MOPG-7: A Multi-Clinic Dental Panoramic Radiograph Dataset with Expert YOLO Labels

Abstract

This dataset is a publicly available, multi-clinic dental imaging database that aims at furthering studies on artificial intelligence (AI), computer vision, and computer-aided diagnosis (CAD) based on panoramic dental radiographs (orthopantomogram or OPG). This database includes 2,095 completely anonymized panoramic dental radiographs gathered retrospectively from four separate dental centers in Bangladesh, namely Sonia Nursing Home, Tangail (1,433 radiographs); Ibn Sina D. Lab & Consultation Center, Dayaganj (533 radiographs); Niramoy Diagnostic Center, Tangail (114 radiographs); and Health City Diagnostic Center, Gaibandha (15 radiographs). Every image comes with bounding box annotations in both YOLO (.txt) and standard COCO JSON format, validated by experts, making the dataset easy to use with frameworks such as Ultralytics YOLO as well as any COCO-compatible model or data loader. Initial bounding boxes were drawn by a licensed dentist and independently reviewed by a second dental professional. A rigorous quality-control process followed, removing 127 substandard bounding boxes, resulting in 9,834 validated bounding boxes. Dataset Classes The dataset includes annotations for seven clinically relevant dental categories: - Missing Teeth: 2,609 annotations - Dental Crown: 1,984 annotations - Root Canal: 1,956 annotations - Caries: 1,410 annotations - Wisdom Teeth: 869 annotations - Broken Down Teeth: 795 annotations - Healthy OPG: 211 annotations Dataset Contents The released dataset includes: - Panoramic dental radiographs (.jpg) - COCO-format JSON annotation files (instances_train.json, instances_val.json, instances_test.json) - YOLO bounding-box annotation files (.txt) - Metadata files (class_definitions.csv, image_metadata.csv, split_assignments.csv) - Documentation (README.md) describing the dataset structure, annotation format, and usage instructions Potential Research Applications This dataset supports multi-class dental object detection, localization of dental abnormalities, CAD, deep learning for medical imaging, computer vision research, transfer learning and foundation models, explainable AI (XAI), medical image analysis, object detection benchmarking, AI-enabled dental diagnosis, dental AI learning, and reproducibility research. Benchmark Performance Three YOLO variants were evaluated to establish a baseline. YOLOv8m achieved the highest precision (71.9%) and mAP@0.5 (72.9%), with the fastest inference latency (1.9 ms per radiograph). YOLOv10m achieved the highest recall (74.2%) and mAP@0.5:0.95 (34.4%), with a latency of 3.4 ms. YOLOv11m reached a precision of 70.9%, recall of 71.6%, mAP@0.5 of 71.7%, and mAP@0.5:0.95 of 33.5%, with a latency of 4.2 ms. These results show the comparative detection performance and computational efficiency of the evaluated YOLO variants on this dataset.