AASTFNet: An Attention-Aware Spatial-Temporal Fusion Network for Enhanced Pain Intensity Evaluation in Facial Image

Video-based pain intensity assessment offers continuous discomfort quantification, overcoming limitations of conventional subjective methods through automated facial expression analysis. However, current deep learning architectures, despite excelling in extracting spatiotemporal pain features from facial expressions, struggle to localize transient pain signals across consecutive frames and specific facial regions, limiting detection accuracy and clinical utility. We propose the Attention-Aware Spatiotemporal Fusion Network (AASTFNet). This framework combines a temporal sub-network with feature gating to focus on video frames containing pain-related information, and a spatial sub-network with attention mechanisms to highlight localized pain-related facial regions. Additionally, it integrates facial landmark geometric information to compensate for global spatial feature loss. Feature vectors from both sub-networks are fused via an adaptive feature fusion network with multi-head attention mechanisms to capture global and local dependencies for pain intensity regression analysis. Tested on the UNBC-McMaster Shoulder Pain Expression Archive Database, AASTFNet achieves MAE $=0.31$, MSE $=0.52$, and $\text{PCC}=0.89$, surpassing state-of-theart methods.