Related Experiment Video
Updated: Sep 16, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
GH-UNet: group-wise hybrid convolution-VIT for robust medical image segmentation
Shengxiang Wang1, Ge Li2, Min Gao3
1School of Data Science and Artificial Intelligence, Wenzhou University of Technology, Wenzhou, China.
NPJ Digital Medicine
|July 10, 2025
Summary
GH-UNet improves medical image segmentation by combining convolutional and Transformer networks for better long-range dependency capture. This efficient model achieves superior performance with fewer parameters and computations.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Medical image segmentation is crucial for clinical diagnosis.
- U-Net models excel but face challenges with complex anatomical structures and long-range dependencies.
Purpose of the Study:
- To introduce GH-UNet, a novel Group-wise Hybrid Convolution-Vision Transformer (ViT) model designed to enhance medical image segmentation.
- To improve the capture of both local details and global context in medical images.
Main Methods:
- Developed GH-UNet, integrating a hybrid convolution-Transformer encoder for local and global feature extraction.
- Incorporated a Group-wise Dynamic Gating (GDG) module for adaptive feature weighting.
- Utilized a cascaded decoder for multi-scale feature integration, ensuring modularity for various backbones.
Main Results:
- GH-UNet demonstrated superior performance across five public and one private medical image segmentation datasets.
- Achieved significant gains in DICE (1.37%) and IOU (1.94%) on the ISIC2016 dataset compared to H2Former.
- Showcased remarkable efficiency, using only 38% of parameters and 49.61% of FLOPs.
Conclusions:
- GH-UNet effectively addresses the limitations of U-Net models in capturing long-range dependencies.
- The proposed model offers a computationally efficient and high-performing solution for medical image segmentation.
- The modular design allows for flexible integration with different network architectures.
