Enhancing Community Vision Screening -- AI Driven Retinal Photography for Early Disease Detection and Patient Trust

ECVS, an AI-driven retinal photography solution, enhances community vision screening by detecting pathology-based visual impairment. It offers a one-stop workflow for large-scale settings, reducing traditional screening tests from five to two, chair time from 40 minutes to 5, and referral time from 2-4 weeks to 10-20 minutes. The system uses deep learning models like ResNet, EffcientNet, ViT, Swin Transformer, and RetFound for classification and U-Net for segmentation. It includes a photo quality assessment and a pathology visual impairment classification model. The ECVS framework comprises various models for detecting vision impairment, diagnosing eye diseases, and segmenting lesion regions. The RETQA model, using Swin V2, EfficientNet b7, and ResNet50, achieved high AUC scores on both internal and external datasets. The PVI models achieved high AUC scores on both internal and external datasets, enabling context-specific outputs with high sensitivity or specificity. The multi-label eye disease classification model, akin to a junior ophthalmologist's diagnosis, provided acceptable results during screening. The segmentation model's ROI heatmap, more aligned with human perception, showed promise with a DICE score of 0.482. The ECVS, comprising four deep learning models, was cost-efficient, with manageable computational overhead.