[2Win5] Poster session 2

[2Win5-60] Exploring AI-Based Prediction of Standing Posture from Foot Sole Images

AI's Potential for Indirect Inference from Images

〇Hiroko Awata¹, Junki Yamamura², Motoshige Takagi³, Mitsuhiro Zeida¹, Yuichi Suzuki¹ (1.Suntory Global Innovation Center Limited, 2.Suntory Wellness Limited, 3.Suntory Holdings Limited)

Keywords:Foot sole image, Standing posture, Deep learning model

Foot contact condition is closely linked to standing posture, suggesting potential as a simple method for posture assessment. This study investigates the feasibility of AI-based standing posture prediction using foot sole images. We collected foot sole and lateral posture images from 781 subjects, calculating neck, shoulder, and lumbar angles to classify posture as normal or abnormal (forward head, kyphosis, swayback). Three deep learning architectures were evaluated: AlexNet, ResNet50, and MLP-Mixer, with and without additional foot morphological features including angular parametrs and foot width-to-length ratios. In binary classification between normal and abnormal postures, our modified AlexNet achieved the highest performance with an average AUC 0.70 and maximum AUC of 0.89. For specific posture abnormalities, MLP-Mixer showed superior performance in detecting forward head posture (AUC 0.77), while ResNet50 performed better for kyphosis (AUC 0.64) and swayback (AUC 0.61). The effectiveness of additional foot morphological features varied by architecture and posture type. These results demonstrate the potential of AI-based posture assessment through foot sole images, while highlighting areas for improvement. Future work will focus on enhancing detection accuracy through architectural innovations and feature engineering.