A Comparison of Deep Learning vs. Dental Implantologists in Cone-Beam Computed Tomography-Based Bone Quality Classification.

Abstract

Bone quality assessment is crucial for pre-surgical implant planning, influencing both implant design and drilling protocol selection. The Lekholm and Zarb (L&Z) classification, which categorizes bone quality into four types based on cortical bone width and trabecular bone density using cone-beam computed tomography (CBCT) data, lacks quantitative guidelines, leading to subjective interpretations. This study aimed to compare the performance of deep learning (DL)-based approaches against human examiners in assessing bone quality, according to the L&Z classification, using CBCT images. A dataset of 1100 CBCT cross-sectional slices was classified into four bone types by two oral and maxillofacial radiologists. Five pre-trained DL models were trained on 1000 images using MATLAB^®, with 100 images reserved for testing. Inception-ResNet-v2 achieved the highest accuracy (86.00%) with a learning rate of 0.001. The performance of Inception-ResNet-v2 was then compared to that of 23 residency students and two experienced implantologists. The DL model outperformed human assessors across all parameters, demonstrating excellent precision and recall, with F1-scores exceeding 75%. Notably, residency students and one implantologist struggled to distinguish bone type 2, with low recall rates (48.15% and 40.74%, respectively). In conclusion, the Inception-ResNet-v2 DL model demonstrated superior performance compared to novice implantologists, suggesting its potential as an supplementary tool for cross-sectional bone quality assessment.