Kaili Du, Yuqiang Zheng, Zhuoshi Wang
RetinalVNG-Net demonstrates promising robustness and generalizability for simultaneous multimodal risk stratification of vascular and neurodegenerative retinal changes across diverse devices and settings, supporting further evaluation as a tool for risk stratification of retinal manifestations associated with systemic vascular and neurodegenerative disease. Prospective longitudinal studies would be required to establish value for early or predictive detection.
BACKGROUND: We developed and validated RetinalVNG-Net, a multimodal deep learning framework for simultaneous risk stratification of hypertensive retinopathy, diabetic retinopathy, and neurodegenerative-associated retinal changes from integrated fundus photography, optical coherence tomography, and clinical metadata.
METHODS: This retrospective multi-center study included 2,740 subjects from three independent ophthalmology centers (January 2019-December 2023). Centers A and B (n = 2,220) constituted the development set, with a stratified 15% subset (n = 333) reserved for hyperparameter tuning and the remainder (n = 1,887) used for five-fold cross-validation. Center C (n = 520) served as a geographically distinct, device-heterogeneous external test set. RetinalVNG-Net employs a RETFound ViT-Large fundus encoder, a dual-stream Optical Coherence Tomography (OCT) branch (ResNet-3D-18 for volumetric B-scans and 2D-CNN for layer thickness maps), and a tabular transformer for metadata, fused via cross-modal attention. An auxiliary regression head outputs a continuous Retinal Biological Age Gap (RBAG) score as an interpretable severity biomarker.
RESULTS: Internal cross-validation yielded macro-averaged AUC-ROC 0.957 (±0.008), sensitivity 0.913, specificity 0.941, and F1 0.908 across four classes. On the external test set, macro-averaged AUC-ROC reached 0.944 (95% CI 0.922-0.958), with per-class AUCs of 0.963 (hypertensive retinopathy), 0.951 (diabetic retinopathy), 0.924 (neurodegenerative changes), and 0.938 (controls). RetinalVNG-Net significantly outperformed the best single-modality fundus model (macro-AUC 0.944 vs. 0.901; p < 0.001).
CONCLUSION: RetinalVNG-Net demonstrates promising robustness and generalizability for simultaneous multimodal risk stratification of vascular and neurodegenerative retinal changes across diverse devices and settings, supporting further evaluation as a tool for risk stratification of retinal manifestations associated with systemic vascular and neurodegenerative disease. Prospective longitudinal studies would be required to establish value for early or predictive detection.