1357734 - Assessing Image Quality of Cardiac Magnetic Resonance (CMR) Perfusion Studies Reconstructed with a Deep Learning algorithm (AIRTM Recon DL).

Adenosine stress CMR has demonstrated good diagnostic and prognostic utility for assessing CAD¹.  To avoid motion-induced dark-rim artifacts, perfusion images need to be acquired with high spatial and temporal resolution. However, increasing spatial and temporal resolution results in a reduction in the SNR.  Deep learning (DL) reconstruction techniques have been widely adopted for image denoising, and the commercially available AIR^TM Recon DL² on the GE MR platform offers the effective way to denoise image while preserving the image quality. The software uses a deep convolutional neural network (CNN) based algorithm embedded in the MRI reconstruction pipeline, generating high fidelity images in the end. The purpose of this study was to investigate and assess the image quality of resting perfusion CMR scans using different denoising levels in AIR^TM Recon DL.



Methods:

Resting perfusion imaging was performed on a 3T Premier MR scanner (GE Healthcare, Waukesha, WI) after the injection of 0.075 mmol/kg of Gadovist in 24 patients undergoing clinically ordered CMR studies using a prototype perfusion pulse sequence.  60 images were acquired at 3 slice locations.  Typical sequence parameters included: TR=2.7ms, TE=1.1 ms, resolution 1.9 mm x 2.25 mm, temporal footprint 125 ms, ASSET R=2. Images were reconstructed using 4 denoising levels by AIR^TM Recon DL levels (DL0%, DL25%, DL35%, DL50%), where DL0% means no DL was applied, and DL35% and DL50% are the GE suggested “low” and “medium” denoising levels, respectively. Images were graded by two experienced cardiovascular imagers on a 5-point scale (1= poor image quality, 5 = excellent).



Results:

Figure 1 shows perfusion images from 1 patient reconstructed with DL0%, DL25%, DL35%, and DL50% respectively. As the DL level is increased the images have a higher apparent SNR and image quality.  The average scores for the DL0%, DL25%, DL35%, and DL50% reconstructions were 3.3 [3, 3.6], 3.9 [3.8, 4], 4 [4, 4.4], and 4.4 [4, 4.5] respectively.  DL25%, DL35%, and DL50% had improved image quality as compared to DL0% (p< 0.05, t-test) (Figure 2).  DL50% had better image quality than DL25% (p< 0.05, t-test) but not better than DL35%.  Figure 3 shows a case with a resting perfusion defect with evidence of a prior myocardial infarction. The highest image quality score for this case was for DL35% which was given an image quality score of 5.



Conclusion:

AIR^TM Recon DL significantly improved image quality for myocardial first pass perfusion imaging.  The highest image quality was obtained for AIR^TM Recon DL levels of 35% and 50%. Resting perfusion abnormalities were seen with all AIR^TM Recon DL levels.