Conditional Latent Diffusion for Controllable Palpebral Conjunctiva Image Generation Toward Hemoglobin Assessment

Background/Objectives: Deep learning-based medical imaging applications often require large and diverse datasets to achieve reliable performance. However, publicly available palpebral conjunctiva datasets for non-invasive hemoglobin assessment remain limited in both size and demographic diversity. The objective of this study was to develop and evaluate a conditioning-guided Latent Diffusion Model (LDM) for controllable palpebral conjunctiva image synthesis under limited-data conditions. Methods: The proposed framework employs an image-to-image latent diffusion strategy conditioned on continuous hemoglobin values together with gender and country information to generate realistic synthetic conjunctiva images with controllable clinical and demographic characteristics. The proposed LDM was compared with GAN-based approaches, including cDCGAN and StyleGAN2-ADA, using evaluation criteria covering image realism, diversity, conditioning consistency, and computational efficiency. Analyses of frequency-domain characteristics, zero-shot cross-population evaluation, and blinded clinical assessment were performed for the proposed LDM. Results: The proposed LDM achieved the lowest FID score (15.86±0.23) among the evaluated models, indicating superior image realism relative to cDCGAN and StyleGAN2-ADA, while maintaining image diversity comparable to StyleGAN2-ADA and substantially outperforming cDCGAN. Conditioning evaluation demonstrated strong consistency between the target hemoglobin values and the generated images, achieving a Pearson correlation coefficient of r=0.910±0.025. Frequency-domain analysis, zero-shot cross-population evaluation, and blinded clinical assessment further supported the realism, structural consistency, and clinical plausibility of the generated images. Conclusions: The findings demonstrate the potential of conditional latent diffusion models for controllable palpebral conjunctiva image synthesis under limited-data conditions. The proposed framework provides a promising approach for generating realistic synthetic conjunctiva images with controllable clinical and demographic characteristics.