'All-in-One' Screening for Pathology-Related Visual Impairment using Artificial Intelligence-Integrated Retinal Imaging.

Abstract

Purpose : Efficiency and coverage remains key challenges for successful implementation of community vision screening programs. The emergence of deep learning (DL) technology offers new opportunities to revolutionize clinical practice. This study aimed to evaluate the performance of a newly developed DL system (DLS) for detection of pathology-related visual impairment (VI).

Methods : We retrospectively included 17,547 individuals from 4 population-based studies, namely the Singapore Epidemiology of Eye Diseases (SEED) Study, Beijing Eye Study (BES), Central India Eye and Medical study (CIEMS), and Blue Mountains Eye Study (BMES). Macular-centred retinal photos and best-corrected visual acuity (BCVA) were collected. Pathology-related VI was defined as BCVA worse than 20/40. 7,793 individuals’ data (15,175 fundus images) from the SEED study were used for development of the DLS. ResNet and DenseNet convoluted neural networks (CNNs) were used for training of the DLS model. Another independent set of 1,950 individuals’ data (3,803 fundus images) from SEED study were used as primary validation. The BES (n=3,323; 6,284 images), CIEMS (n=3,444; 6,554 images), and BMES (n=1,037; 2,031 images) were used as external validation datasets. DLS performance in detection of pathology-related VI was evaluated based on area under the curve (AUC), sensitivity, and specificity levels.

Results : In the primary validation dataset, the AUC for detection of pathology-related VI was 0.92 (95% confidence interval [CI], 0.90-0.93) with sensitivity of 85.8% (95% CI, 81.7-89.3%), and specificity of 85.8% (95% CI, 84.6-86.9%). Across the external validation datasets, the BES, CIMES and BMES recorded AUCs of 0.90 (95% CI, 0.89-0.92), 0.85 (95% CI, 0.84-0.87) and 0.81 (95% CI, 0.77-0.84), respectively. Furthermore, BES showed sensitivity of 83.2%, specificity of 83.9%; CIEMS showed sensitivity of 68.7%, specificity of 89.6%; and BMES showed sensitivity of 66.5% and specificity of 82.5%. Based on the generated saliency maps, the regions on fundus photo in which the CNN model focused on for prediction, were consistent with typical clinical observations made by ophthalmologists (Figure 1 & 2).

Conclusions : We developed a novel DLS which shows promising performance as an 'all-in-one screening’ for pathology-related VI. This new screening modality will help to improve current screening capacity and coverage among elderly.

This is a 2020 ARVO Annual Meeting abstract.