PATNet: Permute attention and transformer-enhanced network for segmentation of musculoskeletal ultrasound images

Abstract

Musculoskeletal ultrasound, due to its non-invasive nature, real-time feedback, and low cost, has been widely used for the evaluation of neuromuscular systems. However, owing to the structural complexity of muscle fibers, traditional image segmentation methods still face significant challenges in accurately identifying subtle pathological regions. To address the above issues, this paper proposes a deep learning model called PATNet (Permute Attention and Transformer-Enhanced Network for Segmentation of Musculoskeletal Ultrasound Images). The model is built upon the classical U-Net architecture to enhance the perception of microstructural features in muscle fibers.The Permute Spatial Attention (PSA) module reconstructs spatial information into the channel dimension, improving the model's sensitivity to muscle fiber orientation, density, and fine-grained texture variations. Meanwhile, the Permute Channel Attention (PCA) module models inter-channel dependencies, which helps suppress fat artifacts and background noise while highlighting anatomically relevant features of muscle tissue. In addition, PATNet incorporates Transformer modules to capture long-range dependencies in the image and leverages skip connections and multi-scale feature fusion mechanisms to enhance feature interaction and representation across different levels. Experimental results on multiple musculoskeletal ultrasound datasets demonstrate that the proposed PATNet model achieves excellent segmentation performance, significantly enhancing both the accuracy and robustness of muscle structure recognition.