3D CNN Distinguishes TMJ Disorders from Healthy Joints on CBCT with Moderate Accuracy

A 3D deep learning model differentiated temporomandibular joint disorders (TMD) from healthy condyles on CBCT with a moderate F1-score of 0.65. Performance was higher for specific TMD subtypes, reaching 0.76–0.88. This retrospective study indicates potential but requires larger,…

• Design: A four-stage retrospective deep learning study developed on an unspecified number of temporomandibular joint CBCT scans.
• Key finding: For differentiating five TMD types from healthy condyles, 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.
• Limitation: The study notes the need for larger, more balanced datasets, indicating a likely single-center, retrospective design with potential selection bias that limits generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.