Photorealistic Synthesis of Oral Lichen Planus Lesions: Code and Example Dataset

This dataset accompanies the manuscript "Photorealistic Synthesis of Oral Lichen Planus and Lichenoid Lesions Enhances Deep-Learning Segmentation in Intra-Oral Photographs" submitted to Computerized Medical Imaging and Graphics.

Research Context: Medical image analysis for oral potentially malignant disorders (OPMDs) is hindered by limited annotated datasets. This work addresses data scarcity by combining CycleGAN with a Realistic Enhancement Algorithm (REA) to synthesize photorealistic oral lichen planus/lichenoid lesion (OLP/OLL) images on healthy buccal mucosa photographs.

What This Dataset Contains:

1. codes.zip — Source code for the CycleGAN+REA synthesis pipeline: buccal mucosa segmentation, CycleGAN lesion generation, REA histogram matching, Laplacian pyramid blending, QC screening, and segmentation inference scripts.
2. models.zip — Pretrained weights: CycleGAN generator, TensorFlow-based buccal mucosa segmentation, OLP lesion segmentation (QC), and three Experiment 1 architectures (MANet+MiT-B2, FPN+MiT-B0, PSPNet+EfficientNet-B0).
3. data.zip — De-identified example images: healthy buccal mucosa inputs and real lesion images with ground-truth masks.

Key Findings:

* REA combines masked histogram matching with Laplacian pyramid blending to reduce CycleGAN artifacts (color drift, boundary discontinuities)
* 520 dental professionals achieved only 58.22% discrimination accuracy (chance=50%), indicating high realism
* Synthetic augmentation improved segmentation by up to +2.08 mIoU points

How to Use:
1. Unzip codes.zip, data.zip, and models.zip
2. Install dependencies from requirements_simplified.txt
3. Run realistic_lesion_pipeline.py for synthesis
4. Run oral_lesion_segmentation.py for segmentation inference

Data Restrictions: Due to ethics restrictions, the full clinical dataset cannot be released. This repository provides minimal examples to demonstrate pipeline functionality.

Associated Publication: Theppitak S, Wongsapai M, Jaidee E, Sakdapreecha C, Ittichaicharoen J, Pongsiriwet S, Warin K, Suebnukarn S, Wuttisarnwattana P. "Photorealistic Synthesis of Oral Lichen Planus and Lichenoid Lesions Enhances Deep-Learning Segmentation in Intra-Oral Photographs", Computerized Medical Imaging and Graphics. [Under Review]