Can Deep Learning Methods Differentiate Temporomandibular Joint Disorders From Healthy Joints? A 3D Artificial Intelligence Algorithm Study Based on CBCT Images.
Authors
Affiliations (6)
Affiliations (6)
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Eskisehir Osmangazi University, Eskisehir, Turkey.
- Division of Oral and Maxillofacial Radiology, Department of Care Planning and Restorative Sciences, University of Mississippi Medical Center School of Dentistry, Jackson, Mississippi, USA.
- Department of Periodontology, Faculty of Dentistry, Eskisehir Osmangazi University, Eskisehir, Turkey.
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Kocaeli University, Kocaeli, Turkey.
- Artificial Intelligence Research and Application Center, Anadolu University, Eskisehir, Turkey.
- Department of Oral and Maxillofacial Radiology, Faculty of Dentistry, Ankara University, Ankara, Turkey.
Abstract
Given the high difficulty of interpreting complex bone-related temporomandibular joint disorders (TMD) on CBCT scans, developing advanced artificial intelligence models has become essential for accurate diagnosis and detailed subcategorization. This study aimed to develop deep learning models capable of performing a wide range of tasks in the evaluation of the temporomandibular joint on CBCT scans. This study followed a four-stage design. First, an nnU-Net v2-based model was developed for automatic segmentation of mandibular condyles on CBCT images. Second, a 3D CNN model was trained to classify healthy and pathological condyles. Third, a 3D CNN was used to differentiate five TMD types. In the fourth stage, for exploratory analysis purposes, an nnU-Net v2-based model was implemented to grade the severity of erosion, osteophyte formation and sclerosis. The segmentation model achieved automatic segmentation of healthy and pathological condyles with 0.87 DSC and 0.77 IoU values. The classification model achieved an F1-score of 0.65 in distinguishing healthy condyles from TMDs. For differentiating the five TMD types from healthy condyles, the F1-scores were 0.76 for erosion, 0.88 for flattening, 0.86 for osteophyte formation, 0.88 for sclerosis and 0.85 for subchondral cysts. In grading analysis, the highest performance was observed for Grade 1 erosion, Grade 2 osteophyte formation and Grade 3 sclerosis. Deep learning algorithms can achieve clinically relevant performance in the segmentation of the mandibular condyle and in differentiating TMDs from healthy condyles. However, further studies with larger and more balanced datasets are needed to improve the ability of these models to classify TMD grades accurately.