Automated Classification of Bone and Air Volumes for Hybrid PET-MRI Brain Imaging

Abstract

In clinically applicable structural magnetic resonance images (MRI), bone and air have similarly low signal intensity, making the differentiation between them a very challenging task. MRI-based bone/air segmentation, however, is a critical step in some emerging applications, such as skull atlas building, MRI-based attenuation correction for Positron Emission Tomography (PET), and MRI-based radiotherapy planning. In view of the availability of hybrid PET-MRI machines, we propose a voxel-wise classification method for bone/air segmentation. The method is based on random forest theory and features extracted from structural MRI and attenuation uncorrected PET. The Dice Similarity Score (DSC) score between the segmentation result and the 'ground truth' obtained by thresholding Computed Tomography images was calculated for validation. Images from 10 subjects were used for validation, achieving a DSC of 0.83±0.08 and 0.98±0.01 for air and bone, respectively. The results suggest that structural MRI and uncorrected PET can be used to reliably differentiate between air and bone.