Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Image Priors for Magnetic Resonance Fingerprinting with pretrained Bloch-consistent denoising autoencoders

Perla Mayo, Matteo Cencini, Ketan Fatania, Carolin M. Pirkl, Marion I. Menzel, Bjoern H. Menze, Michela Tosetti, Mohammad Golbabaee

TL;DR

This work tackles the challenge of recovering quantitative MR parameter maps from undersampled MRF data by integrating a Deep Image Prior with a Bloch-consistent denoising autoencoder (B_DAE). The proposed BARDIP framework combines a pretrained Bloch-consistent autoencoder with a DIP-based Unet prior and a data-consistency plus Bloch-regularized multitask loss, enabling ground-truth-free reconstruction with substantially faster convergence than prior DIP-MRF approaches. Quantitative results on simulated and in-vivo brain data show BARDIP achieving lower $T1$ and $T2$ errors (MAPE) and competitive $PD$ quality, often stabilizing within 1k–10k iterations. The method offers a practical, ground-truth-free alternative for fast, reliable multi-parametric mapping in MRI, with potential impact on clinical workflows due to reduced scan times and improved robustness.

Abstract

The estimation of multi-parametric quantitative maps from Magnetic Resonance Fingerprinting (MRF) compressed sampled acquisitions, albeit successful, remains a challenge due to the high underspampling rate and artifacts naturally occuring during image reconstruction. Whilst state-of-the-art DL methods can successfully address the task, to fully exploit their capabilities they often require training on a paired dataset, in an area where ground truth is seldom available. In this work, we propose a method that combines a deep image prior (DIP) module that, without ground truth and in conjunction with a Bloch consistency enforcing autoencoder, can tackle the problem, resulting in a method faster and of equivalent or better accuracy than DIP-MRF.

Deep Image Priors for Magnetic Resonance Fingerprinting with pretrained Bloch-consistent denoising autoencoders

TL;DR

This work tackles the challenge of recovering quantitative MR parameter maps from undersampled MRF data by integrating a Deep Image Prior with a Bloch-consistent denoising autoencoder (B_DAE). The proposed BARDIP framework combines a pretrained Bloch-consistent autoencoder with a DIP-based Unet prior and a data-consistency plus Bloch-regularized multitask loss, enabling ground-truth-free reconstruction with substantially faster convergence than prior DIP-MRF approaches. Quantitative results on simulated and in-vivo brain data show BARDIP achieving lower and errors (MAPE) and competitive quality, often stabilizing within 1k–10k iterations. The method offers a practical, ground-truth-free alternative for fast, reliable multi-parametric mapping in MRI, with potential impact on clinical workflows due to reduced scan times and improved robustness.

Abstract

The estimation of multi-parametric quantitative maps from Magnetic Resonance Fingerprinting (MRF) compressed sampled acquisitions, albeit successful, remains a challenge due to the high underspampling rate and artifacts naturally occuring during image reconstruction. Whilst state-of-the-art DL methods can successfully address the task, to fully exploit their capabilities they often require training on a paired dataset, in an area where ground truth is seldom available. In this work, we propose a method that combines a deep image prior (DIP) module that, without ground truth and in conjunction with a Bloch consistency enforcing autoencoder, can tackle the problem, resulting in a method faster and of equivalent or better accuracy than DIP-MRF.
Paper Structure (10 sections, 4 equations, 3 figures, 1 table)

This paper contains 10 sections, 4 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. BARDIP Overview
  4. Training Models
  5. Numerical Experiments
  6. Results
  7. Discussion
  8. Conclusion
  9. Compliance with Ethical Standards
  10. Acknowledgments

Figures (3)

  • Figure 1: (a) Proposed pipeline composed of three neural networks, two of them pretrained. The computation of the forward operator as well as the use of the $\textbf{B}_{\textbf{DAE}}$ module enforce the consistency with the data acquisition model and the Bloch equations. (b) Unet used for the experiments reported.
  • Figure 2: T1/T2 reconstruction errors vs iterations for learned approaches on simulated data. Dashed lines indicate the lowest value found across all iterations for the respective approach. Top: SNR 40. Bottom: SNR 35.
  • Figure 3: Tissue maps of the estimations of both approaches on the single slice of real data assessed with L=1000 and L=500. T1 and T2 are in seconds.