Optimizing Magnetic Resonance Imaging for Image-Guided Radiotherapy

Abstract

Magnetic resonance imaging (MRI) is playing an increasingly important role in image-guided radiotherapy. MRI provides excellent soft tissue contrast, and is flexible in characterizing various tissue properties including relaxation, diffusion and perfusion. This thesis aims at developing new image analysis and reconstruction algorithms to optimize MRI in support of treatment planning, target delineation and treatment response assessment for radiotherapy.

First, unlike Computed Tomography (CT) images, MRI cannot provide electron density information necessary for radiation dose calculation. To address this, we developed a synthetic CT generation algorithm that generates pseudo CT images from MRI, based on tissue classification results on MRI for female pelvic patients. To improve tissue classification accuracy, we learnt a pelvic bone shape model from a training dataset, and integrated the shape model into an intensity-based fuzzy c-menas classification scheme. The shape-regularized tissue classification algorithm is capable of differentiating tissues that have significant overlap in MRI intensity distributions. Treatment planning dose calculations using synthetic CT image volumes generated from the tissue classification results show acceptably small variations as compared to CT volumes. As MRI artifacts, such as B1 filed inhomogeneity (bias field) may negatively impact the tissue classification accuracy, we also developed an algorithm that integrates the correction of bias field into the tissue classification scheme. We modified the fuzzy c-means classification by modeling the image intensity as the true intensity corrupted by the multiplicative bias field. A regularization term further ensures the smoothness of the bias field. We solved the optimization problem using a linearized alternating direction method of multipliers (ADMM) method, which is more computational efficient over existing methods.

The second part of this thesis looks at a special MR imaging technique, diffusion-weighted MRI (DWI). By acquiring a series of DWI images with a wide range of b-values, high order diffusion analysis can be performed using the DWI image series and new biomarkers for tumor grading, delineation and treatment response evaluation may be extracted. However, DWI suffers from low signal-to-noise ratio at high b-values, and the multi-b-value acquisition makes the total scan time impractical for clinical use. In this thesis, we proposed an accelerated DWI scheme, that sparsely samples k-space and reconstructs images using a model-based algorithm. Specifically, we built a 3D block-Hankel tensor from k-space samples, and modeled both local and global correlations of the high dimensional k-space data as a low-rank property of the tensor. We also added a phase constraint to account for large phase variations across different b-values, and to allow reconstruction from partial Fourier acquisition, which further accelerates the image acquisition. We proposed an ADMM algorithm to solve the constrained image reconstruction problem. Image reconstructions using both simulated and patient data show improved signal-to-noise ratio. As compared to clinically used parallel imaging scheme which achieves a 4-fold acceleration, our method achieves an 8-fold acceleration. Reconstructed images show reduced reconstruction errors as proved on simulated data and similar diffusion parameter mapping results on patient data.