Craniofacial Biomedicine Using Machine Learning From CBCT Images

Objectives: Cone beam computed tomography (CBCT) has increased its applicability in assessing skeletal variations in patients with craniofacial anomalies. Yet there is a limitation regarding excessive time required for segmentation, which prohibits quantitative diagnosis. We hypothesize that artificial intelligent, machine learning, can provide accurate auto-rendition of 3D bony structures for anomalies.
Methods: Two patient cohorts, 30 unilateral canine impaction subjects and 60 unilateral cleft palate, were recruited for the pilot study. In addition, 30 healthy patients were served as the control. A machine learning algorithm utilizing the random forest method was used to auto-landmark and auto-segment the CBCT images to quantify volumetric skeletal maxillae and cleft defects. Prior to learning, CBCT images from 10 canine impaction patients, 10 healthy patients (control), and 30 cleft palate patients were manually segmented by an experienced orthodontist. Using the database of the manually segmented objects, images from all subjects were then auto-segmented using the learning algorithm to calculate the volumes of maxillae and cleft defects. ANOVA and multiple linear regression were used for statistical analyses.
Results: Algorithm reliability was validated, with a minimal mean difference of 2 voxels between manually identified and automatically digitized landmarks. For the canine impaction, no significant difference in bone volume was found between impacted and non-impacted sides of the study group. The maxillae of the impacted patients had a significantly smaller volume (p<0.05) than the control group volume (2.59±0.53x10⁴mm³ and 2.52±0.50x10⁴mm³, respectively). For the cleft patients, defect width was significantly related to maxillary anterior height and width; and defect volume was significantly related to maxillary anterior height.
Conclusions: The average processing time was about 15min for segmentation of a typical set of CBCT images using machine learning, which enlightens the potential for clinical studies with a large sample size. 3D characterization of craniofacial anomalies will aid treatment modalities.