Magnetization Transfer Magnetic Resonance Fingerprinting for Characterization of Myocardial Scar

Background: While quantitative mapping has proven valuable for differentiating between various myocardial diseases^1–4, existing techniques, including MOLLI and cardiac Magnetic Resonance Fingerprinting (cMRF), ignore the effects of Magnetization Transfer (MT) to simplify signal modeling. However, recent studies have suggested that neglecting MT may reduce the accuracy of tissue property measurements^5–7, whereas leveraging MT effects may enable scar identification without the need for gadolinium-based contrast agents (GBCA)^8–11. This study investigates how incorporating MT into the signal model influences the ability of measurements made with cMRF to distinguish scar from normal myocardium.

Methods: cMRF^12–14 data were acquired in 3-9 short-axis slices in 4 cardiac patients receiving GBCA and 1 renal patient undergoing a contrast-free scan at 1.5T (MAGNETOM Sola, Siemens Healthineers). LGE images were acquired in the patients who received GBCA ~10min after contrast injection, and a T_1ρ map was acquired in the renal patient. T₁ and T₂ maps were reconstructed from the cMRF data using a deep image prior approach,^15,16 with tissue property values estimated using either a conventional EPG model (ignoring MT effects) or a signal model including MT effects. For the MT model, exchange between free- and bound-water pools was simulated using EPG-X¹⁷, assuming a fixed rate of MT between pools (k_f=4.2Hz)¹⁰, T_1,f=T_1,b, and T_2,b=12μs. Following cMRF reconstruction, T₁ and T₂, as well as bound pool fraction (f) values (when MT is included), were measured by computing the mean and standard deviation within ROIs manually drawn within the appropriate American Heart Association (AHA)¹⁸ segments in the slice that contained the largest area of LGE enhancement. Paired t-tests between values measured with and without modeling MT effects were performed. T₁ and T₂ values from additional ROIs drawn in regions of scar (defined by LGE where available or in regions of suspected scar defined by elevated T₁ and T_1ρ values) and normal-appearing myocardium were used in a linear discriminant analysis (LDA) classification. Model performance was evaluated using a leave-one-out cross-validation approach. LDA was conducted both with and without MT modeling and accuracy of the classification was evaluated.

Results: T₁, T₂, and bound pool fraction (f) maps obtained from each patient reconstructed both ignoring and including MT effects in the signal model, as well as the corresponding LGE images, are shown in Fig. 1. Fig. 2 shows average T₁, T₂, and f values measured in each of the AHA segments for both types of reconstructions. Elevations in both T₁ and T₂ (p< 0.001) were observed when modeling MT. Fig. 3 shows results of the LDAs. The models yield an average out-of-sample predictive accuracy of 61.8% when ignoring MT effects and 82.4% when including MT effects.

Conclusion: Including MT effects in the cMRF signal model improves the ability of native tissue property measurements to differentiate scar and normal appearing tissue.

Figure 1. T1 (top), T2 (middle), and bound pool fraction (bottom-left) maps for each patient derived from the cMRF data reconstructed using a signal model that includes MT effects (left) and ignores MT effects (right). The corresponding LGE images are shown at the bottom-right.
Figure 2. Bullseye plots for mean (top) T1, (middle) T2, and (bottom) bound pool fraction values measured in the AHA segments for each patient cMRF map reconstructed without (left) and with (right) including MT effects. Note the higher measured values of T1 when MT is included in the model.
Figure 3. Leave-one-out linear discriminant analysis (LDA) using T1 and T2 values from ROIs in regions of scar (blue) and normal appearing myocardium (red) both ignoring (A) and including (B) MT effects in the signal model. Accuracy for testing data prediction is reported in each plot. The different shadings of each color indicate values from different patients.