2136

Application of Deep Learning Reconstruction for Denoising of Compressed Sensing non-contrast coronary MRA images to achieve improved Diagnostic Confidence.

Yoko Kato¹, Bharath Ambale-Venkatesh², Yoshimori Kassai³, John Pitts⁴, Larry Kasuboski⁴, Jason Ortman¹, Shelton Caruthers⁴, and Joao A.C. Lima¹

¹Cardiology, Johns Hopkins University, Baltimore, MD, United States, ²Radiology, Johns Hopkins University, Baltimore, MD, United States, ³Canon Medical Systems Corporation, Otawara, Japan, ⁴Canon Medical Research USA, Inc., Cleveland, OH, United States

Synopsis

Non-contrast Magnetic resonance coronary artery (MRCA) image acquisition has technical limitations of long acquisition time or reduced image resolution. We explore the use of a denoising approach with deep learning image reconstruction (dDLR) from k-space data. We investigate the effect of various levels of dDLR on Compressed Sensing non-contrast MRCA (CS-MRCA) images and optimize dDLR algorithms that achieve the best diagnostic confidence (DC) and a high signal-to-noise-ratio (SNR).

INTRODUCTION

Non-contrast MR coronary angiography (MRCA) is a candidate modality to replace coronary CT, although it is still challenging because of its time-consuming characteristics and rather low spatial resolution in comparison to CT. Improving spatial resolution is a challenge because smaller pixel sizes tend to produce lower contrast-to-noise and signal-to-noise ratios (SNR) resulting in reduced diagnostic confidence (DC). Compressed sensing (CS) is a method of accelerated image [1–4]. The optimization process of CS is based on finding the best regularization parameter lambda (λ=10^(r-4)) to balance the two terms of data fidelity and transform sparsity [5, 6]. The weakness of CS is its noise-like granular characteristics due to its random k-space data acquisition. The denoising approach with deep learning reconstruction (dDLR) is a noise adaptive algorithm. [7] The denoise level of small batch area depends on soft shrinkage model adaptively. It is promising to improve image quality of CS-MRCA, however too much image smoothing may lead to unfavorable image blurring or produce artefactual smoothing leading to removal of clinically meaningful signal intensity drops across the artery. Hence, merely assessing the noise level is insufficient. In this study, we aim to assess the feasibility of dDLR on CS-MRCA at different resolutions using a visual quality validation scale of DC and a quantitative SNR.

METHODS

Healthy participants underwent non-contrast whole heart MRI with T2-prepared segmented fast low-angle shot 3D spoiled gradient echo sequence with ECG-gating, diaphragm navigator-gating, and fat suppression. All the scans were performed with a Vantage Galan 3T scanner (Canon Medical Systems). The image acquisition parameters are shown in Table 1. For each subject, low and high resolution images were acquired with PI and CS-MRCA with different lambda (λ=10^(r-4) from r = 0.8 to 1.6). DC was defined as the confidence level of the MRCA clinical reading based on an expert, with 1 the least and 4 the most confident. After identification of the lambda level with highest DC, the corresponding CS-MRCA images were chosen to undergo the dDLR process which assumed the input (Gaussian distribution) noise levels varied from 2 to 50% of the original images.These dDLR-CS-MRCA images were scored with DC and quantitative metrics of SNR were also calculated from the signal intensity (SI) in the aorta divided by the standard deviation (SD) of the background (air and myocardium). All the analysis was performed by a single expert reader with 7 years of experience. Data were expressed as mean±SD. For the comparison of DC scores, Wilcoxon signed-rank test was performed. Statistical significance is defined as P<0.05.

RESULTS

Twelve normal participants were recruited for this IRB approved study. The mean age was 32.8 ±14.6 (y/o), 33% were male, and the mean body mass index (BMI) was 24.5±4.7. A representative case is shown in Figure 1. Low resolution images showed significantly higher DC scores than the high resolution both in PI and CS-MRCA images (Figure 2). DC increased and reached a plateau at r = 1.4 (λ=10^(1.4-4)) for the low resolution, while in high resolution r = 1.6 (λ=10^(1.6-4)) showed the highest score. Accordingly, we decided to use r = 1.6 (λ=10^(1.6-4)) CS-MRCA for running our dDLR algorithm. The DC showed highest value at dDLR 16% on low resolution (Figures 3A), and at 20% on high resolution images (Figures 3B). SNR (Air) increase rate was steeper than SNR (Myo) but both the SNRs reached plateau around dDLR 10%. (Figures 3A and B). The peak DC level in high resolution image was comparable to that of the low resolution image (Figure 3C). The summary of the study procedure is presented in Figure 4.

