JotlasNet: Joint Tensor Low-Rank and Attention-based Sparse Unrolling Network for Accelerating Dynamic MRI

TL;DR

This work tackles the challenge of accelerating dynamic MRI by formulating a joint tensor low-rank and attention-based sparse reconstruction. It introduces JotlasNet, a deep unrolled network derived from a composite splitting algorithm, that learns transform domains via CNNs and employs an attention-based soft thresholding operator to provide per-channel sparsity adaptivity. The approach lever transformed tensor nuclear norm regularization and CNN-learned sparsity, organized into a five-layer, parallel unrolling with gradient descent and Nesterov acceleration. Experiments on OCMR and CMRxRecon demonstrate superior reconstruction quality over state-of-the-art multi- and single-coil methods, with competitive speed and strong ablation support for the dual-prior design and AST mechanism.

Abstract

Joint low-rank and sparse unrolling networks have shown superior performance in dynamic MRI reconstruction. However, existing works mainly utilized matrix low-rank priors, neglecting the tensor characteristics of dynamic MRI images, and only a global threshold is applied for the sparse constraint to the multi-channel data, limiting the flexibility of the network. Additionally, most of them have inherently complex network structure, with intricate interactions among variables. In this paper, we propose a novel deep unrolling network, JotlasNet, for dynamic MRI reconstruction by jointly utilizing tensor low-rank and attention-based sparse priors. Specifically, we utilize tensor low-rank prior to exploit the structural correlations in high-dimensional data. Convolutional neural networks are used to adaptively learn the low-rank and sparse transform domains. A novel attention-based soft thresholding operator is proposed to assign a unique learnable threshold to each channel of the data in the CNN-learned sparse domain. The network is unrolled from the elaborately designed composite splitting algorithm and thus features a simple yet efficient parallel structure. Extensive experiments on two datasets (OCMR, CMRxRecon) demonstrate the superior performance of JotlasNet in dynamic MRI reconstruction.

Figures (7)

• Figure 1: The proposed JotlasNet. The network is unfolded from \ref{['eq:net']}. The red-marked symbols are learnable parameters. Five layers are included: Gradient Descent (GD), Low-Rank (LR), Sparse (S), combination, and Acceleration (ACC) layers. The LR and S layers are detailed in the text.
• Figure 2: Reconstruction results on the OCMR dataset (single-coil) using radial sampling with 16 lines. The first row shows the reconstructed images at a specific time frame, and the second row shows the corresponding error maps, except for the first image depicted the sampling mask. The third and fourth rows show the x-t images and the corresponding error maps, respectively. The position of x-t images is marked with red lines on the label image. The PSNR values respected to this test image are also listed.
• Figure 3: Reconstruction results on the CMRxRecon dataset (multi-coil) using 4X equispaced sampling. The first row shows the reconstructed images at a specific time frame, and the second row shows the corresponding error maps, except for the first image depicted the sampling mask. The third and fourth rows show the x-t images and the corresponding error maps, respectively. The position of x-t images is marked with red lines on the label image. The PSNR values respected to this test image are also listed.
• Figure 4: Reconstruction results of JotlasNet, CineVN, and T2LR-Net on the OCMR dataset (multi-coil) using 8-fold VISTA sampling.
• Figure 5: The PSNR of JotlasNet under different numbers of iterations (N). The blue line shows the mean PSNRs while the error bars represents the standard deviations.