This presentation will explore the significant potential of AI and deep learning (DL) technologies in screening for diabetic retinal disease (DRD) through ophthalmic imaging modalities such as fundus photography (FP) and optical coherence tomography (OCT). I will share findings from our previous research, which demonstrated the robust performance of DL models in detecting referable diabetic retinopathy (DR) from FPs during external validation using datasets from Hong Kong. Moreover, I will present our development of a three-dimensional (3D) DL multi-task model that effectively identifies and classifies diabetic macular edema (DME) into centre-involved DME (CI-DME) or non-centre-involved DME (non-CI-DME), with excellent performance across various OCT devices and imaging protocols during external testing. Despite the promise of AI and DL technologies for DRD screening, their widespread implementation in Hong Kong faces several challenges. In this talk, I will discuss these challenges, including discrepancies between laboratory and real-world performance, the lack of transparency and explainability of current black-box AI models, and financial sustainability. I will also share our ongoing research efforts to generate evidence supporting the role of AI in DRD screening in real-world settings.
