Development of a Mobile Application for Assessing the Severity of Hallux Valgus Using Smartphone Foot Images

Background: Hallux valgus (HV) is a common forefoot deformity that can lead to pain, altered gait, and musculoskeletal dysfunctions. Accurate severity assessment is essential for clinical decision-making, yet radiographic methods, though accurate—are costly and less accessible. Objects: This study aimed to develop and clinically validate an end-to-end artificial intelligence (AI)-based mobile application for HV severity classification from smartphone-captured dorsal foot photographs. Methods: The study comprised two phases. In Phase 1 (App & Model Development), we developed a mobile application integrating foot Red-Green-Blue (RGB) image capture, HV severity classification, and immediate reporting. Paired (weight-bearing anteroposterior foot) radiographs and smartphone dorsal foot photographs were collected from 180 adults with HV. Radiographic HV angle and intermetatarsal angle were measured to categorize severity (mild, moderate, severe) as ground truth. A MobileNetV2 convolutional neural network (CNN) was trained on dorsal foot images to predict severity. In Phase 2 (External Validation & Usability Assessment), 30 independent participants underwent both radiographic and app-based severity assessments. Diagnostic times were recorded for both assessments. Participants then completed a 10-item Likert-scale usability questionnaire, with internal consistency assessed using Cronbach’s α. Results: The CNN successfully classified HV severity based on radiographic ground truth and showed consistent performance on an external dataset. App-based assessment was on average approximately 12 minutes faster than radiographic evaluation (p < 0.001). Usability evaluation indicated positive user experience (overall mean = 3.84/5, Cronbach’s α = 0.706). Conclusion: This study presents fully operational mobile AI application that enables rapid, accurate, and user-friendly classification of HV severity directly from smartphone photographs. By combining machine learning with an accessible mobile platform, it can support point-ofcare screening, patient self-monitoring, and community-based care where radiographic evaluation is impractical.