Annotated intraoral image dataset for dental caries detection

Abstract

This study introduces the first publicly available annotated intraoral image dataset for Artificial Intelligence (AI)-driven dental caries detection, addressing the lack of available datasets. It comprises 6,313 images collected from individuals aged 10 to 24 years in Mithi, Sindh, Pakistan, with annotations created using LabelMe software. These annotations were meticulously verified by experienced dentists and converted into multiple formats, including YOLO (You Only Look Once), PASCAL VOC (Pattern Analysis, Statistical Modeling, and Computational Learning Visual Object Classes), COCO (Common Objects in Context) for compatibility with diverse AI models. The dataset features images captured from various intraoral views, both with and without cheek retractors, offering detailed representation of mixed and permanent dentitions. Five AI models (YOLOv5s, YOLOv8s, YOLOv11, SSD-MobileNet-v2, and Faster R-CNN) were trained and evaluated, with YOLOv8s achieving the best performance (mAP = 0.841 @ 0.5 IoU). This work advances AI-based dental diagnostics and sets a benchmark for caries detection. Limitations include using a single mobile device for imaging. Future work should explore primary dentition and diverse imaging tools.

Fig. 1

The data production process, including data collection, data filtering and annotations. W/R: With Retractor, W/O-R: Without Retractor.

Fig. 2

The folder structure of the dataset available on Zenodo. W/R: With Retractor, W/O-R: Without Retractor, YOLO: You Only Look Once, PASCAL VOC: Pattern Analysis, Statistical Modeling, and Computational Learning Visual Object Classes, COCO: Common Objects in Context, JSON: JavaScript Object Notion.

Fig. 3

The folder structure of the benchmark dataset. W/R: With Retractor, W/O-R: Without Retractor, YOLO: You Only Look Once, PASCAL VOC: Pattern Analysis, Statistical Modeling, and Computational Learning Visual Object Classes, COCO: Common Objects in Context, JSON: JavaScript Object Notion.

Fig. 4

Various views of intraoral images, both with and without the use of cheek retractors.