Photorealistic synthesis of oral lichen planus and lichenoid lesions enhances deep-learning segmentation in intra-oral photographs.

Medical image analysis often struggles with limited data and scarce expert annotations, particularly for oral potentially malignant disorders (OPMDs). We address this problem with a hybrid augmentation strategy that pairs CycleGAN image-to-image translation with a lightweight multi-scale blending refinement, termed the Realistic Enhancement Algorithm (REA). REA fuses CycleGAN-generated textures of oral lichen planus and oral lichenoid lesions (OLP/OLL) with authentic buccal-mucosa photographs, preserving anatomical detail while reducing common synthesis artifacts such as color drift, seams, and blur. A large-scale perceptual study involving 520 dental professionals found the synthetic images difficult to distinguish from real lesions (overall discrimination accuracy 58.22%; chance = 50%), supporting clinical plausibility. Using a leakage-controlled five-fold cross-validation protocol, we benchmarked 20 decoder-encoder segmentation combinations and then evaluated synthetic augmentation by adding synthetic images exclusively to the training split (never to validation or test). Incorporating synthetic lesions improved segmentation performance and robustness, with the largest gains observed in previously under-performing architectures (e.g., up to +2.08 points in micro mIoU and +2.31 points in macro sensitivity) while maintaining high overall specificity (>97.5%). Although the present work focuses on buccal-mucosa photographs, the approach is directly extendable to other intra-oral sites. To our knowledge, prior work has rarely addressed photorealistic synthesis of OLP/OLL lesions in intra-oral photographs with large-scale expert realism validation; our results suggest that clinically validated generative augmentation can alleviate data scarcity and strengthen AI-driven tools for oral healthcare.