| COMPARATIVE AI METHODS FOR DETECTING POSTURAL ANOMALIES |
| Paper ID : 1208-SPORTCONGRESS (R1) |
| Authors |
|
Parnian Rashidy *1, Nader Rahnama2, Maryam Ghorbani2, Mohammad Jalal Nemat bakhsh3 1Isfahan University 2University of Isfahan 3Sharif University of Technology |
| Abstract |
| Abstract Introduction: Accurate identification of postural anomalies is crucial for early intervention and rehabilitation in sports medicine and physiotherapy. This study presents two complementary, radiation-free artificial intelligence systems for automated posture analysis using RGB images from various angles. Methods: The first approach employs transfer learning on 2D RGB images, utilizing a pretrained CNN backbone fine-tuned to classify eleven distinct postural deformities alongside a normal posture category. Raw images from front, back, and lateral views are processed directly without explicit key point extraction. The second approach uses a clinically interpretable thresholding method, extracting skeletal key points via MMpose, computing standard clinical measurements, and applying empirically derived thresholds for anomaly classification. Results: A dataset of over 2,700 anonymized, ethically approved images was annotated by sports medicine specialists. The end-to-end CNN achieved an overall accuracy of 92.4% in five-fold cross-validation, with per-class sensitivities exceeding 90% for ten out of twelve postures. The thresholding method attained 88.1% overall accuracy, providing increased model transparency and per-class sensitivities ranging from 80% to 95%. Conclusion: Clinical integration was piloted with the Department of Sport Injuries & Corrective Exercises at the University of Isfahan, confirming compatibility with existing workflows and positive feedback from practitioners. Our findings highlight the strengths of both high-performance deep learning and clinically grounded thresholding methods. The deep learning approach excels in accuracy and speed, while the thresholding method offers greater interpretability aligned with clinical standards. Future work will focus on real-time mobile deployment, longitudinal outcome studies, and extending the analysis to three-dimensional pose estimation |
| Keywords |
| Keywords: Postural anomalies, artificial intelligence, sports medicine, deep learning, thresholding method. |
| Status: Abstract Accepted (Oral Presentation) |
