Improving Spatial Resolution in Phase Sensitive LGE with Variable-Density Undersampling and Deep Learning Reconstruction

Background: Cardiac magnetic resonance (CMR) remains the gold standard for comprehensive evaluation of hypertrophic cardiomyopathy (HCM), offering precise assessment of left ventricular (LV) hypertrophy, myocardial fibrosis through late gadolinium enhancement (LGE), and dynamic left ventricular outflow tract (LVOT) obstruction^1-3.

Late gadolinium enhancement is a crucial part of routine CMR imaging in patients for evaluating the myocardium post contrast injection. The main approach currently used in the clinic is a phase sensitive LGE (PS LGE) sequence that is more robust to the selection of the inversion time from patient to patient.

While the spatiotemporal resolution of current single-shot (ssh) PS LGE sequences is adequate for routine clinical use, improvement in this area would be very beneficial for patient care. Recent developments in 2D accelerated techniques can help overcome current resolution limitations^4-6. The purpose of this study was to assess a highly undersampled high in-plane spatial resolution ssh PS LGE to the current clinical standard.

Methods: 57 subjects with a referring diagnosis of cardiomyopathy were recruited. All images were acquired on a 1.5T SIGNA Artist (GE HealthCare, Waukesha, WI) as part of a routine protocol. The patients underwent 2 cardiac gated ssh PS LGE acquisitions: 1 with a standard acquisition using 2-fold ARC acceleration; and 1 using a fast, highly undersampled (4x) acquisition, which utilizes variable density undersampling and online DL-based (Sonic DL) reconstruction^4-6. Both acquisitions were acquired temporally adjacent to each other. The standard acquisition had ~3x3x8 mm³ voxels, while the Sonic DL version had ~2x2x8 mm³. The two acquisitions had similar temporal resolutions and overall scan times.

Images were then assessed by one cardiologist and one diagnostic radiologist using the suiteHEART software package. A Likert scale (1-5) was used to assess the image quality of each acquisition. Presence of artifact (e.g. banding, blurring, motion) and acquisition preference was also recorded. Wilcoxson ranked signed test weas used for statistical analysis of the qualitative assessments.

Results: There was no statistically significant difference between the two acquisitions in terms of image quality, with average Likert scores of ~4.1-4.2 from each reader for each acquisition. Artifact presence was also similar between the two acquisitions. The highly accelerated version was relatively more preferred by both readers (Reader 1: 18 vs. 12 with 27 no preference; Reader 2: 23 vs. 11 with 23 no preference). Figures 1 and 2 shows example images for patients with and without any visible abnormalities, respectively.

Conclusion: This study showed the value in utilizing a high undersampled acquisition and Sonic DL reconstruction for routine PS LGE acquisitions. There was a ~2-fold improvement in spatial resolution with preserved diagnostic quality, which can directly translate to improved patient care and comfort.

Figure 1: Comparison of Sonic DL and standard PS LGE acquisitions in a patient with an abnormality. The higher spatial resolution of the SONIC DL acquisition helps improve visualization. A representative slice is shown for each acquisition from the same patient.
Figure 2: Comparison of Sonic DL and standard PS LGE acquisitions in a patient without an abnormality. A representative slice is shown for each acquisition from the same patient.