Accuracy and Reliability of Deep Learning-based 2D Posture Analysis

This study assessed the accuracy and reliability of a 2D image-based deep learning algorithm for posture analysis by comparing it with a 3D motion capture system. Twenty healthy adult males participated, and nine balance parameters were measured using both methods: body tilt (ML/AP), shoulder tilt, pelvis tilt (ML/AP), knee tilt, left/right varus/valgus, and forward head posture. We evaluated agreement and reliability using root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficients, and intraclass correlation coefficients (ICC). Most parameters exhibited RMSE and MAE within 3°, while forward head posture, pelvis tilt (AP), and varus/valgus had errors below 10°. High correlations were found for shoulder tilt (r = 0.886) and forward head posture (r = 0.681), whereas knee tilt and left varus/valgus showed lower correlations due to methodological differences. Both methods demonstrated high repeatability (3D: ICC > 0.90, 2D: ICC > 0.80), with moderate-to-high agreement between methods (ICC ≥ 0.5 for most parameters). Shoulder tilt (ICC = 0.919) and forward head posture (ICC = 0.799) showed particularly high agreement. These findings indicate that 2D image-based posture analysis can provide accurate and reliable assessments comparable to 3D motion capture, presenting a more accessible and cost-effective alternative for posture evaluation in clinical and research contexts.