DEEP LEARNING-BASED ASSESSMENT OF JAWBONE DENSITY FOR DENTAL IMPLANT PLANNING: A DIAGNOSTIC ACCURACY STUDY | ||
Alexandria Dental Journal | ||
Articles in Press, Accepted Manuscript, Available Online from 06 October 2025 PDF (654.11 K) | ||
Document Type: Original Article | ||
DOI: 10.21608/adjalexu.2025.372617.1614 | ||
Authors | ||
Sara Madian* 1; Hassan Abo ElKheir2; Marwan Torki3; Noha Elkersh4 | ||
1Oral Medicine, Periodontology, Oral Diagnosis and Oral Radiology department, Faculty of Dentistry, Alexandria University, Alexandria, Egypt | ||
2Professor of Oral Medicine, Periodontology, Oral Diagnosis and Oral Radiology Department, Faculty of Dentistry, Alexandria University | ||
3Department of Computer and Systems Engineering, Faculty of Engineering, Alexandria University, Alexandria, Egypt. | ||
4lecturer of Oral Medicine, Periodontology, Oral Diagnosis and Radiology Department, Faculty of Dentistry, Alexandria University, Alexandria, Egypt | ||
Abstract | ||
Introduction: Dental implants have transformed restorative dentistry, offering a reliable solution for tooth replacement. Their success depends on primary implant stability, which is closely tied to bone density. Misch’s classification system provides a precise method for assessing bone density. Cone Beam Computed Tomography (CBCT) has emerged as a preferred tool due to its lower radiation and cost-effectiveness. Recent advancements in deep learning, particularly Vision Transformers, show promise in analyzing CBCT images for bone density classification. Aim: This research focused on designing and evaluating Vision Transformer (ViT) models to classify jawbone density from CBCT scans. Materials and Methods: A comprehensive dataset of 5,545 CBCT images, extracted from 500 scans, was organized into training, validation, and testing groups. Binary masks were utilized to isolate regions of interest, and the images were categorized into five density types following the Misch classification. Several ViT architectures were trained and assessed, with performance evaluated using key metrics, including accuracy, sensitivity, specificity, loss, and area under the curve (AUC). Results: The SwinV2 model delivered the best overall performance, achieving the highest accuracy (85.65%) and specificity (90.13%), along with a strong AUC (0.73) and the lowest loss (0.8905). The ViTamin model excelled in sensitivity, while the XciT model also performed well, showcasing its reliability. The integration of binary masks improves model outcomes, emphasizing their value in refining classification tasks. Conclusions: The SwinV2 model proved to be the most effective for jawbone density classification. The use of binary masks significantly enhanced model accuracy. | ||
Keywords | ||
Artificial Intelligence; Cone Beam Computed Tomography; Convolutional Neural Networks; Deep Learning; Density Classification | ||
Statistics Article View: 15 PDF Download: 4 |