with Global Feature Aggregation (GFA) for localization, Graph-based Dynamic Feature Fusion (GDFF) for noise suppression, and Uncertainty-based Boundary Refinement (UBR) for boundary delineation.
- GFA employs self-attention for efficient target localization. GDFF uses graph attention to fuse features and suppress noise while preserving details.
- UBR incorporates an uncertainty quantification strategy and auxiliary loss to refine ambiguous boundaries.
Main Results:
- GTBA-Net demonstrated superior performance compared to existing methods on five diverse medical imaging datasets (X-ray, CT, endoscopic, ultrasound).
- Ablation studies confirmed the individual contributions of GFA, GDFF, and UBR modules to localization, noise suppression, and boundary refinement.
Conclusions:
- GTBA-Net shows significant potential for broad application in medical image segmentation.
- The model is particularly effective in scenarios with complex backgrounds, noise, and ambiguous boundaries.
