Deep Learning for Ultrasonographic Assessment of Temporomandibular Joint Morphology

Abstract

Background: Temporomandibular joint (TMJ) disorders are a significant cause of orofacial pain. Artificial intelligence (AI) has been successfully applied to other imaging modalities but remains underexplored in ultrasonographic evaluations of TMJ.

Objective: This study aimed to develop and validate an AI-driven method for the automatic and reproducible measurement of TMJ space width from ultrasonographic images.

Methods: A total of 142 TMJ ultrasonographic images were segmented into three anatomical components: the mandibular condyle, joint space, and glenoid fossa. State-of-the-art architectures were tested, and the best-performing 2D Residual U-Net was trained and validated against expert annotations. The algorithm for joint space width measurement based on TMJ segmentation was proposed, calculating the vertical distance between the superior-most point of the mandibular condyle and its corresponding point on the glenoid fossa.

Results: The segmentation model achieved high performance for the mandibular condyle (Dice: 0.91 ± 0.08) and joint space (Dice: 0.86 ± 0.09), with notably lower performance for the glenoid fossa (Dice: 0.60 ± 0.24), highlighting variability due to its complex geometry. The TMJ space width measurement algorithm demonstrated minimal bias, with a mean difference of 0.08 mm and a mean absolute error of 0.18 mm compared to reference measurements.

Conclusions: The model exhibited potential as a reliable tool for clinical use, demonstrating accuracy in TMJ ultrasonographic analysis. This study underscores the ability of AI-driven segmentation and measurement algorithms to bridge existing gaps in ultrasonographic imaging and lays the foundation for broader clinical applications.

Keywords: artificial intelligence; deep learning; temporomandibular joints; ultrasonography.

Figure 1

Figure 1 presents methodology of examination of the TM Joint. (a). Hockey stick probe alignment parallel to the joint gap. (b). Obtained US image of the joint with indicated articulation surfaces (arrows). (c). Joint gap width–searched parameter (marked with dotted line).

Figure 2

Schematic representation of the Residual U-Net architecture.

Figure 3

The plot visualizes the performance of different hyperparameter combinations for a subset of experiments.

Figure 4

Segmentation performance metrics for the mandibular condyle (MC), joint space (JS), and glenoid fossa (GF) on the test set. Metrics include Dice coefficient, precision, recall, and volume similarity present as boxplots. Circles represent outliers.

Figure 5

Bland–Altman and concordance plots for assessing the agreement between AI-predicted TMJ space width measurements and reference measurements.

Figure 6

The visualization showcases TMJ segmentation and joint space width (yellow dashed line) measurement results, highlighting the mandibular condyle (MC), joint space (JS), and glenoid fossa (GF) with distinct colors: MC in blue, JS in pink, and GF in green. Segmentation performance is demonstrated across three scenarios: (A) the highest Dice (GF: 0.89, JS: 0.96, MC: 0.97), (B) the average Dice (GF: 0.78, JS: 0.91, MC: 0.84), (C) and the lowest Dice (GF: 0.00, JS: 0.65, MC: 0.84).