Enhanced multi-object detection in brachial plexus ultrasound images using transformer-based semantic and spatial feature fusion.
Authors
Affiliations (6)
Affiliations (6)
- School of Design, Shanghai Jiaotong University, Shanghai, 200240, China.
- State Key Laboratory of Media Convergence Production Technology and Systems, Xinhua News Agency, Beijing, 100803, China.
- Network and Data Security Key Laboratory of Sichuan Province, School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, 610054, Sichuan, China.
- Network and Data Security Key Laboratory of Sichuan Province, School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, 610054, Sichuan, China. [email protected].
- YIBIN GREAT Technology Co., Ltd, Yibin, 644000, Sichuan, China. [email protected].
- School of Design, Shanghai Jiaotong University, Shanghai, 200240, China. [email protected].
Abstract
Advances in medical imaging have led to the widespread use of ultrasound for diagnosing and treating peripheral nerve blocks. However, challenges such as speckle noise and low spatial resolution hinder the accurate identification of neural structures. To address these issues, a model named transformer-based detection with semantic and spatial feature fusion (TDSS) was introduced, integrating spatial edge adaptive enhancement, multi-scale semantic fusion, and spatial distillation-assisted detection modules. These components improve edge perception, multi-scale semantic reasoning, and localization accuracy. Experiments on the Ultrasound Brachial Plexus Dataset demonstrated the superior performance of the model: TDSS achieved an AP of 0.407 for nerve detection and an overall mAP of 0.583, providing robust support for clinical medical image detection. The source code for the proposed TDSS framework is publicly available at https://doi.org/10.5281/zenodo.18824199 .