AI-Based Detection of Temporomandibular Disorders on CBCT

This dataset provides the test-set images, expert ground truth annotations, and model predictions generated during the study "Can Deep Learning Methods Differentiate Temporomandibular Joint Disorders from Healthy Joints? A 3D Artificial Intelligence Algorithm Study Based on CBCT Images." The dataset is shared to support the reproducibility and external benchmarking of the segmentation and classification results reported in the original publication. The dataset is organized into three components: Segmentation test set (mandibular condyle): Cone-beam computed tomography (CBCT) images of the test cohort used to evaluate the nnU-Net v2–based mandibular condyle segmentation model, along with the corresponding expert-annotated ground truth masks and the model-predicted segmentation masks. This component allows independent recomputation of the segmentation metrics reported in the main study (DSC, IoU, precision, recall, F1-score). Classification test set (cropped condyles):Cropped, region-of-interest (ROI) NIfTI volumes of the mandibular condyles used to evaluate the 3D-CNN classification models, organized as paired healthy / disorder test samples for each TMD subtype investigated in the study (erosion, flattening, osteophyte formation, sclerosis, and subchondral cyst). Grading test set: CBCT images used to evaluate the multiclass nnU-Net v2 grading models, including the test cases for erosion, osteophyte formation, and sclerosis, together with their grade labels (Grade 1, Grade 2, Grade 3) as defined in the main study. All images are provided in NIfTI format (.nii / .nii.gz) and were anonymized prior to release. The data were originally collected from two centers using multiple CBCT devices and curated by expert oral and maxillofacial radiologists.