Abstract
BACKGROUND: Strabismus is a common ocular misalignment that can impair binocular vision if untreated. Conventional diagnosis and treatment rely on clinical prism diopter (PD) readings, which quantify deviation along with base direction. However, these values are coarse, manual, and subject to inter-clinician variability. METHODS: We present the development and results of StrabNet-CQ (Strabismus Network for Classification and Quantification), a newly developed deep learning-based framework for automated strabismus classification and quantification. Six hundred eye images, with and without strabismus, were analyzed by the model. A YOLOv8 model performs classification into normal and abnormal as well as subsequent classification (normal, esotropia, exotropia, hypertropia, hypotropia), with refined classification via ResNet101 on segmented eye regions from the images. Ocular landmarks are detected using ResNet18, from which horizontal and vertical deviation indices and angular deviation are computed. RESULTS: The system achieved 94% accuracy in strabismus detection and 90% in strabismus classification, with high sensitivity for normal (0.91-1.00), esotropia (0.89), and hypotropia (0.90). The derived parameters show a good correlation with manual PD values(r = 0.733) that can be utilized for quantification of strabismus. CONCLUSION: StrabNet-CQ supports objective diagnosis and holds promise for deployment in clinical settings for strabismus detection as well as quantification.