An accuracy of deep learning-based tool for single implant placement : phase I space analysis
TL;DR
An enhanced deep learning tool for automated detection and segmentation of single edentulous areas on CBCT images demonstrated high accuracy, time efficiency, and consistency, making it a valuable asset for pre-surgical implant planning.
Abstract
Objective: To analyze the accuracy and time efficiency of the enhanced deep learning model compared to experts in detecting and segmenting single edentulous areas. Material and methods: Sixty-six CBCT scans from four machines, comprising 80 single edentulous areas, were examined and split into training and validation sets. Using 3D Slicer, Dentalsegmentator and expert-defined ground truth, a decision tree-based model was developed. Intra- and interobserver consistency was assessed. The model's performance was evaluated by comparing time efficiency, volume difference, surface deviation, and analytical metrics (accuracy, precision, recall, specificity, and F1-score) against manual detection. Results: The enhanced deep learning model achieved high performance in automated detection (accuracy, precision, recall, specificity, and F1-score values of 0.95, 0.95, 0.96, 0.93, and 0.96, respectively), with a median volume difference of 183.31 mm³ and mean surface deviation of 0.575 ± 0.65 mm. Detection and segmentation were also significantly faster than expert analysis, with 37- and 19-fold time reductions (p < 0.001). Conclusions: An enhanced deep learning tool for automated detection and segmentation of single edentulous areas on CBCT images demonstrated high accuracy, time efficiency, and consistency, making it a valuable asset for pre-surgical implant planning.