Text-to-Keratitis: Clinician-Aligned Multimodal Generation of Keratitis Images With Diffusion Models

Generating clinically realistic medical images holds great potential for medical education, diagnostic support, and improving model generalization in medical AI. In this work, we present a prompt-driven diffusion framework tailored for corneal disease image synthesis, capable of generating diverse and high-quality images across multiple disease types and imaging modalities. First, we incorporate a prompt-guided optimization module that refines text embeddings and enforces image-text alignment using CLIP-based supervision. Second, we propose a novel clinician-aligned reward optimization strategy, which integrates expert-provided quality assessments into the sampling process via reward gradients, thereby encouraging clinically plausible outputs. Extensive experiments conducted on our newly constructed corneal disease dataset, which covers five categories of keratitis, demonstrate that our method consistently outperforms existing approaches. Furthermore, classification networks trained with our generated images achieve improved diagnostic performance, confirming the effectiveness of the proposed framework and its practical value for data augmentation in medical AI.