A Clinical Image Dataset for AI-Based Segmentation of Three Oral Diseases

The rapid progress of deep learning has significantly improved AI-aided medical diagnosis, where model performance critically relies on high-quality annotated datasets. Regrettably, such datasets remain scarce for oral diseases due to obstacles in standardized data collection and expert annotation. To mitigate this gap, we introduce a comprehensive dataset named OralTriD, encompassing 808 high-resolution clinical images covering three prevalent oral diseases: oral lichen planus, oral leukoplakia and oral benign ulcers. The dataset comprises images annotated by board-certified dental specialists, providing diagnostic classifications and precise pixel-level lesion segmentation masks. The dataset captures a diverse range of clinical manifestations across disease stages and anatomical sites, with balanced representation of all pathological categories. Benchmark evaluations employing five state-of-the-art deep learning models demonstrate the dataset’s efficacy, achieving a Dice coefficient of 0.7947 for segmentation tasks and a classification accuracy of 96.0%. This resource supports development of robust AI tools for simultaneous lesion localization and disease identification, while serving as a standardized benchmark for oral healthcare research.