Combined Denoising and Suppression of Transient Artifacts in Arterial Spin Labeling<scp>MRI</scp>Using Deep Learning
Patrick W. Hales, Josef Pfeuffer, Chris A. Clark
Abstract
BACKGROUND: Arterial spin labeling (ASL) is a useful tool for measuring cerebral blood flow (CBF). However, due to the low signal-to-noise ratio (SNR) of the technique, multiple repetitions are required, which results in prolonged scan times and increased susceptibility to artifacts. PURPOSE: To develop a deep-learning-based algorithm for simultaneous denoising and suppression of transient artifacts in ASL images. STUDY TYPE: Retrospective. SUBJECTS: 131 pediatric neuro-oncology patients for model training and 11 healthy adult subjects for model evaluation. FIELD STRENGTH/SEQUENCE: 3T / pseudo-continuous and pulsed ASL with 3D gradient-and-spin-echo readout. ASSESSMENT: A denoising autoencoder (DAE) model was designed with stacked encoding/decoding convolutional layers. Reference standard images were generated by averaging 10 pairwise ASL subtraction images. The model was trained to produce perfusion images of a similar quality using a single subtraction image. Performance was compared against Gaussian and non-local means (NLM) filters. Evaluation metrics included SNR, peak SNR (PSNR), and structural similarity index (SSIM) of the CBF images, compared to the reference standard. STATISTICAL TESTS: One-way analysis of variance (ANOVA) tests for group comparisons. RESULTS: The DAE model was the only model to produce a significant increase in SNR compared to the raw images (P < 0.05), providing an average SNR gain of 62%. The DAE model was also effective at suppressing transient artifacts, and was the only model to show a significant improvement in accuracy in the generated CBF images, as assessed using PSNR values (P < 0.05). In addition, using data from multiple inflow time acquisitions, the DAE images produced the best fit to the Buxton kinetic model, offering a 75% reduction in the fitting error compared to the raw images. DATA CONCLUSION: Deep-learning-based algorithms provide superior accuracy when denoising ASL images, due to their ability to simultaneously increase SNR and suppress artifactual signals in raw ASL images. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY STAGE: 1.