Dataset for Automating Dental Condition Detection on Panoramic Radiographs

Should you use the dataset in any scientific papaer, please also cite the paper it is derived from: Mureșanu, S., Hedeșiu, M., Iacob, L., Eftimie, R., Olariu, E., Dinu, C., Jacobs, R., & on behalf of Team Project Group. (2024). Automating Dental Condition Detection on Panoramic Radiographs: Challenges, Pitfalls, and Opportunities. Diagnostics, 14(20), 2336. https://doi.org/10.3390/diagnostics14202336 Panoramic Image Selection and Dataset The training dataset consisted of 1628 panoramic images from the Department of Maxillofacial Surgery and Radiology at the Iuliu Hațieganu University of Medicine and Pharmacy, Cluj-Napoca, Romania. The images were retrospectively collected from the department’s database of patients who visited the hospital between April 2021 and December 2023. The images were acquired using a Vatech PCH-2500 machine (Vatech, Hwaseong, Republic of Korea) and stored in jpeg format. A second set consisting of 180 radiographs was acquired from different medical centers in Cluj-Napoca, Romania, and was used as external validation. Patient consent was waived due to the retrospective nature of the study by the Ethics Committee of the Iuliu Hațieganu University of Medicine and Pharmacy, reference number 117/04.06.2024. All images were de-identified before analysis. The study was conducted in accordance with the principles of the Declaration of Helsinki. Panoramic images of diagnostic quality were selected based on the following criteria: adult patients with permanent dentition. Images containing large cysts, tumors, metallic artifacts, blurring, or severe technique errors that would impede the radiological interpretation were excluded. Image Annotation The panoramic images were manually annotated by three calibrated researchers using an open-source labeling platform (Makesense.ai v1.11.0). The inter-observer agreement was calculated using the interclass correlation coefficient (ICC), to ensure examiner reliability and achieved a value of 0.91. Bounding boxes were created for the collected images to identify the following dental conditions: prosthetic restoration, dental implant, dental filling, endodontic treatment, caries, periapical lesion, periodontal bone loss, impacted tooth, root fragment, root resorption, orthodontic treatment (brackets), and surgical devices (i.e., osteosynthesis plates, orthodontic temporary anchorage devices). Finally, the annotations were exported in the YOLO label format. nc: 14 names: ['IMP', 'PRR', 'OBT', 'END', 'CAR', 'BON', 'IMT', 'API', 'ROT', 'FUR', 'APS', 'ROR', 'ORD', 'SRD'] segment_names_to_labels = [("Implant", 0), ("Prosthetic restoration", 1), ("Obturation", 2), ("Endodontic treatment", 3), ("Orthodontic device", 12), ("Surgical device", 13), ## Low risk ("Carious lesion", 4), ("Bone resorbtion", 5), ("Impacted tooth", 6), ("Apical surgery", 10), ## Medium risk ("Apical periodontitis", 7), ("Root fragment", 8), ("Furcation lesion", 9), ("Root resorption", 11)## High risk ] LABEL_NAMES = [ 'IMP', # Implant -> LOW risk 'PRR', # Prosthetic Restoration -> LOW risk 'OBT', # Obturation -> LOW risk 'END', # Endodontic Treatment -> LOW risk 'CAR', # Carious Lesion -> MEDIUM risk 'BON', # Bone Resorbtion -> MEDIUM risk 'IMT', # Impacted Tooth -> MEDIUM risk 'API', # Apical Periodontitis -> HIGH risk 'ROT', # Root Fragment -> HIGH risk 'FUR', # Furcation Lesion -> HIGH risk 'APS', # Apical Surgery -> MEDIUM risk 'ROR', # Root Resorbtion -> HIGH risk 'ORD', # Orthodontic device -> LOW risk 'SRD' # Surgical Device -> LOW risk ]