MRI is a widely-used imaging modality which offers unique soft-tissue contrast and provides a wealth of anatomical and functional information. However, MRI is inherently slow and signal-to-noise ratio (SNR)-limited, resulting in variable diagnostic image quality and limiting statistical power for research studies. Particularly clinically relevant SNR-starved applications are diffusion MRI (dMRI) and functional (fMRI) for surgical planning (e.g., in functional neurosurgery and in brain tumors). dMRI suffers from long scan times, low resolution and subject motion; BOLD fMRI response signal changes are only about 3% using 3T MRI. State-of-the-art denoising methods, based on image models or smoothing, result in partial-volume effects and loss of fine anatomical detail. We have identified an untapped reserve for significant noise reduction in clinically feasible MRI protocols resulting in SNR increase and Rician MRI noise floor decrease by factors of up to 5-fold, using a model-free noise reduction (denoising) and image reconstruction technique, based on random matrix theory. It does not rely on user-specific input, and outperforms state-of-the-art denoising methods. Our method allows us to identify and remove a pure thermal noise contribution in the principal component analysis (PCA) representation of an MRI data matrix. Remarkably, while noise enters randomly in each voxel's signal, its contribution to the principal components becomes deterministic, when signals from large number of voxels and inequivalent acquisitions (e.g., q-space, time-domain, coils) are combined, which allows us to identify and remove pure- noise components. The key to our MP-PCA method is acquisition redundancy, such that the bulk of the PCA spectrum is dominated by the noise, whose contribution can then be identified and removed. While we initially exploited redundancy in the dMRI q-space, our preliminary findings show it is also present in multi-coil arrays, and in the temporal domain of fMRI. The main goals of this study are: To develop and optimize the MP-PCA denoising framework at the level of multi-coil image reconstruction and to evaluate its accuracy and precision in dMRI (Aim 1); to evaluate its clinical utility for increasing dMRI resolution in functional neurosurgery, based on the ground-truth derived from MR-guided ultrasound intra-operative feedback (Aim 2); and to evaluate its clinical utility for decreasing fMRI scan time in preoperative planning of brain tumor resections (Aim 3). Fundamentally, this project will establish an objective framework to quantify the information content of different MRI modalities, by separating between the signal and the noise. Its applications to dMRI and fMRI, together with using multi-coil redundancy, will lead to maximal possible SNR, thereby reducing scan time, and improving resolution, precision, sensitivity and diagnostic utility of clinically relevant MRI protocols.
MRI, while offering unique soft-tissue contrast, anatomical and functional information, remains inherently slow and signal-to-noise ratio (SNR)-starved, which limits its spatial resolution, lengthens scans, and makes them prone to motion artifacts, thus reducing diagnostic quality. We have identified an untapped reserve for significant noise reduction in clinically feasible MRI acquisitions resulting in SNR increase by up to 5-fold, using a model-free post-processing noise reduction technique, based on random matrix theory. The main goals of this study are to evaluate the robustness and utility of our noise-reduction and reconstruction framework in terms of increasing resolution and shortening scan time, as well as to translate it into functional neurosurgery and preoperative planning for brain tumors.