Feasibility of generating synthetic CT from T1-weighted MRI using a linear mixed-effects regression model

Abstract

Generation of synthetic computed tomography (sCT) for magnetic resonance imaging (MRI)-only radiotherapy is emerging as a promising direction because it can eliminate the registration error and simplify clinical workflow. The goal of this study was to generate accurate sCT from standard T1-weighted MRI for brain patients. CT and MRI data of twelve patients with brain tumors were retrospectively collected. Linear mixed-effects regression models were fitted between CT and T1-weighted MRI intensities for different segments in the brain. The whole brain sCTs were generated by combining predicted segments together. Mean absolute error (MAE) between real CTs and sCTs across all patients was 71.1 ± 5.5 Hounsfield Unit (HU). Average differences in the HU values were 1.7 ± 7.1 HU (gray matter), 0.9 ± 5.1 HU (white matter), −24.7 ± 8.0 HU (cerebrospinal fluid), 76.4 ± 17.8 HU (bone), 20.9 ± 20.4 HU (fat), −69.4 ± 28.3 HU (air). A simple regression technique has been devised that is capable of producing accurate HU maps from standard T1-weighted MRI, and exceptionally low MAE values indicate accurate prediction of sCTs. Improvement is needed in segmenting MRI using a more automatic approach.