Simultaneous T1, T2, and T1ρ mapping of the myocardium using cardiac MR fingerprinting with a deep image prior reconstruction

Abstract

Purpose: To develop a cardiac MR fingerprinting (cMRF) approach using deep image prior reconstruction (DIP) to simultaneously map T₁, T₂, and T_1ρ, and assess T_1ρ in healthy subjects and patients with areas of enhancement on late gadolinium enhancement.

Methods: A 2D electrocardiogram-triggered cMRF sequence was developed to measure T₁, T₂, and T_1ρ simultaneously. DIP reconstruction was evaluated for noise and artifact reduction compared to a low-rank reconstruction. Measurements were assessed in simulation and phantom for accuracy. T_1ρ-cMRF maps were generated in 10 healthy subjects and six patients under evaluation for cardiomyopathy, myocarditis, and amyloidosis receiving gadolinium-based contrast agent-enhanced CMR at 1.5 T.

Results: T_1ρ-cMRF maps showed excellent agreement with ground truth (RMS error = 3.0% ± 5.3%) and conventional methods (R² = 0.99) in simulations and phantom experiments. Measured values in healthy subjects were consistent with literature (T₁ = 1051 ± 63 ms, T₂ = 41.4 ± 3.3 ms, and T_1ρ = 45.5 ± 2.4 ms). DIP reconstruction reduced noise, indicated by lower coefficient of variation (Δ = 4.7), compared to low-rank reconstruction. Mean differences of 10.2 ms (p = 0.02) in T_1ρ and 6.9 ms in T₂ maps were observed between areas of enhancement on late gadolinium enhancement and normal-appearing myocardium in patients. Within individual patients, significant differences (p < 0.01) in T₁, T₂, and T_1ρ were observed between American Heart Association segments with and without contrast enhancement.

Conclusion: The proposed T_1ρ-cMRF sequence using DIP reconstruction enables simultaneous quantification of T₁, T₂, and T_1ρ with decreased coefficient of variation compared to low-rank reconstruction. Simulation and phantom studies show good agreement with references. In vivo measurements were made in healthy subjects, and areas of contrast enhancement in patients showed elevated T₂ and T_1ρ relative to remote myocardium.

Keywords: MR fingerprinting; T1ρ; cardiac MRI; multiparametric mapping.

FIGURE 1

The proposed T_1ρ‐cMRF sequence and flip angle pattern. Preparations consist of an inversion pulse, no preparation, T_1ρ spin lock pulse (TSL = 30, 50, or 60 ms), or T₂ preparation pulses (duration 30, 50, or 80 ms), followed by a 250 ms FISP readout in which 47 spiral arms are collected. cMRF, cardiac MR fingerprinting; FISP, fast imaging with steady state precession; TSL, spin lock time.

FIGURE 2

DIP reconstruction framework. The IRN outputs a set of K subspace images. These images are used to simulate the forward encoding model, including multiplication by coil sensitivity maps, NUFFT, and projection from the SVD subspace to the time domain to obtain an estimated k‐space. Training is completed in a self‐supervised manner by computing the MSE loss of the k‐space estimate with the acquired data. Concurrently, the spatial basis images are input into the PEN, which outputs the estimated T₁, T₂, T_1ρ, and M₀ parameter maps. These maps are used to simulate the signal evolution by the pretrained FGN, which are subsequently compressed to the low‐rank synthetic spatial basis. Training is again self‐supervised by computing the MSE loss between the synthetic and input subspace images. DIP, deep image prior; FGN, fingerprint generator network; IRN, image reconstruction network; MSE, mean squared error; NUFFT, nonuniform fast Fourier transform; PEN, parameter estimation network; SVD, singular value decomposition.

FIGURE 3

XCAT phantom experiments simulating (A) a healthy heart and (B) a heart with a region of fibrosis. The white arrow indicates the simulated area of fibrosis. Ground truth (left) tissue property maps are compared to those generated using the proposed T_1ρ‐cMRF scheme reconstructed with a low‐rank (middle) and DIP (right) approach. RMSE values for healthy (H) and fibrotic (F) myocardium are reported RMSE, RMS error.

FIGURE 4

(A) Correlation plots of the T₁ (top), T₂ (middle), and T_1ρ (bottom) values measured using conventional methods (green) and T_1ρ‐cMRF reconstructed with a low‐rank (blue) and DIP (red) approach compared to the NIST reference values for T₁ and T₂, and reference T_1ρ FLASH scan for T_1ρ. The gray line indicates unity and error bars show the SD over all pixels in each sphere. (B) Bland–Altman plots comparing conventional (left), cMRF with low‐rank (middle) and DIP reconstructions (right) with reference values for T₁ (top), T₂ (middle), and T_1ρ (bottom). On each plot, bias is indicated by the solid line, and dashed lines indicate the 95% limits of agreement NIST, National Institute of Standards and Technology.

FIGURE 5

(A) T₁, T₂, and T_1ρ maps from a single representative healthy subject measured using T_1ρ‐cMRF reconstructed with a low‐rank (left) and DIP (right) approach. (B) Bullseye plots showing the mean and (C) COV for the T₁, T₂, and T_1ρ values measured in the six AHA segments across all 10 healthy volunteers AHA, American Heart Association; COV, coefficient of variation.

FIGURE 6

T₁, T₂, and T_1ρ maps measured using the proposed T_1ρ‐cMRF sequence along with the DIP reconstruction in 10 healthy subjects.

FIGURE 7

(A) T₁, T₂, and T_1ρ maps from a single representative patient measured using T_1ρ‐cMRF reconstructed with a low‐rank (left) and DIP (right) approach. (B) Bullseye plots for the mean and (C) COV of the T₁, T₂, and T_1ρ values measured in the six AHA segments across all six patients.

FIGURE 8

LGE images and native tissue property maps measured using the proposed T_1ρ‐cMRF sequence in six patients. The white and black outlines indicate ROIs drawn in areas of scar and normal‐appearing myocardium, respectively, based on the LGE images. Patients scans referred for cardiomyopathy (purple), amyloidosis (yellow), and myocarditis (pink) are indicated by outline color LGE, late gadolinium enhancement; ROI, region of interest.

FIGURE 9

T₁ (left), T₂ (center), and T_1ρ (right) values measured using T_1ρ‐cMRF with a DIP reconstruction in the healthy‐appearing myocardium of both healthy subjects and patients, as well as in areas of scar in the patients. Significant differences are indicated as * (p < 0.05); ** (p < 0.01).

FIGURE 10

LGE images and bullseye plots for mean T₁, T₂, and T_1ρ values measured in the six AHA segments for each patient. Segments highlighted in red indicate regions of scar identified by an advanced imaging cardiologist. Patients scans referred for cardiomyopathy (purple), amyloidosis (yellow), and myocarditis (pink) are indicated by outline color.