Deep learning for automated detection and classification of temporomandibular joint disorders: A scoping review.

PURPOSE: To map the available evidence on the application of deep learning (DL) in the imaging-based detection and classification of temporomandibular joint disorders (TMDs), and to identify research trends, imaging modalities investigated, and gaps in the current literature.

MATERIALS AND

METHODS

A comprehensive search was conducted in PubMed, IEEE Xplore, Wiley Online Library, ScienceDirect, and Cochrane, covering publications between January 2015 and January 2025, without restrictions on language. Studies applying any DL algorithm to medical imaging for the detection, segmentation, or classification of TMDs were included.

RESULTS

Fifteen studies met the inclusion criteria, comprising 15 retrospective studies. MRI was the most frequently used imaging modality (10 studies), followed by CBCT (3 studies) and panoramic radiography (2 studies). The most commonly applied DL architectures were CNN-based classifiers (9 studies), followed by segmentation frameworks including U-Net and DeepLabV3+ (4 studies), and object-detection models such as YOLO(two studies).

CONCLUSION

This scoping review demonstrates that the available evidence on deep learning applications in TMJ imaging is growing, with CNN-based classifiers and segmentation frameworks applied to MRI and CBCT showing early promise as adjunctive diagnostic tools. However, most studies are retrospective, single-center, and lack external validation. These findings map a rapidly evolving field and highlight the need for multicenter prospective studies, standardized imaging and reporting protocols, and greater integration of explainable AI frameworks to guide future research and support clinical translation.