IADR Abstract Archives
AI-Driven Detection of Alveolar Bone Lesions in CBCT Images
Objectives: 1. To train an online web-based AI model to detect dehiscence and fenestration lesions using a collected dataset of 2D sagittal cross-section radiographs extracted from Cone Beam Computed Tomography (CBCT) scans.
2. To assess the effectiveness of the diagnosing capability of the trained AI model when additional CBCT radiographs were provided.
Methods: A total of 640 CBCT radiographs of mandibular canines were analyzed and classified into "no lesion," "dehiscence," or "fenestration." The AI model was trained using Google Teachable Machine with these images and tested on an additional 420 radiographs. Sensitivity, specificity, accuracy, precision, and F1 score were evaluated by comparing the AI predictions to the true diagnoses.
Results: The code was successfully generated and trained with the 640 CBCT images. 420 images were used to test the machine’s efficiency. The results showed that the system achieved an overall accuracy of 79.47%, a specificity of 79.23% a sensitivity of 79.58%, a precision of 89.49%, and an F1 of 84.25%. When analyzing fenestration lesions, the system achieved an accuracy of 87.27%, a specificity of 85.19%, a sensitivity of 89.29%, a precision of 86.21%, and an F1 of 87.72%. For dehiscence lesions, the system achieved an accuracy of 75.47%, a specificity of 71.88%, a sensitivity of 80.95%, a precision of 65.38%, and an F1 of 72.34.
Conclusions: AI shows promise in detecting alveolar bone defects from CBCT radiographs, particularly fenestration lesions. However, further validation with larger datasets and different tooth types is recommended.
2. To assess the effectiveness of the diagnosing capability of the trained AI model when additional CBCT radiographs were provided.
Methods: A total of 640 CBCT radiographs of mandibular canines were analyzed and classified into "no lesion," "dehiscence," or "fenestration." The AI model was trained using Google Teachable Machine with these images and tested on an additional 420 radiographs. Sensitivity, specificity, accuracy, precision, and F1 score were evaluated by comparing the AI predictions to the true diagnoses.
Results: The code was successfully generated and trained with the 640 CBCT images. 420 images were used to test the machine’s efficiency. The results showed that the system achieved an overall accuracy of 79.47%, a specificity of 79.23% a sensitivity of 79.58%, a precision of 89.49%, and an F1 of 84.25%. When analyzing fenestration lesions, the system achieved an accuracy of 87.27%, a specificity of 85.19%, a sensitivity of 89.29%, a precision of 86.21%, and an F1 of 87.72%. For dehiscence lesions, the system achieved an accuracy of 75.47%, a specificity of 71.88%, a sensitivity of 80.95%, a precision of 65.38%, and an F1 of 72.34.
Conclusions: AI shows promise in detecting alveolar bone defects from CBCT radiographs, particularly fenestration lesions. However, further validation with larger datasets and different tooth types is recommended.