1357558 - CAMWARE: Cascaded Multi-level Wavelet Refinement Neural Network for Accelerated Whole-brain MR Vessel Wall Imaging

Background: 3D MR vessel wall imaging (VWI) is a non-invasive imaging method that can directly probe intracranial arterial wall diseases. However, the needs for large spatial coverage and submillimeter spatial resolution mandate a long acquisition time. Recently, convolutional neural networks^1,2 (CNN) have shown promising results on imaging acceleration in a wide variety of MR applications. In this work, we propose a novel intracranial vessel-dedicated CAscaded Multi-level WAvelet REfine (CAMWARE) network that enables a whole-brain VWI scan within 4 min.

Methods: Network design: CAMWARE adopts a combination of U-Net and multi-channel Data Consistency (DC) Layer1(Figure 1). To enhance the network performance in showing the vessel wall structures, we modify U-Net in two aspects. (a) Pooling and deconvolution operations, which are adopted in conventional U-Net to enlarge receptive field and alter resolution of feature maps, are replaced by discrete wavelet transform and inverse wavelet transform, respectively. (b) Multi-path refinement3 building block is added to fuse coarse features with fine-grained low-level features to restore sharp vessel wall boundaries. We concatenate four “CNN + DC” blocks, such that each additional block acts as an extra step to reduce the learning errors of the former block.

Data preparation: A total of 77 whole-brain VWI datasets, including 47 pre-contrast and 30 post-contrast scans, were acquired on a 3T system (Skyra, Siemens) with a standard 20-channel head-neck coil. The imaging protocol4 included: FOV=248x229x132 mm3, matrix size=448x414x240, 0.55mm isotropic resolution, GRAPPA acceleration of 2x in the phase-encoding direction, 12.1 min acquisition time. Coil-combined complex-valued images reconstructed directly on the scanner were saved as the ground-truth, and thus was considered as a single-coil scenario. A variable density Poisson-disc undersampling mask was retrospectively applied to the k-space data generated by Fourier transforming the complex-valued images to simulate undersampled k-space which corresponded to an acceleration factor 6.5 or a 4-min acquisition. Coil sensitivity masks were estimated using ESPIRIT.

Experiments: We split the 77 VWI sets into 57 for training, 6 for validation and 14 for testing, respectively. CAMWARE was compared to zero-filled reconstruction and compressed sensing (CS) based on L1-SPIRIT reconstruction. All neural networks were trained with 2D transverse slices using a combination of structural similarity index (SSIM) and mean absolute error (MAE) losses with Adam optimizer. The performances were evaluated with SSIM and peak signal-to-noise ratio (PSNR).

Results: The qualitative comparisons and the corresponding error maps for different imaging acceleration methods are shown in Figure 2. Figure 3 shows vessel wall images from two ischemic stroke patients.

Conclusion: In this work, we proposed a novel CAMWARE MR deep-learning reconstruction network that holds the potential to enable 4-min whole-brain VWI.