Model-based deep learning for inverse problems in imaging

Long imaging times have been a limitation for many Magnetic Resonance Imaging (MRI) applications. High resolution brain imaging using MRI is often needed for treating patients with neurodegenerative disorders and it becomes prohibitive due to slowness of the process. Traditionally, the data acquisition process is quickened by collecting fewer samples and that provides a noisy and blurred image of the organ scanned. Researchers have proposed several image reconstruction algorithms to predict a clean and sharp image of the organ, as desired by radiologists for diagnostic purposes.

State-of-the art image reconstruction algorithms are deep learning based and have some limitations. In the context of parallel MRI application, these methods are fast but require extensive memory which discourages high resolution or higher dimensional image reconstruction. In addition, these methods require the MRI scanner's coil sensitivity information explicitly and thus, are vulnerable to motion artifacts introduced due to mismatch between the calibration and main scans. Current approaches are also not fit for deployment in the clinical setup since multiple acquisition specific deep learning models are required to be trained for several acquisition settings which would also require collecting a lot of training data.

In this thesis, we come up with novel ideas to solve these problems. Specifically, in one of the work, we propose a memory-efficient deep learning approach. We show its benefit in reconstructing three-dimensional cardiac MRI data which would not have been possible using the existing deep learning approaches mainly due to memory constraints. An additional benefit of this method is in it being robust to adversarial attacks which is a very common problem with deep learning. In another work, we propose a calibration-less parallel MRI recovery approach that gets rid of the coil sensitivity estimation step and hence provides robustness to mismatch in scans. This approach learns sensitivity information on the y and also performs at par with existing methods. Finally, we work on making the existing techniques adaptive to multiple acquisition settings which are often encountered in a clinical setup. The proposed adaptive method requires much fewer training data than the conventional methods.

Thus, we have taken a big stride towards solving some of the common problems faced by image reconstruction algorithms for MRI. These methods are generalizable and extendable to other imaging applications such as denoising, super-resolution and also other imaging modalities including Computed Tomography (CT).