Machine learning for CBCT segmentation of craniomaxillofacial bony structures

The diagnosis and treatment planning of patients with craniomaxillofacial (CMF) deformities requires a precise three-dimensional (3D) skeletal model. Cone-beam computed tomography (CBCT) is routinely used to this end, by annotating the CMF bones (i.e., maxilla and mandible) from the CBCT volume. However, due to the poor quality of CBCT images, e.g., various image artifacts and very low signal-to-noise ratio, segmentation of CMF bones is a very challenging task, which costs an experienced surgeon roughly 12 h. For more efficient and objective CMF bony segmentation, multiple machine learning methods have been applied to process low-quality CBCT images in a fully automated, data-driven way. In this chapter, we introduce a state-of-the-art machine learning method, i.e., the prior-guided sequential random forests, for CBCT segmentation of CMF bones. Specifically, the method first uses a set of expert-segmented CBCT images as the atlases to perform majority voting for the estimation of the initial segmentation probability maps for an input CBCT image. Guided by the contextual prior provided by the initial probability maps, an auto-context random forest is constructed, which uses the appearance and contextual features extracted from the original CBCT image and initial probability maps to predict updated and more precise probability maps of maxilla and mandible. Such random forests are recursively constructed as a sequence, by extracting updated contextual features from the probability maps estimated by the preceding random forest. They form a deep hierarchy of classifiers for step-by-step improved segmentation of CMF bones. This method was evaluated on 30 CBCT images for patients with nonsyndromic dentofacial deformities, treated with double-jaw orthognathic surgery. The experimental results demonstrated its superior performance in terms of accuracy, compared with previous automated methods. More importantly, it only takes around 20 min to consume a CBCT volume, which is significantly faster than an experienced CMF surgeon, suggesting its practical usage in clinical applications.