Malocclusion Classification on 3D Cone-Beam CT Craniofacial Images Using Multi-Channel Deep Learning Models^.

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Orthodontic Diagnostics

Background:

• Interpreting cone-beam computed tomography (CBCT) for skeletal malocclusions is complex and time-consuming.
• Current diagnostic methods may lack the efficiency and accuracy needed for optimal treatment planning.

Purpose of the Study:

• To develop and evaluate novel multi-channel deep learning (DL) models for automated identification and classification of skeletal malocclusions.
• To compare the performance of multi-channel DL models against single-channel models using 3D CBCT craniofacial images.
• To utilize Class-selective Relevance Mapping (CRM) for model interpretability and identifying key diagnostic features.

Main Methods:

• Two multi-channel DL architectures, 'Ensemble' and 'Synchronized multi-channel', were designed.
• Models processed 2D images derived from three directional views of a single 3D CBCT scan.
• Class-selective Relevance Mapping (CRM) was employed for visualizing model decision-making processes.

Main Results:

• Multi-channel DL models achieved accuracy exceeding 93%, significantly outperforming single-channel models.
• The 'Ensemble' and 'Synchronized multi-channel' architectures demonstrated superior performance in classifying skeletal malocclusions.
• CRM analysis indicated that sagittal-left view 2D images provided the most discriminative information for the DL models.

Conclusions:

• Multi-channel deep learning models offer a highly accurate and efficient method for analyzing 3D CBCT images to detect skeletal malocclusions.
• The proposed DL approach can assist orthodontists in determining appropriate treatment strategies, including orthodontic, surgical, or combined interventions.
• Model interpretability through CRM enhances trust and understanding of the automated diagnostic process.