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

Analyzing and interpreting cone-beam computed tomography (CBCT) images is a complicated and often time-consuming process. In this study, we present two different architectures of multi-channel deep learning (DL) models: "Ensemble" and "Synchronized multi-channel", to automatically identify and classify skeletal malocclusions from 3D CBCT craniofacial images. These multi-channel models combine three individual single-channel base models using a voting scheme and a two-step learning process, respectively, to simultaneously extract and learn a visual representation from three different directional views of 2D images generated from a single 3D CBCT image. We also employ a visualization method called "Class-selective Relevance Mapping" (CRM) to explain the learned behavior of our DL models by localizing and highlighting a discriminative area within an input image. Our multi-channel models achieve significantly better performance overall (accuracy exceeding 93%), compared to single-channel DL models that only take one specific directional view of 2D projected image as an input. In addition, CRM visually demonstrates that a DL model based on the sagittal-left view of 2D images outperforms those based on other directional 2D images.Clinical Relevance— the proposed method aims at assisting orthodontist to determine the best treatment path for the patient be it orthodontic or surgical treatment or a combination of both.