Akademik Veri Yönetim Sistemi
Journal of the College of Physicians and Surgeons Pakistan, cilt.36, sa.4, ss.526-530, 2026 (SCI-Expanded, Scopus)
Objective: To assess the effectiveness of deep learning-based diagnostic software in identifying oral leukoplakia (OL) lesions from intraoral patient photographs. Study Design: An observational study. Place and Duration of the Study: Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Marmara University, Istanbul, Turkiye, from February to November 2024. Methodology: A dataset comprising 222 anonymised retrospective intraoral images, confirmed as OL via incisional biopsy, was labelled using the polygonal annotation tool in CranioCatch software. Expert reviewers validated all annotations. The dataset was split into training (n = 178), validation (n = 22), and test (n = 22) subsets. The models used for lesion detection included YOLOv5, YOLOv8, Mask R-CNN, and U-Net. The model's effectiveness was measured using the confusion matrix. Precision, recall, and F1-score performance provided quantitative assessments of the model's predictive capabilities. Results: YOLOv5 demonstrated the highest performance with an F1-score of 0.845, sensitivity of 0.967, and precision of 0.750. YOLOv8 followed closely with an F1-score of 0.760, sensitivity of 0.871, and precision of 0.675. In contrast, Mask R-CNN and U-Net showed comparatively lower effectiveness, achieving F1, sensitivity, and precision values of 0.408, 0.588, and 0.312 for Mask R-CNN, and 0.629, 0.518, and 0.800 for U-Net, respectively. Conclusion: YOLOv5 and YOLOv8 models outperformed Mask R-CNN and U-Net in accurately detecting OL lesions. These findings highlight the promise of artificial intelligence in facilitating early diagnosis of potentially malignant oral conditions. As datasets expand and algorithms improve, the clinical utility of such AI systems is expected to grow, contributing to timely intervention and improved patient outcomes.