Topology-Aware Conditional Latent Diffusion for Multi-View Fundus Image Synthesis

Fundus imaging in preterm infants is critical for diagnosing retinopathy of prematurity (ROP) but requires capturing multiple fields-of- view (FOVs), imposing significant distress on infants and operational burdens on clinicians. To address this, we propose a Topology- Aware Conditional Latent Diffusion Model (TA-CLDM) that synthesizes patient-specific, multi-view fundus images from a single input, conditioned on textual FOV prompts (e.g., "left eye, nasal view"). Our framework integrates a latent diffusion model (LDM) with a topology-aware loss that leverages persistent homology to preserve vascular continuity-ensuring synthesized vessels retain diagnostically critical structures like branching patterns and connectivity. Trained on a dataset of 2,862 preterm infant fundus images (954 patients, 3 FOVs/patient), TA-CLDM achieves a Fréchet Inception Distance (FID) of 47.93 outperforming baseline LDMs by 19.91 FID. By reducing the need for repeated imaging, TA-CLDM alleviates infant distress and institutional costs while maintaining diagnostic accuracy. This work pioneers the integration of text-guided synthesis and topological preservation in medical imaging, offering a scalable solution for neonatal care.