Impact of AI Denoising on Qualitative and Quantitative Metrics in Accelerated Cardiac CINE MR

Background: Cine scans are a significant component of CMR exams but often require ten or more breath-holds. In this study, we evaluated an AI-denoising technique and progressively increasing compressed sensing factors for their effects on image quality metrics, and morpho-functional assessment, including circumferential strain. We hypothesized that AI reconstruction would enhance image quality while preserving quantitative metrics relative to non-AI reconstruction.

Methods: A total of n=361 short-axis cine acquisitions were acquired in n=53 healthy volunteers (Sex=36F; Weight=79±17 kg; Age=39 ±12 years) and were reconstructed with both a standard non-AI Compressed Sense (CS) reconstruction and a commercially available Artificial Intelligence (AI) denoising reconstruction, resulting in n=722 reconstructed short-axis datasets. Images were acquired on a 3T scanner (Philips Healthcare, Best, The Netherlands) with a 28-channel torso coil. Short-axis cine images were acquired with a range of CS acceleration factors (CS factors 2, 3, 4, 5, 6, 8, 10, and 12).

Cine scans were 2D balanced turbo-field-echo acquisitions with echo train length=8, TR=2.8 TE= 1.38-1.42 ms, 40 cardiac phases, and voxel size 2.1 x 1.7-2.4 x 8 mm. The number of heartbeats for these acquisitions is 13 for CS2, 8 for CS3, 6 for CS4, 5 for CS5, 4 for CS6, 3 for CS8 and CS10, and 2 for CS12.

Three reviewers, blinded to CS acceleration, graded all 722 reconstructed short-axis cine stacks and recorded IQ metrics. IQ metrics were overall IQ, blurring, perceived signal-to-noise ratio (pSNR), aliasing, and susceptibility artifacts. All metrics were scored as 1=excellent, 2=good, 3=fair, 4=non-diagnostic. In all reconstructed images, quantitative metrics were calculated by manual contouring of all 722 reconstructed images. Left- and right-ventricular end diastolic volume (EDV), end systolic volume (ESV), stroke volume (SV), left-ventricular ejection fraction (LVEF), and right-ventricular ejection fraction (RVEF), as well as global circumferential strain were recorded.

Results: Representative images are shown in Figure 1. Myocardial wall boundary and papillary muscles were noted to be generally better defined when comparing AI reconstruction to non-AI methods.

For some image quality parameters (Figure 2), statistically significant improvements to image quality metrics with AI denoising reconstruction compared to standard wavelet reconstruction were detected. This includes overall image quality (CS10 p=0.006), blurring (CS8 p< 0.001, CS10 p< 0.001, CS12 p< 0.001), and pSNR (CS10, p< 0.001). Regarding quantitative parameters, no significant differences were observed between AI and non-AI reconstructed images for any quantitative parameters, LVEF, RVEF, and GCS, for any CS factors 2-12 (Figure 3; p< 0.581).

Conclusion: Quantitative parameters were preserved for AI reconstruction compared to standard non-AI wavelet reconstruction, and image quality metrics, especially image blurring, were found to improve with AI reconstruction.

Figure 1. Representative images from a 32-year-old healthy female. For each compressed sensing factor, images were acquired separately, and non-AI and AI images are reconstructed from the same data set. As acceleration factor increases, details are lost in these images. However, some detail in the papillary muscle and myocardial wall definition is improved by AI denoising, some examples for CS factor 5 and 6 images are demarcated by blue circles. (CS compressed sensing, AI artificial intelligence)
Figure 2. Averaged gradings from four blinded readers for five different image quality metrics in categories overall image quality, blurring, perceived SNR, susceptibility artifacts, and aliasing. Higher acceleration factors were included as this study progressed, resulting in different n for different acceleration factors. Sample sizes indicate data included for both standard and AI reconstructions. Image quality metrics are improved for AI reconstruction, especially for higher compressed sense factors and the blurring metric. (SNR signal-to-noise ratio, AI artificial intelligence)
Figure 3. Quantitative metrics from cine images included left ventricular ejection fraction, right ventricular ejection fraction, and global circumferential strain. The AI-based reconstruction technique did not provide significantly different results from the non-AI based technique for any of the groups (p>0.581). Sample sizes indicate data points included for both wavelet denoising and AI denoising. (AI artificial intelligence)