Alveolar Bone Segmentation in Intraoral Ultrasonographs with Machine Learning.

Area of Science:

• Dentistry
• Medical Imaging
• Machine Learning

Background:

• Intraoral ultrasound imaging offers a portable, non-ionizing, and cost-effective solution for dental care.
• Accurate assessment of alveolar bone is crucial for periodontal diagnosis, but its interpretation in ultrasound images is challenging.
• Alveolar bone supports teeth and is a key structure in the periodontal apparatus.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) approach for automatic segmentation of alveolar bone and localization of the alveolar crest in intraoral ultrasound images.
• To enhance the robustness of ML algorithms through data augmentation techniques.
• To provide a tool assisting dentists in visualizing alveolar bone structures.

Main Methods:

• Training and validation of three convolutional neural network (CNN)-based ML models using intraoral ultrasound images.
• Implementation of data augmentation, including vertical/horizontal shifting and flipping, to synthesize 2100 additional training images.
• Quantitative evaluation of the best performing ML model against expert clinician assessments on 200 images.

Main Results:

• The best ML model achieved an 85.3% Dice score, 88.5% sensitivity, and 99.8% specificity in segmenting alveolar bone.
• The model accurately identified the alveolar crest with a mean difference of 0.20 mm and high reliability (ICC ≥0.98).
• The ML approach provided results in less than one second per image.

Conclusions:

• Machine learning shows significant potential for assisting dentists in the accurate visualization and analysis of alveolar bone from intraoral ultrasound images.
• Automated segmentation and crest localization can improve the efficiency and reliability of periodontal diagnosis.
• This technology could become a valuable tool for both general dentists and specialists in routine dental practice.