Denoising using deep-learning-based reconstruction for whole-heart coronary MRA with sub-millimeter isotropic resolution at 3 T: a volunteer study

PURPOSEThe aim of this study was to assess the usefulness of denoising deep-learning-based reconstruction (dDLR) to improve image quality and vessel delineation in noncontrast 3-T whole-heart coronary magnetic resonance angiography (WHCMRA) with sub-millimeter isotropic resolution (Sub-mm) compared with a standard resolution without dDLR (Standard).METHODSFor 10 healthy volunteers, we acquired the WHCMRA with Sub-mm with and without dDLR and Standard to quantify signal- (SNR) and contrast-to-noise ratio (CNR) and vessel edge signal response (VESR) in all the 3 image types. Two independent readers subjectively graded vessel sharpness and signal homogeneity of 8 coronary segments in each patient. We used Kruskal–Wallis test with Bonferroni correction to compare SNR, CNR, VESR, and the subjective evaluation scores among the 3 image types and weighted kappa test to evaluate inter-reader agreement on the scores.RESULTSSNR was significantly higher with Sub-mm with dDLR (P < .001) and Standard (P = .005) than with Sub-mm without dDLR and was comparable between Sub-mm with dDLR and Standard (P = .511). CNR was significantly higher with Sub-mm with dDLR (P < .001) and Standard (P = .005) than with Sub-mm without dDLR and was comparable between Sub-mm with dDLR and Standard (P = .560). VESR was significantly greater with Sub-mm with (P = .001) and without dDLR (P = .017) than with Standard and was comparable between Sub-mm with and without dDLR (P = 1.000). In the proximal, middle, distal, and all the coronary segments, the subjective vessel sharpness was significantly better with Sub-mm with dDLR than Sub-mm without dDLR and Standard (P < .001, for all) and was comparable between Sub-mm without dDLR and Standard (P > .05); the subjective signal homogeneity was significantly improved from Sub-mm without dDLR to Standard to Sub-mm with dDLR (P < .001). The inter-reader agreement was excellent (kappa = 0.84).CONCLUSIONApplication of dDLR is useful for improving image quality and vessel delineation in the WHCMRA with Sub-mm compared with Standard.

Main points

• A denoising method with deep-learning-based reconstruction (dDLR) uses a deep convolution neural network to reduce image noise in magnetic resonance images without additional scan time.
• dDLR can be applied to improve signal- (SNR) and contrast-noise ratio in noncontrast 3-T whole-heart coronary magnetic resonance angiography (WHCMRA) using a spoiled gradient-echo sequence, which offers lower SNR than a steady-state free precession sequence commonly applied at 1.5 T.
• Combined application of dDLR and sub-millimeter isotropic resolution is useful for improving vessel sharpness, signal homogeneity, and delineation of the coronary arteries in the WHCMRA.

Three-dimensional (3D) whole-heart coronary magnetic resonance angiography (WHCMRA) is a noninvasive imaging method for assessing coronary artery stenosis and is advantageous over coronary computed tomography angiography (CCTA) because it does not require radiation exposure or contrast media administration and is only slightly susceptible to calcium-related artifacts.^1-3 Because WHCMRA is commonly limited in delineation of distal coronary segments and quantification of vessel lumen stenosis due to its insufficient spatial resolution and anisotropy, some investigators have used various techniques to acquire WHCMRA with sub-millimeter isotropic resolution (Sub-mm) whose image quality is not necessarily satisfactory at 1.5 T, within an acceptable acquisition time.^4-6 While a steady-state free precession (SSFP) sequence has been commonly applied in noncontrast WHCMRA at 1.5 T, increased B1 field inhomogeneity, frequency offset from tissue susceptibility variation, and specific absorption rate limit consistency of SSFP at 3 T. As such, the spoiled gradient-echo sequence, which decreases the signal-to-noise ratio (SNR) of the coronary arteries, has often been used at 3 T.^7-10A 3-T clinical MR scanner with a maximum gradient magnetic field of 100 mT/m (the slew rate: 200 mT/m/ms) has been recently introduced and can offer thinner slice images at the same bandwidth (i.e., the same sampling interval/time). With this scanner, a denoising method with deep-learning-based reconstruction (dDLR) has been newly developed using a convolution neural network (CNN) to improve SNR in high-resolution MR images without additional scan time.¹¹ Currently, this dDLR algorithm (Advanced intelligent Clear-IQ Engine [AiCE], Canon Medical Systems Corporation) is clinically available only from the single vendor. While some smoothing filters frequently applied in clinical settings are designed to reduce high-frequency noise at the cost of image blurring, we hypothesized that dDLR should only reduce image noise without any negative influence on the delineation of the coronary vessels in WHCMRA. Thus, we performed volunteer studies to assess usefulness of dDLR to improve image quality and coronary vessel delineation in noncontrast WHCMRA with Sub-mm compared with a standard resolution without dDLR (Standard) using this 3-T scanner.