ARTIFICIAL INTELLIGENCE FOR TEMPOROMANDIBULAR JOINT DETECTION AND ANALYSIS ON CONE BEAM COMPUTED TOMOGRAPHY

Julien Issa

Institute

Department of Diagnostics, University of Medical Sciences, Bukowska 70, 60-812 Poznan, Poland

Abstract

Introduction: The Temporomandibular joint (TMJ) is vital for jaw functions like chewing and swallowing. Unfortunately, temporomandibular joint disorders (TMD) are rising among young adults, causing jaw pain and movement difficulties. Diagnosis involves clinical evaluation and diagnostic imaging, such as cone beam computed tomography (CBCT), for visualizing bony structures, which can be time-consuming and requires specialized training. However, artificial intelligence (AI) advancements have shown potential in automating analysis and saving time, enhancing diagnosis and treatment in dentistry.

Aim of the study: This study aims to develop and validate an AI-based tool for TMJ detection and analysis on CBCT scans, focusing on joint articular space analysis.

Material and methods: We acquired 17 anonymized CBCT scans from the Poznan University of Medical Sciences database, comprising patients with and without TMD. The collected data were preprocessed, and our models were trained and validated using the leave-one-out cross-validation method at the patient level. Initially, we developed an AI-based model utilizing a custom Convolutional Neural Network (CNN) to extract ten sagittal slices suitable for further analysis from the DICOM files. These slices encompassed all components of the TMJ. Subsequently, a second CNN model based on the Resi2 architecture analyzed the ten slices. The model was trained with data annotated by two specialists identifying relevant TMJ landmarks. The results were assessed by specialists and underwent statistical analysis.

Results: The model extracted the suitable slices with an accuracy of 95.13%. For the joint space analysis, the model achieved an overall average of Mean Absolute Error (MAE) and Mean Squared Error (MSE) were 0.684 mm and 1.107 mm2, respectively.

Conclusion: The first model accurately identifies the sagittal slices that are most suitable for further analysis. However, the MAE for the second model was high, indicating that additional training is required with a larger sample size to improve the accuracy of that model.