Abstract
Ultra-high-field fMRI at 7 T enables submillimeter spatial resolution but is fundamentally limited by thermal noise, which constrains temporal signal-to-noise ratio (tSNR) and statistical power in task activation studies. NORDIC, the current standard for thermal noise suppression, performs local, patch-wise decomposition using both spatial and temporal redundancy within each patch, but requires complex-valued (magnitude and phase) data. We present STUNet (SpatioTemporal U-Net), a supervised deep learning model that denoises 7 T task fMRI using magnitude data only, eliminating the need for complex-valued acquisition. STUNet combines a 3D U-Net encoder-decoder with a bidirectional convolutional GRU bottleneck and a lightweight temporal mixing block to jointly exploit spatial and temporal redundancies within a sliding window of consecutive volumes, applying a global, learned, nonlinear spatiotemporal model rather than NORDIC’s local, linear, patch-based decomposition. The model was trained on noisy–clean pairs derived from repeated in-house 7 T acquisitions across seven healthy volunteers and evaluated against NORDIC and unprocessed data on a held-out test scan from an unseen volunteer, paradigm, resolution, and repetition time. STUNet improved temporal SNR relative to unprocessed data (7.89 to 10.23, NORDIC: 12.98) and volume SNR (10.86 to 11.52, comparable to NORDIC’s 11.39), and yielded activation sensitivity in AFNI-derived GLM maps with a peak GLM t-statistic improvement (29.92 to 31.66), closely comparable to NORDIC (32.12), though with spatially over-extended activation clusters relative to NORDIC, attributable in part to the total variation regularization term used during training. Beyond its quantitative performance, STUNet’s magnitude-only design halves the per-session data and storage requirements relative to NORDIC, offering a practical advantage for research and clinical workflows in which phase data are not retained. 7T fMRI Deep Learning Image Denoising Spatiotemporal NORDIC
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RIME_022.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=N7mgl31YwM
BibTex
@InProceedings{LobTho_STUNet_MICCAISAT2026,
author = { Loboy Ramos, Thomas Alan AND Solana, Ana Beatriz AND Fernandez, Brice AND Wiesinger, Florian AND Wiestler, Benedikt AND Menzel, Marion Irene},
title = { { STUNet: A Spatiotemporal Deep Learning Approach for 7T fMRI Denoising } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}