Schrödinger Diffusion Driven Signal Recovery in 3T BOLD fMRI Using Unmatched 7T Observations

TL;DR

The paper addresses the gap between 3T and 7T fMRI quality for retinotopic mapping by aligning 3T and 7T signals in a shared 2D brain-disk domain and applying an unsupervised Schrödinger Bridge to translate unmatched 7T data into high-SNR 3T reconstructions. The approach combines conformal brain-disk parameterization with a discrete, unpaired Schrödinger Bridge diffusion model learned through neural guidance, enabling high-quality fMRI time series from typical 3T acquisitions. Evaluations on real and synthetic data show improved signal fidelity (lower FID, higher SSIM/PSNR) and enhanced pRF fits (higher $R^2$), suggesting that 7T-level quality can be computationally approximated from 3T scans. This method has practical implications for more accessible high-resolution retinotopic analysis and broader downstream analyses like segmentation and classification in neuroimaging.

Abstract

Ultra-high-field (7 Tesla) BOLD fMRI offers exceptional detail in both spatial and temporal domains, along with robust signal-to-noise characteristics, making it a powerful modality for studying visual information processing in the brain. However, due to the limited accessibility of 7T scanners, the majority of neuroimaging studies are still conducted using 3T systems, which inherently suffer from reduced fidelity in both resolution and SNR. To mitigate this limitation, we introduce a new computational approach designed to enhance the quality of 3T BOLD fMRI acquisitions. Specifically, we project both 3T and 7T datasets, sourced from different individuals and experimental setups, into a shared low-dimensional representation space. Within this space, we employ a lightweight, unsupervised Schrödinger Bridge framework to infer a high-SNR, high-resolution counterpart of the 3T data, without relying on paired supervision. This methodology is evaluated across multiple fMRI retinotopy datasets, including synthetically generated samples, and demonstrates a marked improvement in the reliability and fit of population receptive field (pRF) models applied to the enhanced 3T outputs. Our findings suggest that it is feasible to computationally approximate 7T-level quality from standard 3T acquisitions.

Figures (7)

• Figure 1: Overview of the Pipeline: We collect 3T and 7T fMRI data from the NOD and NSD datasets, respectively. The fMRI signals at each time step from all pRF experiments are used to generate low-quality (LQ) and high-quality (HQ) fMRI slices. For synthetic data, the original NSD fMRI serves as the ground truth, while their down-sampled versions as LQ data. All data are fed into our pipeline with three components.
• Figure 2: Disk conformal parameterization. (a) The full fsaverage cortical surface; (b) the ROI subdivision of the full mesh determined by the FreeSurfer vertex labels fischl2012freesurfer; (c) the parameterized planar disk of the ROI obtained from harmonic map $h$; (d) the refined planar disk through resulting conformal mapping $c = r \circ h$.
• Figure 3: The illustration of training pipeline. For a randomly selected time step $t_i\in \mathbf{t}$, we recursively generate samples following Eq. \ref{['eq:SBsim2']} and Eq. \ref{['eq:SBsim3']} to approximate distribution $\hat{x}_{1 \mid t_i} \sim p(x_1,x_{t_i})$ as discussed in Sec. \ref{['sec:Method-sb']}. All three losses $\mathbb{L}_\text{Adv}, \mathbb{L}_\text{SB}$ and $\mathbb{L}_\text{reg}$ are weighted through individual hyper-parameters.
• Figure 4: Illustration of enhanced BDs. We show the fMRI response on the shared parameterized planar disk.
• Figure 5: BOLD signal from synthetic low-quality data, enhanced data and the ground truth. (a) fMRI plots for the vertex with highest $R^2$ show a significant alignment between the ground truth and the enhanced signals, with slight misalignment on valley points. (b) fMRI plots for the vertex with lower $R^2$ show worse alignment, probably caused by the inactive and small visual response on this vertex.