Synthetic data augmentation for improving performance in deep learning models for anatomical landmark localization on the distal upper limb.

Precise anatomical landmark localization is critical for the therapeutic efficacy of Traditional Eastern Medicine, yet manual localization methods show inter-practitioner variability exceeding 5 mm. Deep learning can automate this process, but data scarcity limits model development. MetaHuman-rendered synthetic images offer a scalable alternative, though their texture fidelity and architecture-specific training effects remain underexplored. Six landmark localization models (HRNet-W32/W48; YOLO26-pose: small/medium/large/extra-large) were trained for five anatomical landmarks on the distal upper limb under two configurations: real-only (2,963 images) and real-synthetic mixed (3,863 images). Synthetic image quality was assessed via Local Binary Pattern (LBP) chi-squared distances and Gray-Level Co-occurrence Matrix (GLCM) features. All models were evaluated on 50 held-out real images from 18 participants using localization error (mm) and mean Average Precision (mAP). Real-to-synthetic LBP chi-squared distances (0.0155 ± 0.0042) fell within natural real-image variation (0.0107 ± 0.0117), confirming sufficient texture fidelity. Mixed training reduced the mean localization error across all six models. Under subject-level clustered bootstrap analysis (n = 18 participant clusters, 10,000 iterations), statistically significant improvements were observed in YOLO26x-pose (17.8%; p = 0.002), YOLO26s-pose (14.5%; p = 0.001), and HRNet-W48 (6.3%; p = 0.020). However, because the mixed training set contains approximately 30% more images and gradient updates than the real-only set, contributions from increased training volume, optimization steps, and synthetic data content cannot be fully disentangled. Detection confidence remained stable across conditions. All models achieved sub-3.1 mm accuracy under mixed training, approaching expert-level consistency. These findings offer practical guidance for integrating MetaHuman-rendered synthetic images into deep learning pipelines for Traditional Eastern Medicine applications when real data is scarce.