Poserwgcn: Development and Preliminary Clinical Validation of a Real-Time 3D Human Pose Estimation Method for Radiotherapy Based on Rwkv and Graph Convolutional Networks
Abstract
Purpose
Patient pose variations during radiotherapy can degrade setup accuracy and compromise treatment safety. Existing monitoring approaches often suffer from limited real-time capability and complex system configurations. To address these limitations, this study proposes a markerless, real-time pose monitoring method tailored for radiotherapy, aiming to improve patient positioning accuracy and monitoring efficiency.
Methods
We propose PoseRWGCN, an attention-free dual-stream model that integrates Receptance Weighted Key Value (RWKV) and Graph Convolutional Networks (GCNs) for 3D human pose estimation. The RWKV stream models global temporal dynamics and employs a recursive gating mechanism to efficiently capture long-range dependencies, while the GCN stream extracts local spatial features constrained by the skeletal topology. The two representations are unified via an adaptive fusion strategy. The model supports causal inference for real-time online monitoring and bidirectional inference for offline video analysis. In addition, a multi-scale real-time alerting module is designed to assess pose deviations using keypoint displacement, joint-angle variations, and projected-length changes, and to trigger alerts when predefined thresholds are exceeded.
Results
On Human3.6M, PoseRWGCN achieves MPJPEs of 41.9 mm in the non-causal setting and 42.2 mm in the causal setting. On MPI-INF-3DHP, it attains an MPJPE of 15.4 mm, with a PCK of 98.6 and an AUC of 85.6. In radiotherapy scenario tests, the model achieves an MPJPE of 58.2 mm and a mean joint-angle error of 1.3987°, while running at approximately 20 fps, meeting real-time monitoring requirements.
Conclusion
PoseRWGCN provides accurate and efficient markerless 3D pose estimation, enabling real-time patient pose monitoring and positioning support during radiotherapy. The proposed multi-scale alerting mechanism enhances risk notification capability and may improve treatment safety.