Deep learning for accelerated magnetic resonance imaging

Abstract

Medical imaging has aided the biggest advance in the medical domain in the last century. Whilst X-ray, CT, PET and ultrasound are a form of imaging that can be useful in particular scenarios, they each have disadvantages in cost, image quality, ease-of-use and ionising radiation. MRI is a slow imaging protocol which contributes to its high cost to run. However, MRI is a very versatile imaging protocol allowing images of varying contrast to be easily generated whilst not requiring the use of ionising radiation. If MRI can be made to be more efficient and smart, the effective cost of running MRI may be more affordable and accessible. The focus of this thesis is decreasing the acquisition time involved in MRI whilst maintaining the quality of the generated images and thus diagnosis. In particular, we focus on data-driven deep learning approaches that aid in the image reconstruction process and streamline the diagnostic process. We focus on three particular aspects of MR acquisition. Firstly, we investigate the use of motion estimation in the cine reconstruction process. Motion allows us to combine an abundance of imaging data in a learnt reconstruction model allowing acquisitions to be sped up by up to 50 times in extreme scenarios. Secondly, we investigate the possibility of using under-acquired MR data to generate smart diagnoses in the form of automated text reports. In particular, we investigate the possibility of skipping the imaging reconstruction phase altogether at inference time and instead, directly seek to generate radiological text reports for diffusion-weighted brain images in an effort to streamline the diagnostic process. Finally, we investigate the use of probabilistic modelling for MRI reconstruction without the use of fully-acquired data. In particular, we note that acquiring fully-acquired reference images in MRI can be difficult and nonetheless may still contain undesired artefacts that lead to degradation of the dataset and thus the training process. In this chapter, we investigate the possibility of performing reconstruction without fully-acquired references and furthermore discuss the possibility of generating higher quality outputs than that of the fully-acquired references.