Abstract
Reducing the stage count of a physics-guided unrolled network accelerates four-dimensional (4D) flow magnetic resonance imaging (MRI) reconstruction but may degrade image and flow fidelity. We study stage-aligned knowledge distillation (KD) from a 16-stage FlowVN teacher (S16) to an eight-stage student (S8). S8 copies alternating teacher stages and receives final-output, trajectory, and two-stage-update supervision; a matched supervised control isolates KD. Across two seeds, 16 held-out cases, and accelerations 10–50, full-KD reduced normalized root-mean-square error, relative velocity error, angular error, and VENC-scaled velocity RMSE by 1.91%, 2.90%, 1.31%, and 2.48%, respectively, and increased structural similarity by 0.0037. In a nested ablation, trajectory matching reduced nRMSE by 0.00250 and VENC-scaled velocity RMSE by 0.763 cm/s, whereas output-only KD was inconclusive for nRMSE and update matching increased nRMSE by 0.00127. On 32 Task 2 validation volumes, full-KD S8 reduced the displayed Complex Difference Error from 0.129 to 0.127 relative to supervised S8, but neither met the organizer FlowVN reference, identified as S10 from the released implementation; S16 obtained 0.085. S8 halved parameter count and fixed-shape forward time (2.62 versus 5.24 s). Separately, execution optimization reduced S8 complete-case wall time from 60.02 to 30.44 s and peak GPU memory from 13.72 to 3.44 GiB, with a saved-output relative L_2 difference of at most 7.60×10^-7. KD therefore provides modest quality recovery at fixed S8 cost, but this recovery is small relative to the observed S8–S16 quality gap.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMRxRecon2026_010.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/CMRxRecon2026_010_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=ZmT3bhaD1w
BibTex
@InProceedings{XiaYul_StageAligned_MICCAISAT2026,
author = { Xiao, Yuliang AND Anvari Hamedani, Kian AND Vavasour, Zach AND Graham, Simon J. AND Chiew, Mark},
title = { { Stage-Aligned Knowledge Distillation for Fast and Memory-Efficient 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}
}