SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Hung Do, PhD
Mgr Clinical Collaborations
Canon Medical Systems USA, Inc.
Andrew Yoon, MD
Cardiologist
Long Beach Memorial Medical Center, University of California, Irvine
Krishna Nayak, PhD
Professor
University of Southern California
Background:
Signal-to-noise ratio (SNR) in MRI is proportional to the square root of acquisition time. Therefore, denoising algorithms could potentially allow for shorter image acquisition while keeping image quality comparable to that acquired from longer acquisitions. UNET [1] and the deep residual learning network (dnCNN) [2] architectures are widely used for image generation and denoising tasks. This work proposes a hybrid network, named (dnoiseNET), and demonstrates its improved performance compared to UNET and dnCNN architectures.
Methods:
Gaussian noise was added to raw arterial spin labeled (ASL) images to simulate noisy input within the peak SNR (PSNR) range of 18.28±8.18. A total of 478 ASL images (control and label) from 22 subjects were randomly divided into training and validation sets of 438 and 40 images, respectively. A total of 144 ASL images acquired at rest and during Adenosine stress from 6 “un-seen” heart transplant recipients were reserved for testing [3].
Network architectures are shown in Figure 1. Training was performed on a 12Gb-memory NVIDIA K80 GPU. Common training details include batch size = 12, learning rate = 2e-4 with learning rate decay, 150 epochs with early stopping, loss function = mean absolute error, and Adam optimizer. Additionally, nonlocal mean (NLM) denoiser [4] was also used for comparison. Inference time was 20-30ms/image for DL denoisers and 2.6s/image for NLM.
Mean squared error (MSE), PSNR, structural similarity index (SSIM) [5] and, more importantly, end-point quantitative measures including myocardial blood flow (MBF) and precision (a.k.a. physiological noise) were used as quality metrics for evaluating the performance of the denoising algorithms.
Results:
Representative denoised images are shown in Fig1D (control image) and Fig1E (label image). All denoisers demonstrate superior MSE, PSNR, and SSIM compared to noisy data as seen in Fig2A (p<0.001). Among all the denoisers, dnoiseNET has superior MSE, PNSR, and SSIM compared to that from the other denoisers as seen in Fig2A (p<0.001). Figure 3 shows UNET and dnCNN introduces significant bias to measured MBF, while bias in dnoiseNET is negligible.
Conclusion:
dnoiseNET provides superior performance compared to NLM, UNET, and dnCNN in term of MSE, PSNR, and SSIM. More importantly, dnoiseNET introduces negligible bias to the end-point quantitative MBF measurements.