Automated pediatric TMJ articular disk identification and displacement classification in MRI with machine learning.

OBJECTIVE

To evaluate the performance of an automated two-step model interpreting pediatric temporomandibular joint (TMJ) magnetic resonance imaging (MRI) using artificial intelligence (AI). Using deep learning techniques, the model first automatically identifies the disk and the TMJ osseous structures, and then an automated algorithm classifies disk displacement.

MATERIALS AND

METHODS

MRI images of the TMJ from 235 pediatric patients (470 joints) were reviewed. TMJ structures were segmented, and the disk position was classified as dislocated or not dislocated. The UNet++ model was trained on MRI images from 135 and tested on images from 100 patients. Disk displacement was then classified by an automated algorithm assessing the location of disk centroid and surfaces for bone landmarks.

RESULTS

The mean age was 14.6 ± 0.1 years (Female: 138/235, 58 %), with 104 of 470 disks (22 %) anteriorly dislocated. UNet++ performed well in segmenting the TMJ anatomical structures, with a Dice coefficient of 0.67 for the disk, 0.91 for the condyle, and a Hausdorff distance of 2.8 mm for the articular eminence. The classification algorithm showed disk displacement classification comparable to human experts, with an AUC of 0.89-0.92 for the distance between the disk center and the eminence-condyle line.

CONCLUSION

A two-step automated model can accurately identify TMJ osseous structures and classify disk dislocation in pediatric TMJ MRI. This tool could assist clinicians who are not MRI experts when assessing pediatric TMJ disorders.

CLINICAL SIGNIFICANCE: Automated software that assists in locating the articular disk and surrounding structures and classifies disk displacement would improve the TMJ-MRI interpretation and the assessment of TMJ disorders in children.