IHPE: A Lightweight Hand Pose Estimation Network for Surgical Motion Capture in Medical Education

Currently, most medical education adopts a master-apprentice model combined with video-based instruction; however, this approach is inefficient in correcting apprentices’ mistakes. To address this issue, we propose a low-cost and accurate surgical hand pose estimation method to enhance medical education. Specifically, we design an Integrated Hand Pose Estimation (IHPE) network, which combines hand keypoint detection with deep learning-based object detection to achieve precise hand pose estimation during surgery. IHPE is constructed by employing dynamic convolution and ghost convolution to build a lightweight yet effective backbone. By replacing traditional C2f layers with newly designed In-C2f modules, the framework generates low-cost ghost features and applies dynamically generated weights to improve feature representation while significantly reducing model complexity. We conduct extensive experiments on the Hagrid public dataset and a private surgical action dataset to validate the effectiveness of IHPE. On the Hagrid dataset, IHPE achieves a precision of 86.2%, with only 2.40M parameters and a model complexity of 6.6 GFLOPs. On the private dataset, IHPE achieves a precision of 90.8%, with 2.49M parameters and a complexity of 6.9 GFLOPs, outperforming many existing lightweight detectors. In addition, we construct IHPE-eva, a teacher-student knot-tying action evaluation system based on IHPE. IHPE-eva detects hand keypoints from video streams, reconstructs motion trajectories, evaluates the periodicity of knot-tying actions using Fourier Transform, and assesses movement smoothness via the standard deviation of acceleration. This system enables real-time evaluation and provides valuable feedback for surgical trainees, supporting mobile deployment scenarios.