Multiclass CBCT Image Segmentation for Orthodontics with Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Dentistry
• Orthodontics

Background:

• Accurate segmentation of jaw and teeth in cone beam computed tomography (CBCT) is crucial for orthodontic diagnosis and treatment planning.
• Existing methods often require manual intervention or only segment either the jaw or teeth, lacking simultaneous multiclass segmentation capabilities.

Purpose of the Study:

• To train and validate a mixed-scale dense (MS-D) convolutional neural network for automated multiclass segmentation of jaw, teeth, and background in CBCT scans.
• To compare the performance of multiclass segmentation with binary segmentation approaches for jaw or teeth.

Main Methods:

• A mixed-scale dense (MS-D) convolutional neural network was developed for simultaneous segmentation of jaw, teeth, and background in 30 CBCT scans.
• Manual segmentation by four dentists served as the gold standard for comparison.
• Performance was evaluated using Dice similarity coefficient and surface deviation, with segmented scans converted to 3D models.

Main Results:

• The MS-D network achieved high overlap with gold standard segmentations (Dice similarity coefficient: 0.934 for jaw, 0.945 for teeth).
• Generated 3D models exhibited minor surface deviations (0.390 mm for jaw, 0.204 mm for teeth).
• Automated segmentation took approximately 25 seconds per scan, compared to about 5 hours for manual segmentation.

Conclusions:

• The MS-D network provides accurate multiclass segmentation of jaw and teeth in CBCT scans, comparable to binary segmentation methods.
• This automated approach significantly reduces segmentation time, enhancing the feasibility of patient-specific orthodontic treatment planning.
• The developed MS-D network offers a promising solution for efficient and accurate analysis of orthodontic CBCT data.