A deep learning algorithm for noise reduction in a novel respiratory-triggered single-shot phase sensitive inversion recovery myocardial delayed enhancement cardiac MRI pulse sequence

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Phase-sensitive inversion recovery improves tissue contrast by correcting for imperfect choice of inversion recovery time, however it is challenging to combine with a free-breathing acquisition. Deep learning (DL) algorithms have growing applications in cardiac MRI to improve image quality during image reconstruction. Purpose Here, we introduce a single-shot phase-sensitive myocardial delayed enhancement sequence with respiratory triggering (SShPSMDE-RT) (Figure 1). This novel sequence allows faster free-breathing acquisition of late gadolinium enhancement (LGE) images with reduced motion artifact (Figure 2.a). We combined this with a DL noise reduction algorithm to further improve image quality as compared to a standard segmented breath-hold (BH) PSMDE sequence. Methods 61 subjects (29 male, age 51±15) underwent cardiac MRI with the SShPSMDE-RT sequence and a standard BH sequence. The DL algorithm was applied at increasing levels (DL25, DL50, DL75, DL100) (Figure 2.b). Qualitative metrics were image quality, artifact severity, and diagnostic confidence, graded on a 5-point Likert scale. Quantitative metrics were sharpness of the left ventricle septum border and the LGE region (distance in mm for signal intensity to drop from 80% to 20%), blood-myocardium contrast-to-noise ratio (CNR), LGE-myocardium CNR, LGE signal-to-noise ratio (SNR), and LGE burden. 324 slices were included in the analysis. The sequences were compared via paired T-test. Results 27 subjects had positive LGE as determined by CMR experts. The average time to acquire a slice for SShPSMDE-RT is 4–7 seconds versus ∼30–40 seconds for the BH scan. The single-shot sequence had significantly better image quality (SShPSMDE-RT 2.1±0.8 vs. BH 1.5±0.6, p<0.001), less artifact (1.2±0.5 vs. 2.6±1.1, p<0.001), and better diagnostic confidence (3.4±0.7 vs. 2.6±0.8, p<0.001). Septum sharpness was slightly worse in SShPSMDE-RT images (4.1±1.7 mm vs. 3.8±1.6 mm, p = 0.008), but the DL algorithm improved sharpness of SShPSMDE-RT images such that there was no significant difference compared to BH images (p>0.5). There was no significant difference in LGE sharpness between the sequences. The SShPSMDE-RT images had superior blood-myocardium CNR (17.2±6.9 vs. 16.4±6.0, p = 0.040), LGE-myocardium CNR (12.1±7.2 vs. 10.4±6.6, p = 0.054), and LGE SNR (59.8±26.8 vs. 31.2±24.1, p<0.001); these metrics all improved with application of the DL algorithm. There was no significant difference in measured LGE burden (p = 0.12). Conclusions Our novel SShPSMDE-RT sequence significantly reduces scan time and motion artifact. This free-breathing sequence combined with a DL noise reduction algorithm provides better or similar image quality on both qualitative and quantitative metrics as compared to a standard BH PSMDE sequence. This technique can be used to obtain LGE imaging in patients who are unable to breath-hold or tolerate longer scan times.