A Single-TR-Resolved Deep Image Prior Reconstruction for Cardiac MR Fingerprinting

Presenting Author(s)

• JH

  Jesse I. Hamilton, PhD

  Assistant Professor
  University of Michigan
  Ann Arbor, Michigan, United States

Primary Author(s)

• CS

  Calder D. Sheagren, PhD

  Postdoctoral Fellow
  University of Michigan
  Ann Arbor, Michigan, United States

Co-Author(s)

• ZL

  Zhongnan Liu, MSc

  Graduate Student Research Assistant
  University of Michigan, United States
• ZL

  Zexuan Liu, MSc

  Graduate Student Research Assistant
  University of Michigan, United States
• GL

  Gastao Lima Da Cruz, PhD

  Scientist
  University of Michigan
  Ann Arbor, Michigan, United States

Background: Cardiac Magnetic Resonance Fingerprinting (MRF) enables multiparametric mapping but can be limited by motion during long breathholds and cardiac acquisition windows.1 Various reconstruction methods have been proposed to shorten acquisition times in MRF, including a low-rank Deep Image Prior (LR-DIP) method.2 However, low-rank constraints may lead to image artifacts or biased tissue property estimates due to unmodeled effects, such as residual motion. We propose a DIP method that reconstructs TR-resolved MRF time-series images without low-rank constraints, providing a flexible framework for accelerated and potentially motion-corrected cardiac MRF.

Methods: This study adapted a Time-Dependent DIP reconstruction originally developed for real-time cine imaging to cardiac MRF, yielding time-series images at a temporal resolution of a single TR (5.5 ms) that are then matched to an MRF dictionary to obtain T1, T2, and M0 maps.3–5 Time-Dependent DIP uses a fixed nonlinear manifold to encode temporal variations; this manifold is sampled for specific time points and input to a convolutional neural network (CNN) to generate time-resolved images (Fig. 1). The following manifold designs were investigated here: (1) a multifrequency manifold containing sinusoids spanning a range of frequencies (first developed to model physiological motion in real-time cine scans)5, (2) an MRF-basis manifold containing basis functions from the SVD of the MRF dictionary, and (3) a combination of both manifolds. First, the XCAT phantom was used to simulate a 2D breathheld ECG-triggered MRF scan, comparing reconstruction accuracy to both LR-DIP and a sparse and locally low-rank reconstruction (SLLR).6 Next, in vivo evaluation was performed in 10 healthy participants and 5 cardiomyopathy patients, with myocardial T1 and T2 values measured at a mid-short axis slice.

Results: In simulations, the proposed TR-DIP reconstruction with the combination manifold yielded NRMSE values comparable to SLLR, although slightly higher than LR-DIP. Moreover, convergence was generally faster with this manifold compared to the multifrequency and MRF-basis manifolds (Fig. 2). In healthy participants, TR-DIP yielded mean myocardial T1 values comparable to those from LR-DIP (974±102ms vs. 981±82ms). Mean myocardial T2 values were significantly higher compared to LR-DIP (49±5ms vs. 44±4ms), although the magnitude of this difference was relatively small (Fig. 3). The same trend was observed in cardiomyopathy patients, where TR-DIP yielded T1 and T2 values of 996±130ms and 54±9ms compared to 1009±66ms and 46±5ms with LR-DIP.

Conclusion: This work demonstrates the feasibility of a Time-Dependent DIP reconstruction adapted for cardiac MRF that produces single-TR time-series images. By avoiding low-rank constraints, a key advantage of this approach is the potential to derive motion fields from TR-resolved images for motion correction, which will be explored in future work.

Figure 1. A. Illustration of basis elements for the multifrequency, combination, and MRF manifolds. B. Training scheme of TR-DIP, where the manifold is sampled at a single time point and passed through the CNN to generate a complex-valued image frame, which is compared to the acquired data in k-space. C. Inference illustration of TR-DIP, where the CNN is applied to all time points in the manifold to generate TR-resolved images, which are matched to the MRF dictionary to obtain quantitative T1, T2, and M0 maps.
Figure 2: XCAT simulation results showing RMSE for T1 and T2 as a function of training epochs using the TR-resolved DIP reconstruction with multifrequency, MRF, and combination manifold designs. For reference, the RMSE values from LR-DIP and SLLR reconstructions are indicated by dashed lines.
Figure 3: A. Representative tissue property maps in a healthy participant using TR-DIP (bottom row) and LR-DIP (top row) reconstructions. B. Linear regression, Pearson r, and Bland-Altman analysis comparing TR-DIP and LR-DIP tissue property values in healthy participants and patients.