Related Experiment Video
Updated: Jun 25, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Local global feature enhanced transformer with attention pruning for precise medical image segmentation
Ning Xu1, Qun Song2, Hongxu Yin2
1The Third Affiliated Hospital of Kunming Medical University, Yunnan Cancer Hospital, Peking University Cancer Hospital Yunnan, Kunming, China.
Iscience
|June 24, 2026
Summary
This study introduces the Local Feature Enhanced AgentTopk UNet (LFEAT-UNet) for medical image segmentation. It improves accuracy by enhancing local details and reducing redundant computations, offering a robust solution for clinical applications.
Area of Science:
- Medical Image Analysis
- Artificial Intelligence
- Computer Vision
Background:
- Vision Transformer (ViT) methods are effective for global modeling in medical image segmentation.
- Challenges include capturing fine-grained local details and mitigating attention redundancy in ViT models.
Purpose of the Study:
- To propose a novel hybrid Transformer network, LFEAT-UNet, for enhanced medical image segmentation.
- To improve the representation of detailed structures and optimize computational efficiency.
Main Methods:
- Introduced Local Feature Enhancement (LFE) module for detailed structure information.
- Implemented Token Selection and Filtering (TSF) mechanism to prune background tokens.
- Utilized Cascaded Multi-scale Patch Perception (CMSPP) module for improved feature fusion.
Main Results:
- Achieved Dice scores of 83.12% on the Synapse dataset and 92.20% on the ACDC dataset.
- Demonstrated strong robustness and superior performance on a private adrenal nodule dataset.
- LFEAT-UNet effectively delineates complex clinical targets.
Conclusions:
- LFEAT-UNet offers an effective, sparsity-aware solution for medical image segmentation.
- The proposed method shows significant potential for clinical applications.
- Addresses limitations of existing ViT-based methods in medical imaging.