Transformer based tooth classification from cone-beam computed tomography for dental charting.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Dental charting is crucial for various applications but manual methods present challenges like inaccuracy and burden.
• Automated tooth classification from 3D cone-beam computed tomography (CBCT) images offers a solution to these limitations.

Purpose of the Study:

• To develop a deep neural network for accurate automatic tooth classification from 3D CBCT image patches.
• To address the limitations of Transformer networks in 3D medical image analysis by proposing an improved architecture.

Main Methods:

• A novel Grouped Bottleneck Transformer architecture combining CNN and Transformer advantages was developed.
• The network accepts 3D CBCT image patches containing the region of interest (ROI) for tooth classification.
• Experiments were conducted on a clinical dataset (450 training, 104 testing samples) and the MedMNIST3D dataset.

Main Results:

• The proposed network achieved 91.3% classification accuracy and a 99.7% AUC score on the clinical dataset.
• On the MedMNIST3D dataset, the network outperformed existing methods on 5 out of 6 3D medical image subsets.
• The Grouped Bottleneck Transformer mitigates the need for large datasets and reduces computational complexity.

Conclusions:

• The developed deep neural network effectively automates tooth classification from 3D CBCT images.
• The Grouped Bottleneck Transformer architecture shows promise for 3D medical image analysis, overcoming Transformer limitations.
• This automated approach has significant potential to improve dental charting accuracy and efficiency in clinical and forensic settings.