ABSTRACT

　Foot posture assessment is recognized as a screening tool in the prevention of sportsrelated injuries. However, the practical application of clinical foot posture evaluation indices in field settings remains limited due to the need for specialized knowledge and training among medical professionals. This study aimed to develop and validate a deep learning-based model that classifies foot posture using images of the hindfoot. The classification was based on indicators derived from the Foot Posture Index, focusing on the relative alignment of Abduction/Adduction of the forefoot on the rearfoot, as well as Frontal plane alignment and position of the calcaneus. The model was designed to categorize foot posture into five types: adduction, abduction, varus, valgus, and neutral. A total of 750 posterior foot images were collected and divided into training （70%）， validation （20%）， and test （10%） datasets. During training, the model demonstrated progressive improvement, with increasing precision and decreasing loss values across epochs, indicating effective learning. The model achieved a mean Average Precision （mAP（ B）） of 0.72 at IoU thresholds ranging from 0.50 to 0.95（mAP50-95（B））， and 0.80 at an IoU threshold of 0.50（mAP50（B））， reflecting high overall accuracy. Evaluation using the test dataset revealed that while the classification accuracy for the neutral foot posture was relatively low （accuracy: 0.26）， the model performed well in identifying the other four postures, with accuracy ranging from 0.69 to 1.00. These results suggest that the model has practical potential for use in sports and clinical settings, particularly for non-specialists who require accessible and automated tools for foot posture assessment. Future work will focus on improving the classification of neutral posture and expanding the dataset to enhance generalizability.

DECENTE SPORTS SCIENCE Vol.47/The DESCENTE AND ISHIMOTO MEMORIAL FOUNDATION FOR THE PROMOTION SPORTS SCIENCE