Scan-Adaptive Dynamic MRI Undersampling Using a Dictionary of Efficiently Learned Patterns

TL;DR

The learned sampling approach improves reconstruction quality across multiple acceleration factors on public and in-house cardiac MRI datasets, including PSNR gains of 2-3 dB, reduced NMSE, improved SSIM, and higher radiologist ratings.

Abstract

Cardiac MRI is limited by long acquisition times, which can lead to patient discomfort and motion artifacts. We aim to accelerate Cartesian dynamic cardiac MRI by learning efficient, scan-adaptive undersampling patterns that preserve diagnostic image quality. We develop a learning-based framework for designing scan- or slice-adaptive Cartesian undersampling masks tailored to dynamic cardiac MRI. Undersampling patterns are optimized using fully sampled training dynamic time-series data. At inference time, a nearest-neighbor search in low-frequency $k$-space selects an optimized mask from a dictionary of learned patterns. Our learned sampling approach improves reconstruction quality across multiple acceleration factors on public and in-house cardiac MRI datasets, including PSNR gains of 2-3 dB, reduced NMSE, improved SSIM, and higher radiologist ratings. The proposed scan-adaptive sampling framework enables faster and higher-quality dynamic cardiac MRI by adapting $k$-space sampling to individual scans.

Figures (10)

• Figure 1: Schematic of the proposed dynamic MostNet reconstruction framework. At each iteration, a dual-domain CRNN denoiser operating in the spatiotemporal (x-t) and temporal-frequency (x-f) domains is followed by a conjugate-gradient data-consistency (DC) update. The CRNN blocks follow the complementary dual-domain design introduced in CTFNet qin2021complementary. The notation "3x" indicates that the corresponding block is repeated three times within each denoising module. The learnable parameters $\bm{\theta}\xspace$ reside in the denoising blocks, while the DC layer enforces fidelity to the acquired multi-coil measurements.
• Figure 2: Ground truth reconstruction pipeline for the MCardiac dataset. ESPIRiT uecker2014espirit-based coil sensitivity maps are computed from the ACS region, and SENSE reconstruction is performed on zero-padded $k$-space to obtain reference images.
• Figure 3: Comparison of Cartesian undersampling masks for the MCardiac dataset. The grid displays two baseline sampling masks - Equispaced (left) and VDRS (middle) alongside the proposed dSUNO masks (right) at 4$\times$, 8$\times$, and 12$\times$ acceleration factors. These represent the retrospective undersampling patterns selected from the originally acquired prospective 2$\times$ grid.
• Figure 4: ROI-based qualitative comparison of different sampling techniques and reconstructions on two anatomically distinct slices at 4$\times$ acceleration. Each figure includes: (i) reconstructed image cropped over the region-of-interest (ROI) around the heart, (ii) spatio–temporal (x-t) profile from a horizontal stripe within the ROI, and (iii) error maps showing the absolute difference from the ground truth over the ROI.
• Figure 5: Qualitative comparison of dSUNO with other baseline sampling techniques at 8$\times$ (top panel) and 12$\times$ (bottom panel) acceleration factors for two different slices. Each panel includes the (i) reconstructed images and (ii) error maps computed against ground truth for the same frame. As the undersampling becomes more aggressive, reconstructions using the equispaced and VDRS mask degrade noticeably, while dSUNO continues to preserve anatomical details.