Related Experiment Video
Updated: Jun 6, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
CC-TransXNet: a hybrid CNN-transformer network for automatic segmentation of optic cup and optic disk from fundus
Zhongzheng Yuan1, Jinke Wang2, Yukun Xu1
1Department of Software Engineering, Harbin University of Science and Technology, Weihai, 264300, China.
Medical & Biological Engineering & Computing
|November 27, 2024
Summary
CC-TransXNet, a novel hybrid network, enhances optic disk and optic cup segmentation for glaucoma diagnosis by combining CNN and transformer strengths. It improves accuracy and generalization, especially on small datasets, while managing model parameters effectively.
Area of Science:
- Ophthalmology and Medical Imaging
- Artificial Intelligence in Healthcare
- Computer Vision for Medical Diagnosis
Background:
- Accurate optic disk (OD) and optic cup (OC) segmentation is crucial for glaucoma diagnosis.
- Current Convolutional Neural Network (CNN) models lack global information and generalize poorly on small datasets.
- Transformer models capture global features but struggle with medical image segmentation due to high parameters and insufficient local information.
Purpose of the Study:
- To develop an innovative hybrid network, CC-TransXNet, integrating CNN and transformer advantages for improved OD and OC segmentation.
- To enhance generalization ability and accuracy in medical image segmentation, particularly for small datasets.
- To create an efficient model with controlled parameters to prevent overfitting.
Main Methods:
- Proposed CC-TransXNet, a W-shaped hybrid network utilizing TransXNet and improved ResNet for feature extraction, capturing both local and global information.
- Integrated Convolutional Block Attention Module (CBAM) within residual structures to enhance OD and OC recognition via channel and spatial attention.
- Employed Contextual Attention (CoT) self-attention mechanism in skip connections for adaptive contextual information allocation, boosting segmentation accuracy.
Main Results:
- CC-TransXNet demonstrated superior segmentation accuracy compared to traditional U-Net, CNN, and transformer-based networks across four public datasets (REFUGE 2, RIM-ONE DL, GAMMA, Drishti-GS).
- The model exhibited significantly enhanced generalization ability, particularly on small sample datasets.
- CC-TransXNet effectively controlled model parameters, mitigating overfitting risks and indicating potential for efficient medical image segmentation.
Conclusions:
- The proposed CC-TransXNet effectively addresses limitations of existing CNN and transformer models for optic nerve head segmentation.
- CC-TransXNet offers a promising solution for accurate and generalizable glaucoma diagnosis through improved OD and OC segmentation.
- The hybrid architecture provides an efficient and robust approach for medical image segmentation tasks.
