Reducing geometric distortions of Diffusion-Weighted Imaging using Compressed Sensing

Abstract

Diffusion-weighted imaging (DWI) is a type of contrast imaging used in Magnetic Resonance Imaging (MRI) that visualizes the amount of diffusion of water molecules in tissue. Tumors are well visible on DWI images. DWI is often acquired with Echo Planar Imaging (EPI) techniques. Unfortunately, these techniques lead to geometric distortions in the diffusion-weighted images. This causes problems in locating the position of the tumor exactly, which is required for radiotherapy. In this thesis, an approach called Compressed Sensing (CS) was investigated as a technique to reduce the geometric distortions. In theory, the distortions are reduced by obtaining less MR data during scan acquisition (undersampling). By enforcing sparsity of the data in a transform domain, a well reconstructed image can be obtained as the solution of an appropriate minimization problem. The reconstruction algorithm used to solve this problem was cFISTA, a modification of FISTA developed by Beck and Teboulle. Five undersampling strategies were retrospectively used on a DWI patient data set and the best strategy among these five was identified. The reconstruction quality of the whole image and the quality of the tumor reconstruction were assessed using the so-called Structure Similarity Image Measure. A strategy called centerincreased gave the best balance between the average percentage of the MR data required for high quality reconstruction and the variation between the test images, for both the tumor reconstruction and reconstruction of the whole image. High quality reconstructions were obtained for this strategy, when on average only 20% of the MR data was included.