Abstract
Four-dimensional flow magnetic resonance imaging (4D Flow MRI) enables time-resolved quantification of blood-flow dynamics, but long acquisition times limit its routine clinical use. Reconstruction from highly undersampled multi-coil k-space is challenging because complex-valued errors can degrade both anatomical image quality and phase-derived velocity measurements. In this work, we present a training-oriented enhancement of the physics-informed FlowVN framework for the regular tracks of the CMRx4DFlow2026 challenge. The proposed method retains the original unrolled inference architecture and introduces two complementary training strategies: progressive acceleration-factor curriculum learning and ROI-aware joint complex- and magnitude-domain supervision. The curriculum gradually introduces severe undersampling conditions, allowing a single model to learn robust reconstruction priors across acceleration factors from 10× to 50×. The ROI-aware objective emphasizes the aortic region while preserving surrounding anatomical context, improving both structural fidelity and velocity-sensitive phase recovery. Experimental results demonstrate that the proposed strategies provide balanced improvements in magnitude-image and velocity-field reconstruction over the original FlowVN baseline. On the validation set, the proposed method achieved an SSIM of 0.935671, an nRMSE of 0.053034, a RelErr of 0.474395, an AngErr of 37.341468, and a ComplexDiffErr of 0.104072, outperforming the original FlowVN across all five metrics. Since both components are applied only during training, the inference architecture and computational cost remain unchanged, supporting evaluation in both reconstruction-quality and efficiency-oriented settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRxRecon2026_005.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/CMRxRecon2026_005_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=wwynt0I6kd
BibTex
@InProceedings{LiuYaj_Curriculum_MICCAISAT2026,
author = { Liu, Yajun AND Zhu, Pengyu AND Wang, Yinlong AND Meng, Hui},
title = { { Curriculum Learning and ROI-Aware Supervision for Highly Accelerated 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}
}