Explainable gait-based scoliosis detection using dense optical flow

Early detection of scoliosis through non-invasive methods remains a key clinical challenge, especially in adolescent populations. While prior work has demonstrated the feasibility of gait-based analysis for scoliosis classification using silhouette or pose-based representations, these inputs may discard relevant motion cues. In this work, we explore the use of dense optical flow as an alternative input modality to capture fine-grained biomechanical manifestations of spinal deformities. Building upon the ScoNet-MT architecture, we systematically compare optical flow configurations against silhouette-based baselines using the Scoliosis1K dataset. Our results show that optical flow, particularly RAFT computed on full-size RGB frames, substantially improves classification performance across all dataset splits, achieving an accuracy of 93.59% on the clinically realistic 1:1:8 split. Additionally, we introduce OF-ScoNet-MT Lite, a parameter-efficient variant of the original model that achieves 87.85% accuracy, using less than half the parameters of the base model. To enhance clinical trust, we incorporate an explainability framework based on stabilized Element Shapley Values (ESV), which quantifies the temporal contribution of gait phases. By synchronizing model attributions with manually annotated gait cycles, we demonstrate that the model prioritizes the posterior ( Away ) phase, establishing a verifiable link between deep learning predictions and clinical gait observations. These findings underscore the importance of dense motion representations in medical gait analysis and validate the feasibility of lightweight, non-invasive deep learning tools for large-scale scoliosis screening.