Twenty-Fold Accelerated Cardiovascular Phase-Contrast MRI using Channel-Shift CNN (RF_TH_274)

Thursday, January 30, 2025

6:00 PM – 6:10 PM East Coast USA Time

Location: Blue Room PreFunction

Presenting Author(s)

CZ

Chi Zhang, PhD

Postdoc
Stanford University

Primary Author(s)

CZ

Chi Zhang, PhD

Postdoc
Stanford University

Co-Author(s)

JO

Julio A. Oscanoa, MSc

Stanford University
ML

Michael Loecher, PhD

Research Scientist
Stanford University
MM

Matthew Middione, PhD

Research Scientist
Stanford University
Shreyas S. Vasanawala, MD, PhD

Professor
Stanford University
Daniel B. Ennis, PhD

Professor
Stanford University

Background: Phase contrast (PC) MRI is the clinical standard for cardiovascular blood velocity and flow quantification. Although recent advances enabled deep-learning 8-fold accelerated cardiovascular PC-MRI [1], breath holding through multiple heartbeats is still required during the scan. Further acceleration to reduce the scan time to a single heartbeat is important for patients unable to perform breath holding. In this study we propose a channel-shift convolutional neural network (CS-CNN) for deep learning reconstruction of 20-fold accelerated cardiovascular PC-MRI.

Methods: CS-CNN: CS-CNN aims to enlarge the receptive field, which is proven to be beneficial in various computer vision tasks [2], including MRI reconstruction [3]. CS-CNN uses replicas of the input image via circular shifting along the under-sampled direction. These replicas are concatenated with the input image along its channel dimension. This augmented input is subsequently fed to a conventional CNN, whose receptive field is determined by the step size of the circular shifting and the number of shifted replicas. As such, CS-CNN achieves an enlarged receptive field via a fully convolutional architecture. This allows arbitrary input matrix size and can also be efficiently executed on modern hardware owing to a reduced memory footprint (compared to infeasible fully convolutional networks) and fast processing speed.
In-Vivo Experiment: 158 fully sampled single slice 2D PC-MRI data were acquired on a 3T MRI (GE Healthcare) using a commercial ECG-gated spoiled gradient echo sequence: TE = 3.0 ± 0.2ms, TR = 5.3 ± 0.3ms, temporal resolution = 83.4 ± 6.1ms, flip angle = 22 ± 3.9°, slice thickness = 9.1 ± 1.6mm. Subjects were consented under IRB approval. 141 slices were used for training and 17 for testing. Retrospective k-space undersampling at 8 and 20-fold was generated using a variable density pattern. The proposed CS-CNN was used as the regularizer in an unrolled network [4] for reconstruction.

Results: Fig. 2 shows representative results using the method in [1] (ResNet) at 8 and 20-fold acceleration and using our CS-CNN for 20-fold acceleration. At a relatively low 8-fold acceleration, the ResNet baseline produces sharp images and accurate total flow. At 20-fold acceleration, the ResNet baseline exhibits blurrier images and increased flow errors, while the CS-CNN shows improved image quality (less artifacts and sharper details), as well as more accurate total flow. Table 1 shows the statistics of image quality metrics (SSIM and NMSE) and flow quantifications (NMSE in %VENC), with respect to fully sampled acquisition. Reconstruction using the CS-CNN outperforms its ResNet counterpart at 20-fold for reduced errors.

Conclusion: The proposed CS-CNN provides improved image quality and flow metrics at 20-fold acceleration. A future study will validate the method via prospective acquisition on patients.

Grant support: R01 HL171515, U01 EB029427, and GE Healthcare A-129