OCTFlow: Unified conditional generation for balanced classification and annotated segmentation in retinal OCT

The scarcity of annotated medical images, compounded by severe class imbalance and labor-intensive pixel-wise annotation, fundamentally constrains the development of robust diagnostic systems in retinal OCT analysis. To address these challenges, we propose OCTFlow, a unified conditional synthesis framework for retinal OCT augmentation under class-label or segmentation-mask guidance. Unlike conventional Latent Diffusion Models (LDMs) that use Variational Autoencoders (VAEs) for perceptual compression, OCTFlow operates in the representation space of a frozen ophthalmic foundation model such as RETFound, incorporating domain-specific retinal representations into latent generation. The framework operates consistently across a range of these encoders. We then employ latent Flow Matching (FM) for efficient conditional synthesis. For task-specific conditioning, we introduce two strategies: (1) Representation-Space Classifier Guidance (RSCG), which uses gradients from a noise-aware latent classifier to encourage target-class generation for underrepresented diseases; and (2) Dynamic Region Masking (DRM), which applies stochastic pathology-mask perturbations to promote diverse paired image-mask synthesis. Extensive experiments on a large-scale curated multi-source OCT classification dataset and two segmentation benchmarks demonstrate that OCTFlow performs best on the reported synthesis metrics among the evaluated generative baselines. More importantly, augmenting the training sets with OCTFlow-synthesized samples improves downstream performance under the evaluated protocols, particularly for rare disease identification and pathological lesion segmentation. These results support OCTFlow as an effective augmentation framework for data-constrained retinal OCT analysis. Our code is publicly available at https://github.com/Endoscope-Lab/OCTFlow.git.