The application of deep learning in intelligent assessment of horizontal bone loss in mandibular first molars.

OBJECTIVE: This study aims to develop an assessment tool that utilizes deep learning (DL) techniques to automatically detect the extent of horizontal bone loss in mandibular first molars on cone-beam computed tomography (CBCT) images. METHODS: A total of 505 CBCT images of patients with mandibular first molar horizontal bone loss were included in this study. Clinical experts annotated the horizontal bone loss areas, and a 3D-ResNets model was employed to classify the degree of horizontal bone loss. The diagnostic and classification performance of the model was then evaluated. RESULTS: The 3D-ResNets model demonstrated accurate diagnosis and classification of healthy teeth, degree I/II horizontal bone loss teeth, and degree III horizontal bone loss teeth (Precision: 0.814 vs. 0.778 vs. 0.906; Recall: 0.921 vs. 0.737 vs. 0.829; F1-score: 0.864 vs. 0.757 vs. 0.866; Accuracy: 0.901 vs. 0.838 vs. 0.919). The model demonstrated reliable diagnostic performance, achieving an overall accuracy (OA) of 0.829, comparable to the dentist's OA of 0.820 (no significant difference). CONCLUSION: The 3D-ResNets model demonstrates high accuracy, sensitivity, and specificity in detecting the extent of horizontal bone loss in mandibular first molars on CBCT images, showing promising potential for clinical application. CLINICAL SIGNIFICANCE: The current study provides preliminary experimental support for the development of a clinical auxiliary diagnostic system to automatically detect the degree of horizontal bone loss in mandibular first molars.