Early diagnosis of oral potentially malignant disorders (OPMDs) may be hindered by similar clinical presentations shared between benign oral lesions and OPMDs. The goal of this retrospective pilot study was to assess the use of machine learning (ML) as an adjunctive evaluation in conjunction with conventional comprehensive oral examination of OMPDs. Digital images of 80 deidentified intraoral lesions (40 benign intraoral lesions and 40 OPMDs) were collected. The images, which were previously identified independently by experienced oral pathologists, were used to create 3 datasets: raw images, grayscale images, and enhanced color images. The datasets were subsequently divided into training (n = 60), test (n = 10), and validation (n = 10) groups so that class labels (benign lesion or OMPD) were distributed equally in each group. A cross-validated grid search was used to optimize the hyperparameters of the Extreme Gradient Boosting (XGBoost) classifications model. Predictions were made on the test group and used to optimize the prediction threshold. The final results were validated by predictions based on the validation group. The XGBoost classification model was able to differentiate between benign intraoral lesions and OPMDs with a mean classification accuracy of 70%, sensitivity of 80%, and specificity of 60% when grayscale and enhanced color intraoral images were used. A mean classification accuracy of 50%, sensitivity of 40%, and specificity of 60% were observed when raw intraoral images were used. The results demonstrated that ML may be a promising tool for the diagnosis of OPMDs when used as an adjunct to conventional comprehensive oral examination.
Keywords: early diagnosis; leukoplakia; machine learning; oral diagnosis; oral potentially malignant disorders; retrospective studies.