Development of a deep learning model for automated diagnosis of neuromuscular diseases using ultrasound imaging

Neuromuscular diseases (NMDs) pose significant diagnostic challenges due to their heterogeneous clinical manifestations and the limitations of traditional diagnostic tools. While musculoskeletal ultrasound has become a promising non-invasive modality for evaluating muscle pathology, its diagnostic accuracy remains heavily dependent on the operator's expertise. To address this, we propose a lightweight and interpretable deep learning model to enable automated classification of ultrasound images in NMD screening.

We developed a novel model, termed NMD-AssistNet, which integrates GhostNet as the backbone with CBAM attention modules and depthwise separable convolutions to enhance both efficiency and discriminative capacity. The model was trained and evaluated on a public dataset containing 3,917 annotated ultrasound images of various muscle groups. Mixup augmentation, label smoothing, and SWALR learning rate scheduling were applied to improve generalizability. Performance was benchmarked against CSPNet, EfficientNet, GhostNet, HRNet, and Vision Transformer.

NMD-AssistNet achieved the highest performance among the evaluated models, reaching a classification accuracy of 0.9502 and an area under the curve (AUC) of 0.9776. Grad-CAM visualizations revealed that the model effectively focused on clinically relevant muscle regions, highlighting its potential interpretability.

NMD-AssistNet demonstrates strong diagnostic capability, computational efficiency, and model interpretability and offers a promising solution for real-time, automated NMD screening. This framework has the potential to be deployed in portable ultrasound systems or edge AI devices to assist clinicians in both hospital and community healthcare settings.