Reliable Markerless Pose Estimation Involving Clinically Relevant Finer Anatomical Landmarks for Infant General Movements Analysis

Objective: Automated markerless infant General Movements Assessment (GMA) requires accurate pose estimation (PE) to reliably reflect anatomical trajectories, particularly at fine distal joints that underpin clinically meaningful movement features. The performance of state-of-the-art markerless PE at metacarpophalangeal (MCP) and metatarsophalangeal (MTP) landmarks, however, remains unexplored. This study systematically evaluated the joint localization reliability of three advanced 2D markerless PE frameworks under clinical conditions. Results: DeepLabCut, Lightning Pose, and ViTPose were trained and evaluated under identical protocols using a clinically motivated 24-keypoint infant configuration spanning proximal and distal joints. All frameworks generalized well to unseen infant videos and achieved high aggregated accuracy at relaxed spatial thresholds ( (aPCK@0.10: 0.991, 0.978, and 0.969, respectively). However, normalized mean-absolute-error revealed systematic differences in spatial precision, with persistent errors at distal joints for Lightning Pose and ViTPose, while DeepLabCut achieved consistently lower absolute localization error. Conclusion: High threshold-based accuracy and visually smooth trajectories can overestimate functional precision at clinically sensitive distal landmarks. Anatomically comprehensive evaluation using complementary metrics is essential for reliable automated GMA and informed selection of PE frameworks for clinical applications.