DISCUSSION

dDLR processing achieved the same level of DC even in high resolution images, which was not possible in PI or in CS-MRCA images without dDLR. The denoise strength depends on structure and complexity of images so that SNR, i.e. SNR of “Air” quickly improved across dDLR strength based on adaptive algorithm. Selecting optimal dDLR strength based on the resolution is likely an important consideration. The significant discordance between the DC level and noise level by SNR can be explained by the increased blurriness by the oversmoothing with high dDLR strength. When assessing dDLR processed images, it is important to assess not only the granularity, but also the blurriness. DC and SNR worked complementary so that the usage of both metrics is recommended to assess the image quality.

CONCLUSION

dDLR achieved the same level of DC in CS-MRCA with different resolution. Combining DC and SNR was recommended to assess both the noise and blurriness derived from dDLR.

Acknowledgements

This study was funded by Canon Medical Systems Corporation.

References

1. Akçakaya M, Basha TA, Chan RH, et al.: Accelerated contrast-enhanced whole-heart coronary MRI using low-dimensional-structure self-learning and thresholding. Magn Reson Med 2012; 67:1434–1443.

2. Akçakaya M, Basha TA, Chan RH, Manning WJ, Nezafat R: Accelerated isotropic sub-millimeter whole-heart coronary MRI: compressed sensing versus parallel imaging. Magn Reson Med 2014; 71:815–22.

3. Nam S, Akçakaya M, Basha T, et al.: Compressed sensing reconstruction for whole-heart imaging with 3D radial trajectories: A graphics processing unit implementation. Magn Reson Med 2013; 69:91–102.

4. Nakamura M, Kido T, Kido T, et al.: Non-contrast compressed sensing whole-heart coronary magnetic resonance angiography at 3T: A comparison with conventional imaging. Eur J Radiol 2018; 104(February):43–48.

5. Otazo R, Kim D, Axel L, Sodickson DK: Combination of compressed sensing and parallel imaging for highly accelerated first-pass cardiac perfusion MRI. Magn Reson Med 2010; 64:767–776.

6. Forman C, Wetzl J, Hayes C, Schmidt M: Compressed Sensing : a Paradigm Shift in MRI. MAGNETOM Flash 2016; 66:8–13.

7. Isogawa K, Ida T, Shiodera T, Takeguchi T: Deep Shrinkage Convolutional Neural Network for Adaptive Noise Reduction. IEEE Signal Process Lett 2018; 25:224–228.

Figures

Table 1. Image parameters for the Parallel Imaging and Compressed Sensing MRCA acquisition in high and low resolution images. MRCA, magnetic resonance coronary angiography; TR, repetition time; TE, echo time; FOV, field of view; PI, parallel imaging; CS, compressed sensing; SPAIR, spectral attenuated inversion recovery.

Figure 1. Representative case of quantitative analysis on low and high resolution images scanned with PI, CS λ=10^(1.6-4), and CS λ=10^(1.6-4) with DLR 16% or 20%. ROIs are placed on the aorta, myocardium, and the air. PI, parallel imaging; CS, compressed sensing; DLR, deep learning reconstruction.

Figure 2. Normalized Diagnostic Confidence on Parallel imaging and Compressed Sensing with different r value change for the lambda (λ=10^(r-4)) on low resolution and high resolution images. Low resolution images (blue) showed significantly higher DC scores than the high resolution (red) both in PI and CS-MRCA images. DC, diagnostic confidence; PI, parallel imaging; CS, compressed sensing; MRCA, magnetic resonance coronary angiography.

Figure 3. Diagnostic Confidence, SNR(Myo), and SNR(Air) in different dDLR strengh. Panels(A)and(B): Normalized data on low(A) and high resolution(B) images. Panel(C): Combined original data on low and high resolution images. DC are multiplied with 10 to fit the Y-axis. The SNRs reached plateau around dDLR strength of 10, with more obvious change of the curve on low resolution images. The peak DC level in high resolution was comparable to that of the low resolution. In higher dDLR strength significant discordance between the SNR and the DC level were observed. DLR, deep learning reconstruction; DC, diagnostic confidence; SNR, signal-to-noise ratio.

Fig.4 Summary of the Study procedures. CS, compressed sensing; PI, parallel imaging; MRCA, magnetic resonance coronary angiography; DC, diagnostic confidence; DLR, deep learning reconstruction; SNR, signal-to-noise ratio.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)

2136