Temporomandibular joint segmentation in MRI images using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Temporomandibular joint (TMJ) internal derangements (ID) are common TMJ disorders (TMD).
• Diagnosis relies on MRI, but low resolution and contrast hinder identification of TMJ articular discs.
• Accurate localization of the articular disc is critical for TMJ-ID diagnosis.

Purpose of the Study:

• To apply convolutional neural networks (CNNs) for segmenting mandibular condyle, articular eminence, and TMJ disc in MRI images.
• To evaluate the performance of CNN models against human raters with varying expertise.

Main Methods:

• Two CNN models were trained on MRI images from 100 patients and validated on 40 patients.
• 2D slices and 3D volume data were used as input for model validation.
• Data augmentation and five-fold cross-validation were employed; model accuracy was compared to four human raters.

Main Results:

• Both CNN models achieved high segmentation accuracy: Dice coefficient ~0.7 for articular disc, >0.9 for mandibular condyle, and Hausdorff distance ~2mm for articular eminence.
• Models demonstrated near-expert performance for articular disc segmentation and expert-level performance for mandibular condyle and articular eminence.
• CNN models outperformed non-expert raters in segmenting all three TMJ anatomical structures.

Conclusions:

• CNN-based segmentation models are reliable tools for assisting clinicians in identifying key TMJ anatomy on MRI.
• These models offer a foundation for the automatic diagnosis of TMD.
• Automated segmentation can enhance TMJ-MRI interpretation reliability, save time, and aid reader training.