Statistical Methods in Computed Tomography Image Estimation
share
There is increasing interest in computed tomography (CT) image estimations from magnetic resonance (MR) images. This study aims to introduce a novel statistical learning approach for improving CT estimation from MR images. Prior knowledges about tissue-types, roughly speaking non-bone and bone tissue-types from CT images, have been used in collaboration with a Gaussian mixture model (GMM) to explore CT image estimations from MR images. Due to the introduced prior knowledges, GMMs were trained for each of the tissue-type. At the prediction stage, we have no CT image, that is there are no prior knowledges about the tissue-types and thereby we trained RUSBoost algorithm on the training dataset in order to estimate the tissue-types from MR images of the new patient. The estimated RUSBoost algorithm and GMMs were used to predict CT image from MR images of the new patient. We validated the RUSBoost algorithm by applying 10-fold cross-validation while the Gaussian mixture models were validated by using leave-one-out cross-validation of the datasets from the patients. In comparison with the existing model-based CT image estimation methods, the proposed method has improved the estimation, especially in bone tissues. More specifically, our method improved CT image estimation by 23 Hounsfield units (HU) and 6 HU on average for datasets obtained from nine and five patients, respectively. Bone tissue estimations have been improved by 107 HU and 62 HU on average for datasets from nine and five patients, respectively. Evaluation of our method shows that it is a promising method to generate CT image substitutes for the implementation of fully MR-based radiotherapy and PET/MRI applications. Keywords: Computed tomography; magnetic resonance imaging; CT image estimation; supervised learning; Gaussian mixture model
READ FULL TEXT