Treffer: Cutting 10 Hours to 30 Seconds: Deep Learning Precision Cropping of Strabismus Eye Regions for Workflow Optimization and Standardized AI Model Training Data Preprocessing (Preprint)

BACKGROUND Privacy protection for strabismus patients necessitates de-identification of facial information in ocular gaze photographs used for AI model development, outpatient medical records, telemedicine, and academic exchange. This requires cropping full-face images to include only the eye region. However, conventional preprocessing methods are time-consuming, labor-intensive, and plagued by non-standardized practices, highlighting the need for an efficient and standardized solution. OBJECTIVE This study aimed to develop an AI-driven management platform for strabismus patients, leveraging preprocessing algorithms based on ocular gaze photographs. The platform uses advanced algorithms to optimize eye region cropping, improving accuracy, efficiency, and standardization in both clinical workflows and AI model training data preprocessing. METHODS This retrospective and prospective cross-sectional study used 5832 images from inpatients and outpatients, capturing three gaze positions under varying conditions. The model was evaluated using precision, recall, and mean average precision (mAP) at various Intersection over Union (IoU) thresholds through 5-fold cross-validation on an inpatient dataset and tested on an independent outpatient dataset. Expert validation confirmed alignment with clinical standards, and a control experiment with five optometry specialists compared manual and automated cropping efficiency for real-time applications. RESULTS The model achieved a mean precision of 1, recall of 1, mAP50 of 0.995, and mAP95 of 0.893 across the 5-fold cross-validation set. On the independent test set, the precision was 1, recall was 1, mAP50 was 0.995, and mAP95 was 0.801. Expert validation of 48 images confirmed adherence to clinical and research standards. A control experiment with five optometry specialists demonstrated a reduction in image preparation time from 10 hours to 30 seconds for 900 photos. CONCLUSIONS The AI-driven management platform for strabismus patients integrates an intelligent cropping algorithm, electronic archives for comprehensive patient histories, and a patient-physician interaction module. This platform not only optimizes the eye region cropping process, improving both accuracy and efficiency, but also offers a complete solution for strabismus care. The system accelerates clinical workflows with seamless uploads to electronic health records, while ensuring patient privacy during remote consultations. This innovative approach has the potential to significantly enhance data preparation, patient care, and clinical practices in ophthalmology. CLINICALTRIAL This retrospective and prospective cross-sectional study received approval from the Ethics Committee of West China Hospital, Sichuan University, China (2023 Review No. 1477).