Deep Learning for MRI-Visible TMJ Abnormalities Detection from Panoramic Radiographs

Overview

This study developed an interpretable multimodal deep learning framework integrating anatomically guided attention and clinical metadata to detect MRI-visible temporomandibular joint (TMJ) abnormalities from panoramic radiographs. The ensemble model achieved an area under the curve (AUC) of 0.86 in classifying MRI-positive and MRI-negative TMJ cases, demonstrating promising diagnostic accuracy and interpretability.

Background

Temporomandibular disorders (TMD) are challenging to diagnose early, particularly intra-articular TMJ abnormalities which require magnetic resonance imaging (MRI) for confirmation. Panoramic radiographs are widely used but have limited sensitivity for detecting these abnormalities. Artificial intelligence and deep learning approaches offer potential to improve screening by integrating imaging with clinical data. This study aimed to develop a comprehensive, interpretable AI framework to triage patients for MRI referral based on panoramic radiographs and clinical metadata.

Data Highlights

Key Findings

• The ensemble deep learning framework combined open- and closed-mouth TMJ panoramic radiographs with structured clinical metadata for diagnosis.
• Incorporation of anatomically guided attention mechanisms focused the model on condylar regions relevant to TMJ abnormalities.
• Gradient-weighted Class Activation Mapping (Grad-CAM) confirmed consistent model attention on clinically relevant anatomical areas.
• Ablation studies showed that adding clinical metadata and spatial attention improved diagnostic performance.
• The model achieved balanced classification performance between MRI-negative and MRI-positive TMJ cases with an AUC of 0.86.

Clinical Implications

This interpretable AI framework can assist clinicians in early identification of patients with MRI-visible TMJ abnormalities using widely available panoramic radiographs. By triaging patients for MRI referral more effectively, it supports timely diagnosis and intervention for temporomandibular disorders. Integration of clinical metadata alongside imaging enhances diagnostic accuracy, emphasizing the value of multimodal data in TMJ assessment.

Conclusion

The study presents a promising deep learning-based screening tool that leverages multimodal data and anatomically guided attention to detect MRI-visible TMJ abnormalities from panoramic radiographs. This approach may facilitate early diagnosis and improve clinical decision-making for temporomandibular disorders.

References

1. Choi et al. 2025 -- Artificial intelligence-enhanced diagnosis of degenerative joint disease using temporomandibular joint panoramic radiography and joint noise data
2. Singer & Mupparapu 2023 -- Temporomandibular joint imaging
3. Schiffman et al. 2014 -- Diagnostic Criteria for Temporomandibular Disorders (DC/TMD) for Clinical and Research Applications
4. Zieliński et al. 2024 -- A meta-analysis of the global prevalence of temporomandibular disorders