AI segments zygomatic bones on CBCT scans

Title

A deep learning-based automatic segmentation of zygomatic bones from cone-beam computed tomography images: A proof of concept

Authors

Baoxin Tao; Xinbo Yu; Wenying Wang; Haowei Wang; Xiaojun Chen; Feng Wang; Yiqun Wu

Summary

This proof-of-concept study proposes a deep learning–based framework for automatic segmentation of zygomatic bones from cone-beam computed tomography (CBCT) images. The model combines a slice-level classification network (VGG-16) with a 3D U-Net–based segmentation network enhanced by edge supervision using a Canny operator. Attention maps were generated using Grad-CAM and Guided Grad-CAM to improve interpretability. A total of 130 CBCT scans were randomly divided into training, validation, and test datasets (6:2:2). The model achieved high accuracy and efficiency, requiring ~17 seconds per scan versus ~49 minutes for dentists, and produced higher Dice scores than dentists in a 10-scan comparative evaluation.

DOI

https://doi.org/10.1016/j.jdent.2023.104582

Key Words

Medical imaging; Artificial intelligence; Deep learning; Neural networks; Zygoma; Digital dentistry

Extracted Data

Year: 2023

Modality: Cone-Beam Computed Tomography (CBCT)

Dataset: 130 CBCT scans (adult patients; bilateral zygomatic bones fully included; excluded congenital dysplasia, prior zygomatic reconstruction, and/or maxillofacial trauma).

Dataset Split: Training: 78 (60%); Validation: 26 (20%); Test: 26 (20%). Additionally, 10 scans randomly selected from the test set for comparison with 4 dentists.

Network Architecture: VGG-16 (Classification), 3D U-Net (Segmentation) with edge supervision (Canny operator) and attention modules (Grad-CAM/Guided Grad-CAM)

Metrics: Classification: Accuracy, Precision, Recall, F1-score. Segmentation: Dice coefficient (Dice), Intersection over Union (IoU), Precision, Recall, Average Surface Distance (ASD), 95% Hausdorff Distance (HD).

AP - Professional Qty: 3

AP - Supervisor Presence: No information

AP - Experience Level: No information

AP - Expertise Area: General dentists

AP - Tool or System: ITK-SNAP 3.8

ML Task: Image Classification and Semantic Segmentation

##Project Objective: Develop and validate an accurate and efficient deep learning model to automatically segment zygomatic bones from CBCT, improving speed and consistency versus manual workflows.

Clinical Relevance: Automatic zygomatic bone segmentation can generate accurate 3D models for preoperative digital planning of zygoma reconstruction, orbital surgery, zygomatic implant surgery, and orthodontics; improves efficiency and reproducibility.