Related Experiment Video
Updated: Mar 27, 2026
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
Artificial Intelligence-Driven Segmentation of Three Oral Diseases: Enabling Precision Diagnosis and Decision Support
Baowen Cheng1, Meijuan Xiang2, Zhuofan Zhao1
1Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai, 200050, China.
Introduction:
The advancement of AI-assisted oral disease diagnosis critically relies on the availability of annotated, high-quality datasets and resilient deep learning architectures. Nevertheless, progress in this domain has been impeded by the scarcity of publicly accessible data and inadequate benchmarking of existing models. To address these limitations, we introduced a meticulously curated dataset encompassing three prevalent oral mucosal diseases and undertook a thorough assessment of five deep learning models, focusing on their performance in concurrent lesion segmentation and classification.
Methods:
This study involved the compilation of a dataset comprising 808 high-resolution clinical images, encompassing three distinct oral mucosal pathologies: oral lichen planus, oral leukoplakia, and oral benign ulcers. Data collections were facilitated through clinical practices, and the images were subsequently annotated by three oral medicine specialists, who provided both diagnostic labels and pixel-level segmentation masks. To ascertain the dataset's utility and efficacy, five sophisticated deep learning architectures were benchmarked for their capacity to perform simultaneous segmentation and classification tasks.
Results:
Our benchmark evaluation indicated robust model performance on the proposed dataset. For the task of lesion segmentation, the optimal model achieved a mean Dice coefficient of 0.768 ± 0.031. Concurrently, an accuracy of 0.940 ± 0.013 was attained for disease classification. These findings collectively affirmed the dataset's substantial suitability for the development and rigorous assessment of AI models within authentic diagnostic contexts.
Conclusion:
The proposed dataset presented a standardised benchmark to facilitate the advancement of AI development in dental applications. Combined with rigorous deep learning validation, we provided a reproducible framework for precise lesion localisation and accurate disease identification.
Clinical Relevance:
The proposed integrated dataset-model pipeline facilitates the development of dependable AI tools capable of aiding clinicians in prompt diagnosis, mitigating diagnostic inconsistencies, and augmenting decision-making processes within standard oral health care practices.
