M⁴4TMD: A multimodal, multi-task deep learning framework for comprehensive assessment of TMD-related abnormalities

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Current deep learning (DL) for temporomandibular disorders (TMD) underutilizes MRI, focuses on single tasks, and relies on limited data types.
• Addressing these limitations is crucial for improving diagnostic accuracy and clinical utility of AI in TMD assessment.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning (DL) framework, M44TMD, for the concurrent assessment of TMD-related abnormalities.
• To integrate multi-sequence, multi-slice MRI with clinical data for enhanced diagnostic performance.

Main Methods:

• Collected 12,690 MRI slices and clinical data from 765 participants.
• Developed M44TMD framework utilizing multimodal data for concurrent assessment of degenerative joint disease (DJD), anterior disc displacement (ADD), and effusion.
• Benchmarked M44TMD against existing DL methods and clinicians, including generalization and interpretability experiments.

Main Results:

• M44TMD demonstrated superior internal test performance (ROC-AUC: 0.831, 0.913, 0.961) compared to prior methods.
• The framework's accuracy for DJD, ADD, and effusion surpassed junior dentists and was comparable to senior dentists.
• Validated robustness and interpretability through generalization and visual interpretability experiments.

Conclusions:

• The M44TMD framework effectively integrates multimodal MRI and clinical data for concurrent TMD abnormality assessment.
• M44TMD exhibits diagnostic performance comparable to senior dentists, highlighting its potential for clinical application.
• This framework marks a significant advancement toward integrating DL-based TMD diagnosis into clinical practice.