Electronic Thesis and Dissertation Repository

Degree

Doctor of Philosophy

Program

Medical Biophysics

Supervisor

Drangova, Maria

Abstract

Magnetic Resonance Imaging (MRI) is a tremendously useful diagnostic imaging modality that provides outstanding soft tissue contrast. However, subject motion is a significant unsolved problem; motion during image acquisition can cause blurring and distortions in the image, limiting its diagnostic utility. Current techniques for addressing head motion include optical tracking which can be impractical in clinical settings due to challenges associated with camera cross-calibration and marker fixation. Another category of techniques is MRI navigators, which use specially acquired MRI data to track the motion of the head.

This thesis presents two techniques for motion correction in MRI: the first is spherical navigator echoes (SNAVs), which are rapidly acquired k-space navigators. The second is a deep convolutional neural network trained to predict an artefact-free image from motion-corrupted data.

Prior to this thesis, SNAVs had been demonstrated for motion measurement but not motion correction, and they required the acquisition of a 26s baseline scan during which the subject could not move. In this work, a novel baseline approach is developed where the acquisition is reduced to 2.6s. Spherical navigators were interleaved into a spoiled gradient echo sequence (SPGR) on a stand-alone MRI system and a turbo-FLASH sequence (tfl) on a hybrid PET/MRI system to enable motion measurement throughout image acquisition. The SNAV motion measurements were then used to retrospectively correct the image data.

While MRI navigator methods, particularly SNAVs that can be acquired very rapidly, are useful for motion correction, they do require pulse sequence modifications. A deep learning technique may be a more general solution. In this thesis, a conditional generative adversarial network (cGAN) is trained to perform motion correction on image data with simulated motion artefacts. We simulate motion in previously acquired brain images and use the image pairs (corrupted + original) to train the cGAN.

MR image data was qualitatively and quantitatively improved following correction using the SNAV motion estimates. This was also true for the simultaneously acquired MR and PET data on the hybrid system. Motion corrected images were more similar than the uncorrected to the no-motion reference images. The deep learning approach was also successful for motion correction. The trained cGAN was evaluated on 5 subjects; and artefact suppression was observed in all images.

Share

COinS