🛰️ Daily AI Frontier
‹ back to 2026-08-27

FlowMoDL: Model-Based Deep Learning with Conjugate-Gradient Data Consistency for Highly Accelerated 4D Flow MRI Reconstruction

arXiv cs.CV Medical/Healthcare AI Tristan Gottwald, Michelle Bruch, Mubashir-Ul Hassan, Fatma Alickovic, Milan Kloiber, Daniel Tenbrinck, Torsten Panholzer, Melanie Schaller, Jana Hutter 2026-08-26

TL;DR - FlowMoDL is an unrolled neural network for reconstructing highly accelerated 4D flow MRI while preserving anatomical detail and phase-derived blood-flow velocity accuracy. It supports acceleration factors from 10Ă— to 50Ă— with one model and outperforms classical and deep-learning baselines under a limited gradient-step budget.

  • Alternates a learned (3+1)D spatiotemporal denoiser with conjugate-gradient SENSE data-consistency updates.
  • Uses dual-pathway conditioning to adapt denoiser features and data-consistency weighting across acceleration factors.
  • Trains with deep supervision on magnitude, velocity magnitude, and angular errors, stabilized through curriculum learning.
  • On the multi-center CMRx4DFlow dataset, it leads all evaluated baselines across magnitude SSIM, nRMSE, relative velocity error, and angular error.

view merged work →