A multi-label deep learning system for simultaneous detection of nine fundus conditions.

Abstract

Purpose: To develop and validate a deep learning model for simultaneous detection of nine fundus conditions from a single color fundus photograph.

Methods: A development dataset of 236,242 color fundus images was assembled from 17 heterogeneous sources and partitioned at the patient level into training (70%), tuning (10%), and internal validation (20%) sets; 5-fold cross-validation was applied on the training set. Nine target conditions were labeled using standardized or photographic criteria: diabetic retinopathy (DR), age-related macular degeneration (AMD), myopic macular degeneration (MMD), glaucoma suspect (GS), epiretinal membrane (ERM), retinal vascular occlusion (VO), media opacity, retinal hemorrhages, and any retinal disorder (composite). A multitask ConvNeXt architecture was trained end-to-end, with operating thresholds pre-specified via Youden's index on the tuning set. External validation was performed on an independent cohort of 4055 images from an East Asian health-screening center, excluded from model development.

Results: On internal validation (N = 47,229), AUROCs ranged from 0.943 (AMD) to 0.989 (VO), with sensitivity of 88.1%-97.1% and NPV ≥ 98.8% for eight of nine conditions. On independent external validation (N = 4055), AUROCs ranged from 0.894 (retinal disorder) to 0.983 (MMD), with NPV ≥ 95.6% for eight of nine conditions, including 99.3% for DR, 99.9% for VO, and 99.8% for MMD.

Conclusions: This study presents a validated, criteria-anchored multi-label fundus triage framework capable of simultaneously screening for nine conditions from a single fundus photograph, with performance maintained on external validation at a source-excluded health-screening site within the same national health-screening system.