Evaluation of the Detectability of Oral Potentially Malignant Diseases with a Deep Learning Approach: A Retrospective Pilot Study.

Abstract

Oral potentially malignant diseases (OPMD) may arise during the malignant transformation of the oral mucosa, with cellular changes in these lesions increasing the likelihood of cancer development compared to normal tissues. This study aims to evaluate the performance of a deep learning-based diagnostic software designed to detect OPMD. A total of 358 anonymized retrospective intraoral images from patients histopathologically diagnosed with oral lichen planus, oral leukoplakia, or oral cancer via incisional biopsy were used. The images were annotated using the polygonal labeling method in CranioCatch software (CranioCatch, Eskişehir, Turkey) and reviewed by Oral, Dental, and Maxillofacial Radiologists. The dataset was divided into training (n = 288), validation (n = 35), and test (n = 35) sets. A deep learning model based on the YOLOv8 architecture was developed, and its performance was assessed using a confusion matrix. The model achieved an F1 score of 0.693, a sensitivity of 0.666, and a precision of 0.723. These findings suggest that deep learning and artificial intelligence show promise in the diagnosis of OPMD and that routine oral examinations and early detection of these lesions-especially in high-risk individuals-are essential responsibilities for dental professionals. Larger, multi-center datasets, calibration, and external validation are needed for clinical translation.