Evaluating synthetic anterior segment images for transferable eye disease recognition

To evaluate whether prompt-driven synthetic anterior segment disease (ASD) images preserve disease-relevant semantics for transferable ASD recognition.

A total of 17,853 slit-lamp images from eight ASD categories were retrospectively collected. Expert ophthalmologists provided lesion annotations to construct aligned image-text pairs for fine-tuning a Stable Diffusion-based (SD) model. Generated images were evaluated by expert quality control, Turing test, Fréchet Inception Distance (FID), and Kernel Inception Distance (KID). Their diagnostic utility was assessed on an independent test set of 5,466 images using precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Performance was evaluated across three models, compared with a real-image baseline, and assessed under different synthetic data scales. Feature separability and interpretability were further examined using t-SNE and Grad-CAM.

The SD model generated high-fidelity and semantically consistent ASD images guided by textual prompts. Expert assessment, Turing test, and FID/KID analyses supported the quality of the synthetic images. On test set, the classifier achieved a micro-AUC of 0.977 (95% CI, 0.975-0.979) and a macro-AUC of 0.979 (95% CI, 0.977-0.981), with F1-scores ranging from 89.28% to 99.52% across disease categories. Synthetic-image training outperformed the real-image baseline, and performance improved with increasing numbers of generated training images. t-SNE visualization demonstrated clear inter-class feature separation, and Grad-CAM highlighted lesion regions.

Prompt-driven synthetic anterior segment images preserved disease-relevant semantics and supported synthetic-to-real transfer in eye disease recognition, suggesting their potential value as complementary training resources for clinical artificial intelligence.