Novel Models for High-Dimensional Imaging: High-Resolution fMRI Acceleration and Quantification

TL;DR

This thesis addresses the challenge of achieving high spatial and temporal resolution in fMRI while maintaining strong SNR. It develops three core advances: (1) OSSI, a high-SNR acquisition with oscillating steady-state signals, (2) patch-tensor low-rank reconstruction to exploit local spatiotemporal redundancies for fast 3D OSSI imaging, and (3) a physics-based manifold (OSSIMM) for joint reconstruction and dynamic quantification of $R_2^*$ and related parameters. A voxel-wise temporal attention network is introduced to model dynamic MRI sequences with limited data, offering faster reconstruction and improved functional maps. Together, these approaches enable substantial accelerations (up to 12× in 2D/3D OSSI) and higher activation detection with robust tSNR gains, while also supporting dynamic quantification of tissue properties. The work demonstrates OSSI-based fMRI can achieve high-resolution imaging and quantitative physics maps without requiring higher magnetic field strength, promising practical impact for neuroscience research and clinical imaging.

Abstract

The goals of functional Magnetic Resonance Imaging (fMRI) include high spatial and temporal resolutions with a high signal-to-noise ratio (SNR). To simultaneously improve spatial and temporal resolutions and maintain the high SNR advantage of OSSI, we present novel pipelines for fast acquisition and high-resolution fMRI reconstruction and physics parameter quantification. We propose a patch-tensor low-rank model, a physics-based manifold model, and a voxel-wise attention network. With novel models for acquisition and reconstruction, we demonstrate that we can improve SNR and resolution simultaneously without compromising scan time. All the proposed models outperform other comparison approaches with higher resolution and more functional information.

Figures (71)

• Figure 2.1: Simulation of OSSI spin behavior and signals. (a) and (b) Periodic motion of magnetization through RF pulses (the filled squares are at the end of the RF pulse) and free precession for a gray matter spin at -20 Hz off-resonance frequency, $T_1$ = 1433.2 ms, $T_2$ = 92.6 ms, TR = 15 ms, $n_c$ = 10, and FA = 10$^\circ$ from two different views. (c) Magnitude signal variation of different isochromats (6.67 Hz apart) for the magnetization in (a) and (b) just after the RF pulse, the black dashed line is the Ernst angle signal for spoiled-GRE. (d) Spin positions during free precession for different isochromats (same isocromats as in (c)) leading to phase dispersion and $T_2^*$-weighting. The cyan circles mark the center of the precession interval.
• Figure 2.2: Simulation for signal properties just after the RF pulse, where the pulse duration was adjusted to minimize off-resonance phase accumulation during the RF pulses (TE $<$ 0.02 ms). The left and right panels show simulated OSSI signal magnitude and phase, respectively. (a) and (b) show magnitude and phase responses as a function of off-resonance frequency and time (TR number), observe the periodicity in time ($\text{T}_\text{OSSI} = n_c\text{TR}$) and frequency (1/TR = 66.67 Hz). (c) and (e) are magnitude response of the signal vs. time and frequency, respectively, and (d) and (f) are the phase responses showing phases after correction for the excitation RF phase. The blue and red lines in temporal plots (c) and (d) correspond to two isochromats at off-resonance -33.33 Hz and -32.67 Hz, respectively. It can be seen that an off-resonance amount of less than 1/TOSSI lead to some modest changes in the shape of the response. The green curve in (e) and (f) are the magnitude and phase of the frequency response, respectively, and indicate the manifold on which the steady-state response exists. The blue and red lines connect 6.67 Hz apart samples of the manifold and start from off-resonance -33.33 Hz and -32.67 Hz respectively. Particularly, by comparing (c) and (e), (d) and (f), it is clearly shown that the time and samples of frequency responses have exactly the same shape, only flipped.
• Figure 2.3: Simulation of acquisition parameters for spiral-out readouts (TE = 1.6 ms). (a) to (c) are $T_2^*$ sensitivity defined as Sactivated – Srest in units of $M_0$ = 1. (a) shows the RMS combined magnitude signal as a function of $n_c$ and flip angle for a fixed TR of 15 ms. Notice the bright spot around $n_c$ = 10 and flip angle = 10$^\circ$. We focus on the region denoted by the blue square for OSSI fMRI acquisition parameter optimization, and the results are in (b) to (e). (b) shows how $T_2^*$ sensitivity varies with TR and flip angle for a fixed $n_c$ = 10. The signal is normalized by $\sqrt{(\text{TR}-c)/\text{TR}} \approx \sqrt{\text{T}_\text{A/D}}$ with $c$ = 5 ms for SNR efficiency. (c) shows how $T_2^*$ sensitivity varies with $n_c$ and flip angle for TR = 15 ms. (d) gives off-resonance sensitivity at different TR and flip angles for $n_c$ = 10. (e) gives off-resonance sensitivity at different $n_c$ and flip angles for TR = 15 ms.
• Figure 2.4: Images of steady state with quadratic phase progression ($n_c$ = 10) with (a) balance gradients (OSSI) and (b) spoiling gradients (GRE). Each panel has 10 images across the periodic phase pattern and is shown twice to demonstrate the reproducibility. The 2-norm combined images are given on the right. The OSSI and GRE images are not on the same intensity scale.
• Figure 2.5: Time courses for a 4-voxel ROI in the phantom for OSSI (red) and GRE (blue). Both before and after 2-norm combination, OSSI shows signal strengths roughly two times larger than the spoiled GRE signal.