Deep Learning Reconstruction Improves Image Quality While Preserving T2 Quantification in Cardiac MRI

Background: T2 mapping in cardiac MRI is a key tool for assessing myocardial edema and inflammation (1), but its diagnostic accuracy depends on high-quality, reproducible images. Deep Learning Reconstruction (DLR) algorithms, enhance spatial resolution and signal-to-noise ratio without changing acquisition parameters (2). However, their impact on quantitative cardiac imaging remains insufficiently validated. This study evaluates whether applying DLR alters native myocardial T2 quantification and assesses its potential to improve image quality and precision without affecting diagnostic thresholds.

Methods: Fifty-four patients (56±18 years, 36 males) undergoing clinical cardiac MRI at 1.5T (GE Signa Artist) were prospectively enrolled. Native short-axis T2 mapping was acquired using breath-held multi-echo fast spin echo sequences at basal, mid, and apical levels. Each T2 map was reconstructed using both standard reconstruction (SR) and DLR (AIR™ Recon DL, GEHC), with identical acquisition parameters. A T1MES phantom with known T2 values was scanned for in vitro validation. T2 values were analyzed using CVI42 software, applying identical AHA-segmented ROIs across both reconstructions. Image quality was assessed by an expert reader blinded to reconstruction type, using a 5-point Likert scale evaluating sharpness, homogeneity, and diagnostic quality. Agreement between SR and DLR was evaluated using ICC, Bland-Altman analysis, and coefficient of variation (CoV). Differences in quality scores and T2 precision were assessed with the Wilcoxon signed-rank test.

Results: In-vitro, SR and DLR yielded identical T2 values in clinically relevant phantom tubes (R²=0.94, ICC=0.97 [0.79–1.00], CoV ≤0.82%). In vivo, DLR significantly improved image quality: 94% of cases were rated equal or better than SR, with 69% showing higher scores (p < 0.0001). Mean myocardial T2 values were nearly identical (SR: 57.8±4.8 ms, DLR: 57.9±4.8 ms; ICC=1.00 [1.00–1.00], CoV=0.22%). Segmental agreement was also excellent (ICC=1.00 [1.00–1.00], CoV=0.46%). DLR reduced regional T2 standard deviation by 11.8% on average (mean difference: –0.56±1.24 ms; p< 0.0001), suggesting improved precision without bias. An illustration of this improvement is shown in Figure 1.

Conclusion: DLR significantly enhances image quality and improves the precision of myocardial T2 maps without altering mean T2 values. This algorithm preserves quantitative accuracy, supporting its safe integration into clinical cardiac MRI workflows without redefining reference ranges. Further studies are warranted to assess performance across vendors and field strengths.

Figure 1. Native myocardial T2 mapping using standard reconstruction (SR, panels A–C) and deep learning reconstruction (DLR, panels D–F). A, D: Representative mid-ventricular T2 maps. B, E: Bullseye maps of segmental mean T2 values. C, F: Bullseye maps of T2 standard deviation, showing reduced variability with DLR.