Automated Oral Lesion Detection and Classification Using Computer Vision
Oral lesion detection; YOLOv8; EfficientNet-B3; Grad-CAM; Oral cancer screening; Explainable AI; Tele-dentistry
Authors:AbstractEarly diagnosis of oral lesions is a critical determinant of clinical outcomes in oral cancer management. Conventional screening relies on expert examination, which is prone to inter-observer variability and limited access in rural settings. This study proposes a hybrid deep learning framework integrating YOLOv8 for lesion localization with EfficientNet-B3 for multiclass classification. The pipeline automatically identifies suspicious intraoral lesions and categorizes them as Normal, Benign, Premalignant, or Malignant. Gradient-weighted Class Activation Mapping (Grad-CAM) provides discriminative heatmaps for clinical interpretability. Experiments on 5,200 annotated intraoral images yielded a detection mAP@50 of 96.8%, classification accuracy of 96.1%, and AUC-ROC of 98.2%, outperforming six state-of-the-art baselines. Ablation studies confirm the contribution of each component. The framework is suitable for tele-dentistry deployment with an inference time of 22 ms per image. ••••••••••••••••••••••••••••••• ejprd.org - Published by Riset Publication Services LLC
EJPRD
Copyright ©2026 by RisetPublicationServicesLLC