# Deep learning-based diagnosis of temporomandibular joint osteoarthritis using whole-body bone scans

## TL;DR

This paper shows that a deep learning model can accurately diagnose temporomandibular joint osteoarthritis using targeted bone scans, better than whole-body scans.

## Contribution

A lightweight deep learning model (VGG16-Lite) was developed and shown to outperform existing models for TMJ-OA diagnosis using bone scans.

## Key findings

- VGG16-Lite achieved high diagnostic accuracy (AUC >0.90) for TMJ-OA using head-and-neck bone scans.
- Whole-body scans provided limited predictive value for TMJ-OA (AUC ∼0.65).
- The lightweight model offers high accuracy with low computational cost.

## Abstract

Temporomandibular joint osteoarthritis (TMJ-OA) is a degenerative condition that causes pain and functional limitation, yet its relationship with systemic osteoarthritis (OA) remains unclear. This study developed deep learning models to automatically diagnose TMJ-OA using bone scintigraphy (bone scans) and to evaluate systemic OA features as potential predictors. A dataset of 1,943 patients (3,886 TMJs) was analyzed with three convolutional neural network (CNN) approaches based on the VGG16 architecture. In head-and-neck imaging, the VGG16-Lite model achieved outstanding diagnostic accuracy (AUC >0.90) across age and sex subgroups, outperforming pretrained models. Whole-body scans excluding the head and neck provided only modest predictive value for TMJ-OA (AUC ∼0.65), suggesting limited utility of systemic features alone. These findings highlight the value of targeted bone scans with lightweight deep learning models for robust and efficient TMJ-OA detection, while also underscoring the need for further research into systemic associations.

•Deep learning (VGG16-Lite) enables accurate TMJ-OA diagnosis on bone scans (AUC >0.90)•TMJ-focused imaging outperforms whole-body scans (AUC >0.90 vs. ∼0.65)•Lightweight model achieves high accuracy with low computational cost•AI-assisted diagnostics may improve accessibility for TMJ-OA in clinical practice

Deep learning (VGG16-Lite) enables accurate TMJ-OA diagnosis on bone scans (AUC >0.90)

TMJ-focused imaging outperforms whole-body scans (AUC >0.90 vs. ∼0.65)

Lightweight model achieves high accuracy with low computational cost

AI-assisted diagnostics may improve accessibility for TMJ-OA in clinical practice

Orthopedics; Bioinformatics