A manual acupuncture manipulation recognition deep learning model via self-supervised and adaptive ultrasound image key information extraction methods

The current research on the quantification and identification of manual acupuncture manipulations (MAMs) mainly focuses on acupuncturists’ MAM operations, without clear quantitative model that interprets the effects on subcutaneous muscle tissue movement derived from different MAMs. In MAM teaching and clinical applications, the practitioner can only rely on subjective experience for technique selection, as they are unable to identify the MAMs through feedback from the movement of subcutaneous muscle tissue. Consequently, it is difficult to provide evidence for the correct application of MAMs. To address these issues, we first conduct a real-time ultrasound image dataset of subcutaneous muscle tissue effects for four typical MAMs. Subsequently, a deep learning recognition model is designed based on self-supervised and adaptive key information extraction from ultrasound images for MAMs. To select effective information from lengthy video sequences, we developed a keyframe extraction method based on the variation in optical flow intensity. Additionally, to guide the deep learning network’s focus on muscle movement features in the area surrounding the acupuncture-affected region, we designed a self-supervised and adaptive dynamic region-of-interest (ROI) mask extraction module. Then, we utilize dynamic masks to guide the spatiotemporal attention mechanism of the deep network, to precise capture muscle movement effects under different MAMs. Finally, we conducted experiments involving eighteen experienced acupuncturists, two ultrasound technicians, and 157 healthy subjects from three medical institutions in the ultrasound data collection process. Experimental results indicate that the proposed MAM recognition model achieves an identification accuracy of 92.68 %, confirming its validity and effectiveness.