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

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Generating realistic medical images is crucial for medical education, AI diagnostics, and improving AI model generalization.
• Current methods may lack clinical accuracy and diversity across various corneal diseases and imaging types.

Purpose of the Study:

• To develop a prompt-driven diffusion framework for synthesizing diverse, high-quality corneal disease images.
• To enhance clinical realism and diagnostic utility of synthetic medical images.

Main Methods:

• Implemented a prompt-guided optimization module with CLIP-based supervision for text-image alignment.
• Introduced a clinician-aligned reward optimization strategy integrating expert quality assessments into image generation.
• Utilized a newly constructed corneal disease dataset covering five keratitis categories.

Main Results:

• The proposed framework consistently outperformed existing methods in generating corneal disease images.
• Synthetic images generated by the framework improved the diagnostic performance of classification networks.
• Demonstrated the framework's effectiveness for data augmentation in medical AI.

Conclusions:

• The prompt-driven diffusion framework effectively generates clinically realistic corneal disease images.
• The clinician-aligned reward optimization enhances the plausibility and utility of synthetic medical data.
• The framework shows significant practical value for advancing medical AI and education.