RetiFlow: Generative Completion and Synthesis of Multimodal Retinal Imaging for Inherited Retinal Disease Phenotyping

Inherited retinal diseases (IRDs) are a leading cause of blindness in children and working-age adults, and an increasing number are becoming treatable. Multimodal retinal imaging is routinely acquired and informative, yet conventional classifiers struggle with incomplete examinations, which are common in clinical practice, and cannot reconstruct missing modalities or augment underrepresented classes. We introduce RetiFlow, a multitask latent generative model for multimodal retinal phenotyping. Four frozen variational masked autoencoders compress color fundus photography, wide-field and macular autofluorescence, and OCT into compact latent representations. A single flow-matching model then learns their joint distribution conditioned on pathology, gene, and laterality. The same model reconstructs missing modalities, synthesizes class-conditional images on demand, and supports pathology and gene classification. On a held-out cohort of 129 patients, modality completion improves weighted-AUROC for gene prediction from 0.8713 to 0.8898, while class-conditional synthesis reaches 0.8906. Averaging the predictions from the two configurations yields the best performance (0.8998) while requiring no additional training. RetiFlow thus unifies completion, synthesis, and phenotyping within a single framework, offering a practical path toward earlier, more reliable IRD diagnosis.