Abstract / Summary
Early detection and intervention of strabismus are crucial for effective treatment, yet traditional diagnostic methods require extensive expertise, often causing diagnostic delays in underserved areas. In this study, we propose and validate a “two-tier” stratified diagnostic system, supported by the largest known dataset of strabismus gaze photographs to date (84,897 photographs from 9433 individuals for cross-validation and 10,458 photographs from 1162 individuals for prospective testing). This system serves two primary purposes: (1) enabling accessible home-based screening via primary-gaze photographs, and (2) assisting community physicians in making accurate preliminary diagnoses using Strabismus-Net, a specialized nine-gaze deep learning architecture incorporating dual-stream cross-attention and adaptive feature aggregation. The screening model achieved an area under the curve (AUC) of 99.53%. For subtype classification, Strabismus-Net achieved AUCs of 97.15% for orthotropia, 99.19% for exotropia, 99.11% for esotropia, and 90.20% for vertical strabismus. Statistically benchmarked against clinicians, the model performed comparably to three senior specialists across all subtypes ( P > 0.05). In clinically challenging scenarios—specifically identifying vertical strabismus and confirming orthotropia—it significantly outperformed three junior residents ( P < 0.001 and P < 0.05, respectively). Together with an integrated management platform, this digital health solution achieves state-of-the-art performance, optimizing workflows and streamlining specialist referrals.