Automated Diagnosis of Oral Diseases Using ROI Detection and Transfer Learning on Photographic Dental Images

Resumo

The early and precise diagnosis of oral diseases is essential for effective treatment and improved patient outcomes. This research presents a novel deep learning–based framework that leverages region of interest (ROI) detection and transfer learning to automate the classification of six common dental conditions using high-resolution photographic images. A dataset of 1,800 images, annotated by certified dental practitioners, was compiled from local clinics. The proposed pipeline begins with a binary classifier to determine whether ROI localization is required. For images needing localization, a lightweight YOLOv4-tiny detector was employed to isolate the relevant regions. These cropped ROIs were then used to fine-tune multiple pre-trained convolutional neural networks (CNNs), including VGG16, VGG19, InceptionV3, DenseNet201, and ResNet50. Among these, the VGG19 model achieved the high performance, with an accuracy of 95.85%, precision of 95.47%, and recall of 96.18%. The results demonstrate that combining efficient ROI detection with transfer learning significantly enhances diagnostic accuracy while maintaining computational efficiency, making the approach suitable for integration into mobile and clinical diagnostic systems. This work highlights the potential of AI-driven dental diagnostics to support early detection, reduce consultation delays, and improve accessibility to oral healthcare.