Comparative Study of Deep Learning Models for Dental Caries Detection

Abstract

Dental caries is a dental disease that is considered a global public health problem and requires detection that is friendly to remote areas.  This study presents a comparison of the evaluation results of deep learning models for detecting dental caries from early to extensive levels using YOLOv11, Faster R-CNN and RetinaNet models. The dataset contains 1,036 images divided into 4 classes (healthy teeth, early caries, moderate caries and extensive caries). As a result, YOLOv11 produced the highest mean average precision (mAP) of 79.2%. In addition, balanced precision (70.9%), recall (76.6%) and f1 score (73.6%), high average precision (AP) per class (healthy teeth: 85.5%, early caries: 66.9%, extensive caries: 91.6% and moderate caries: 72.6%), a 5.6 ms inference time and 5 MB model size are featured by YOLOv11 which is suitable to be implemented into various devices to support medical personnel in detecting dental caries in remote areas.