Abstract
Four-dimensional (4D) flow MRI encodes three velocity components over the cardiac cycle in a 10–20 minute scan; reaching the recommended 5–10 minutes calls for high undersampling, where velocity-derived measures degrade far faster than image magnitude. We present XF-FlowNet, our entry to the CMRx4DFlow2026 challenge, which keeps the widely used serial unrolled cascade skeleton — a residual regularizer followed by a data-consistency (DC) step, repeated K times — and adds three components targeted at the velocity field: refinement in the temporal-frequency (x–f) domain with an explicit |f| positional encoding, conditioning of the regularizer on the sampling mask spectrum and the current DC residual, and a zero-initialized gated conjugate-gradient DC block. On 30 cases at five acceleration rates (R=10–50, 150 reconstructions) it improves over the same skeleton without these components by +0.0156 SSIM, -0.0113 nRMSE, -0.0611 RelErr and -5.20° angular error, with every metric improving in essentially all 150 reconstructions (p<0.001, paired Wilcoxon). Run with half the cascades, the same weights clear the challenge quality gate on all 32 official cases while reducing inference time by 40%.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRxRecon2026_008.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=l5zis7GRS9
BibTex
@InProceedings{JunJi_Separating_MICCAISAT2026,
author = { Jung, Ji-Hoon AND Kang, Jihun AND Ha, Hojin AND Lee, June-Goo},
title = { { Separating Aliasing from Flow: Temporal-Frequency Regularization for Unrolled 4D flow MRI Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17267},
month = {pending},
page = {pending}
}