Title: Deep Learning-based Reconstruction of Accelerated, Multi-Coil 3D MR Imaging and MR Spectroscopic Imaging Data

Brief Description: This work proposes developing an efficient 3D, complex-valued reconstruction foundation model that can reconstruct multi-parametric, multi-coil 3D MR imaging volumes for both standard MRI and MR spectroscopic imaging (MRSI).

This project explores deep learning methods for reconstructing accelerated, multi-coil 3D MR imaging and MR spectroscopic imaging data. The long-term goal is to develop a robust reconstruction framework that improves image fidelity, supports physically consistent outputs, models the imaging system, and generalizes across contrasts and acquisition settings.

Purpose: My primary research focus has been on MR spectroscopy. Current research is trending towards MRSI with increasingly higher spatial resolutions. These acquisitions have very long scan times making them impractical without acceleration techniques. This work seeks to create a reconstruction model that can accurately and faithfully reconstruct 3D imaging volumes from accelerated acquisitions, which will ultimately lead to better clinical translation of MRSI techniques. My specific interests are in developing these techniques for studying and modeling brain tumors with the end goal of improving personalized precision oncology to yield better noninvasive diagnostics and treatment planning using the metabolic information from MRSI.

Motivation: Modern 3D MRI and MRSI can produce highly informative data, but accelerated acquisitions often create challenging inverse problems. My goal is to study reconstruction strategies that make better use of the available measurement data while preserving clinically and scientifically meaningful structure. Once a reconstruction strategy has been validated, it can be expanded to further reduce scan times by accurately reconstructing more strongly accelerated acquisitions.

Direction: The project will investigate a compact but flexible reconstruction framework for complex-valued 3D data. The emphasis is on methods that remain stable across varying acquisition conditions and that can adapt to different resolutions, contrasts, and coil configurations. The work will begin with standard 3D MRI using the fastMRI Brain dataset and will later extend toward MRSI.

Novel Proposals: This work plans to use a dual-domain (image and frequency) model which has been shown to outperform single domain models. To do so, the initial architecture proposal will use convolutions and mamba blocks to model the specific nature of features and how they are related in their respective domains. Complex-valued reconstructions have also been shown to outperform real-valued real/imag reconstructions. To implement complex reconstructions, specialized loss functions must be test to ensure the balanced preservation of both magnitude and phase information. And finally, most approaches use denoising methods to reconstruct imaging data. In the multi-coil setting, it would be helpful to leave behind residual Gaussian noise that can be further reduced during coil-combination. So several loss functions will be proposed and studied for their ability to control residual noise statistics.

Building on existing works:

This effort builds on my Augmentrum project, which has been accepted for ISMRM 2026. The associated paper is currently in preparation and will be submitted soon. Public details will be expanded as the project develops. The purpose of Augmentrum is to provide a unified framework for processing and augmenting MRI, MRS, and MRSI data using spectro-temporal, spatial, and k-space augmentations to make more efficient use of limited in vivo datasets. The modeling capabilities of Augmentrum will help train models that can generalize across contrasts and k-space trajectories.

This work will also build on my work titled "Denoising for Physical Fidelity: Feature Enhancement via Spatial Conditioning on Multi-Coil Transients Improves Metabolite Quantification of Noisy Spectra". Preliminary work was presented at the ISMRM Spectroscopy Workshop in Boston in 2024 and the manuscript is currently wrapping up. The reconstruction proposal will expand this successful denoising work from the spectral domain to the imaging domain.

Current State:

Development of the deep learning framework (data loader, preprocessing, and augmentation pipeline, the network architecture, and the optimization procedures) has already begun but has not been finalized. This development is part of what highlighted the need for this donation proposal in order to continue. 

This work is computationally demanding and benefits from a system with substantial unified memory and fast local storage. A compact AMD-based machine would provide an efficient development platform for model prototyping, training, and validation. 3D models already have large memory footprints, but multi-coil 3D data volumes are even larger and have heavier requirements. 

In addition to advancing my own research, this project will support student thesis opportunities. I plan to offer several master theses to help me explore different combinations of loss functions, neural network architectures, and training paradigms, such as diffusion and diffusion bridges, unrolled models, and equilibrium implementations. This will provide students with hands-on learning experience working on state-of-the-art research using clinical data and cutting edge deep learning models.