Diagnostic Accuracy of Artificial Intelligence in Radiographic Assessment of Bony Changes in the Temporomandibular Joint

Abstract

This study aimed to evaluate the diagnostic accuracy of the large language model (Gemini) in detecting osseous changes of the Temporomandibular Joint in Cone Beam Computed Tomography (CBCT) images and in identifying the specific type of these changes, in comparison with an oral and maxillofacial radiologist whose diagnosis was considered the gold standard. This study included 240 sections of the mandibular condyle obtained from CBCT images. These sections were classified by the specialist according to the presence or absence of osseous changes and subsequently categorized by type of change (flattening, erosion, osteophyte formation, sclerosis, pseudocyst formation, or mixed changes). The image sections were then entered into the model in .PNG format, and the model was provided with a standardized prompt to evaluate all images. The results demonstrated that the model achieved moderate diagnostic performance in the binary classification task of distinguishing between the presence and absence of osseous changes. The overall accuracy was 75.0%, sensitivity was 70.2%, and specificity was 77.6%, with moderate agreement with the gold standard (κ=0.465, p<0.001). However, its performance declined markedly when performing the more complex task of classifying the type of osseous change. In this task, the overall accuracy reached 42.9%, with relatively low agreement with the gold standard (κ=0.242, p<0.001). The findings of this study suggest that multimodal models may hold promising potential in supporting dental radiology practice. Nevertheless, their current use should remain within the framework of assistive tools rather than serving as a substitute for specialized clinical expertise.