Abstract
BACKGROUND: The diagnosis of temporomandibular joint (TMJ) disc displacement relies on clinical symptoms and magnetic resonance imaging (MRI), which is complex, costly and time-consuming. Although cone-beam computed tomography (CBCT) reveals indirect signs suggestive of TMJ disc displacement, manual interpretation remains expertise-dependent, thereby limiting its use in clinical practice. This study aims to predict the presence of TMJ disc displacement risk in CBCT images using deep learning techniques. METHODS: By leveraging the CBCT images of 330 patients, a two-stage TMJ disc displacement screening model was developed. In the first stage, an object-detection model was trained on YOLOv11, using 30 manually annotated CBCT images as reference. A total of 5,238 TMJ Regions of Interest (ROIs) were identified, among which 2,260 showing signs of TMJ disc displacement. Subsequently, these detected images were used to train a FastViT-t8-based binary-classification model, with diagnostic results of two experienced oral and maxillofacial radiologists based on MRI set as the ground truth. RESULTS: The object-detection model achieved a Precision of 0.986, a Recall of 0.982, an mAP50 of 0.988, and an mAP50-95 of 0.534. The binary-classification model achieved an AUC of 0.733 (95% CI: [0.713-0.756]), an AUPR of 0.716 (95% CI: [0.685-0.745]), and an accuracy of 0.669. CONCLUSIONS: The proposed model demonstrates preliminary screening capability for TMJ disc displacement using CBCT images. While its current performance precludes standalone diagnostic use, the model may serve as a practical triage tool in orthodontic settings, assisting in the early identification of patients who should be referred for confirmatory MRI and offering references for related research.