Abstract
4D flow MRI measures blood velocity over the cardiac cycle, but its long scan time forces heavy undersampling, so images must be reconstructed by an algorithm. A reconstruction is judged on two things: the anatomical (magnitude) image, and the velocity, which lives not in any image but in the phase difference between four velocity-encoded'' acquisitions of the same anatomy. This paper makes one simple point: when a learned network mixes those four encodings inside its convolutional filters (the default channel-mixing design), it tends to disturb the phase differences that carry the velocity, because an unconstrained weighted sum across encodings is not phase-preserving. We build a complex-valued, spatio-temporal unrolled variational network for the CMRx4DFlow2026 challenge and show that denoising each encoding separately (per-encoding’’) instead of jointly cuts the mean velocity-direction error from 29.6° to 18.9° at acceleration R=20 on held-out data, a ∼11° gain from one design choice, where every other change we tried (including matched-capacity variants) moved it by about 1°. We retrain the challenge’s FlowVN baseline and improve it on all four official metrics, and a test-time magnitude–phase fusion lowers the mean angular error to 17.96°, below either alone. For the special tasks (no matching training data or ground truth) we adapt the network at test time by self-supervision. We also decompose the held-out-to-leaderboard gap, and show that conditioning a single per-encoding model on the acceleration beats the retrained FlowVN on all four metrics under the full official protocol.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRxRecon2026_011.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=H8LEE3u7Ag
BibTex
@InProceedings{KolGov_Keep_MICCAISAT2026,
author = { Kolli, Govinda AND Dukre, Adinath Madhavrao AND Razzak, Imran},
title = { { Keep the Encodings Apart: Per-Encoding Complex-Valued Variational Networks for Accelerated 4D Flow MRI } },
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}
}