Deep Learning-Based Screening of Alport Syndrome Using Retinal Optical Coherence Tomography Images

Background: Prior studies indicate that temporal retinal thinning demonstrated by optical coherence tomography (OCT) is associated with Alport syndrome (AS). We aim to develop a deep-learning (DL) algorithm (termed DEEP-ASK) to differentiate between retinal images of individuals with AS vs other causes of glomerular hematuria. Methods: Data sets of three-dimensional (3D) OCT scans were collected from patients in the Shanghai Registry of Alport syndrome. The structure of the DEEP-ASK network is constructed upon a contrastive DL model pre-trained on 2D scans and a transformer framework. The area under the curve (AUC) was calculated with the probabilities for overall prediction results of the DEEP-ASK model. Results: 21120 OCT scans images from 93 AS patients (mean age 33.44±15.80), along with age- and sex-matched control participants (48 glomerulonephritis patients with primary cause and 24 normal subjects) collected between June 2024 to May 2025 were included. The DEEP-ASK model achieved a mean AUC of 0.95 for AS diagnosis, outperforming that of the temporal retinal thickness index (0.63). Conclusion: Our preliminary results support the potential value of DEEP-ASK model in noninvasive screening and early diagnosis of AS. Funding: Clinical Revenue SupportFigure:Overall study design of the development and validation of the DEEP-ASK system