Robust and Efficient Learning of High-Resolution Brain MRI Reconstruction from Small Referenceless Data

PROJECT SUMMARY/ABSTRACT Neuropsychiatric (mental, behavioral and neurological) disorders are increasingly dominating the burden on US healthcare. Yet, our understanding of such disorders is largely restricted to a description of symptoms, and the treatments remain palliative. Several large-scale efforts, including the Human Connectome Project (HCP) and the BRAIN Initiative call for the development of technologies to map brain circuits to improve our understanding of brain function. Magnetic resonance imaging (MRI) plays a central role in these initiatives as a powerful non-invasive methodology to study the human brain, including anatomical, functional and diffusion imaging. Yet, MRI methods have major limitations on achievable resolutions and acquisition speed. These affect both high resolution whole brain acquisitions that aim to image voxel volumes that contain only a few thousand neurons for improved understanding of the brain, and also the more commonly utilized research and clinical protocols. This, in turn, necessitates improved reconstruction methods to facilitate faster acquisitions. Several strategies have been proposed for improved reconstruction of MRI data. Recently, deep learning (DL) has emerged as an alternative for accelerated MRI showing improved quality over conventional approaches. However, it also faces challenges that hinder its utility, especially in high-resolution brain MRI, including need for large databases of reference data for training, concerns about generalization to unseen pathologies not well-represented in training datasets, robustness issues related to recovery of fine structures, and difficulties in training networks for processing multi-dimensional image series. In this proposal, we will develop and validate robust and efficient learning strategies for high-resolution brain DL MRI reconstruction without large databases of reference data. We will develop self-supervised learning methods for training with small referenceless databases or in a scan-specific manner. We will augment these with uncertainty-guided training strategies for improved recovery of areas with high uncertainty, methods for synergistically combining random matrix theory based denoising with DL reconstruction, and memory-efficient distributed learning techniques to process large image series. Our developments will enable at least a two-fold improvement in acceleration rates over existing protocols, and at higher resolutions. They will be validated on HCP-style acquisitions with extensive anatomical, functional and microstructural evaluation at multiple resolutions. Finally, we will curate a whole brain sub-millimeter HCP-style database for studying functional and structural connectivity at the level cortical layers and columns, while also facilitating technical developments for new modeling, image processing and reconstruction algorithms. Successful completion of this project has the potential to transform the scales that can be imaged with MRI, improve the quality of existing protocols and/or significantly reduce scan times, leading to reductions in healthcare costs, improved diagnosis and/or increased patient throughput.