Multimodal OCT–OCTA Classification of Dry AMD, Active MNV, and Inactive MNV Using Frozen Vision Transformer Features: A Patient-Level Nested Cross-Validation Study

Background/Objectives: Distinguishing dry age-related macular degeneration (AMD) from active and inactive macular neovascularization (MNV) may benefit from integrated interpretation of optical coherence tomography (OCT) and OCT angiography (OCTA). However, the added value of OCTA under artifact-prone conditions remains uncertain. We compared OCT-only, direct OCTA fusion, and OCT–OCTA multimodal classification using frozen Vision Transformer features. Methods: This retrospective study included 368 eyes from 200 patients. OCT B-scans and outer retina and choriocapillaris OCTA images were analyzed. A pretrained Vision Transformer B/32 (ViT-B/32) generated modality-specific embeddings. Class-weighted linear support vector machine (SVM) models were developed using patient-grouped repeated nested cross-validation (10 repetitions; five outer and four inner folds). The primary gradable cohort comprised 320 eyes from 185 patients; severe OCTA artifacts and non-exudative MNV were assessed separately in predefined challenge cohorts. Results: Direct multimodal fusion achieved an accuracy of 0.753, balanced accuracy of 0.679, macro-F1 of 0.685, and macro area under the receiver operating characteristic curve (AUROC) of 0.887. It significantly outperformed direct OCTA fusion across five prespecified metrics; differences from OCT alone were not statistically significant. Class-specific sensitivities were 0.916 for dry AMD, 0.709 for active MNV, and 0.412 for inactive MNV. Performance estimates were similar in the real-world cohort (accuracy, 0.758; macro-AUROC, 0.886). Whole-modality occlusion suggested predominant OCT reliance for dry AMD and greater outer retina OCTA reliance for active and inactive MNV. Conclusions: Frozen Vision Transformer-derived multimodal features supported three-class OCT–OCTA classification under patient-level validation. Structural OCT provided most of the discriminative information. OCTA influenced model predictions in occlusion analyses, but its incremental classification benefit was not established. External multicenter validation is required before clinical use.