Re-Visible Dual-Domain Self-Supervised Deep Unfolding Network for MRI Reconstruction

TL;DR

This work tackles rapid MRI reconstruction under the practical constraint of lacking fully-sampled training data. It introduces a re-visible dual-domain self-supervised framework that jointly leverages under-sampled k-space data in both spatial and frequency domains, along with a CP-PPA-based deep unfolding network (DUN-CP-PPA) that embeds imaging physics and learned priors. The core innovations are the re-visible dual-domain loss, which utilizes all under-sampled data during training through a dual-branch k-space loss and image-domain losses, and the SFFE-enhanced ProxNet_x within a CP-PPA unfolding, enabling end-to-end reconstruction with strong generalization and robustness to pattern shifts and noise. Empirical results on single- and multi-coil fastMRI and IXI datasets show competitive performance against supervised methods and clear gains over existing self-supervised approaches, with ablations confirming the value of the dual-domain design and the CP-PPA-based architecture for physics-informed reconstruction.

Abstract

Magnetic Resonance Imaging (MRI) is widely used in clinical practice, but suffered from prolonged acquisition time. Although deep learning methods have been proposed to accelerate acquisition and demonstrate promising performance, they rely on high-quality fully-sampled datasets for training in a supervised manner. However, such datasets are time-consuming and expensive-to-collect, which constrains their broader applications. On the other hand, self-supervised methods offer an alternative by enabling learning from under-sampled data alone, but most existing methods rely on further partitioned under-sampled k-space data as model's input for training, resulting in a loss of valuable information. Additionally, their models have not fully incorporated image priors, leading to degraded reconstruction performance. In this paper, we propose a novel re-visible dual-domain self-supervised deep unfolding network to address these issues when only under-sampled datasets are available. Specifically, by incorporating re-visible dual-domain loss, all under-sampled k-space data are utilized during training to mitigate information loss caused by further partitioning. This design enables the model to implicitly adapt to all under-sampled k-space data as input. Additionally, we design a deep unfolding network based on Chambolle and Pock Proximal Point Algorithm (DUN-CP-PPA) to achieve end-to-end reconstruction, incorporating imaging physics and image priors to guide the reconstruction process. By employing a Spatial-Frequency Feature Extraction (SFFE) block to capture global and local feature representation, we enhance the model's efficiency to learn comprehensive image priors. Experiments conducted on the fastMRI and IXI datasets demonstrate that our method significantly outperforms state-of-the-art approaches in terms of reconstruction performance.

Figures (10)

• Figure 1: The framework of our proposed re-visible dual-domain self-supervised learning is shown. We utilize DUN-CP-PPA for reconstruction with image physics incorporated. During training, we adopt a dual-branch structure. The first branch reconstructs $\tilde{k}_p$ to boost the model’s generative capability and the second branch reconstructs $\tilde{k}$ to mitigate the input distribution shift caused by further partitioning. The model is trained using the proposed re-visible dual-domain loss. During inference, the trained network can directly reconstruct the image from $\tilde{k}$.
• Figure 2: The overall structure of the proposed Chambolle and Pock Proximal Point Algorithm based Deep Unfolding Network (DUN-CP-PPA) is shown in the first line, which includes two parts: update $x$ module and update $y$ module. Among them, the update $y$ module is in an analytical form and is computed according to the given equation. The update $x$ module includes $\operatorname{ProxNet}_x$, which is shown in the second line, and extracts multi-scale features using the U-Net architecture. In $\operatorname{ProxNet}_x$, there is a Spatial-Frequency Feature Extraction Block that captures comprehensive feature representations.
• Figure 3: Visual comparison of methods for 4$\times$ acceleration using a 1D equispaced subsampling mask on fastMRI dataset. First row: Reconstructed images; second row: Zoomed region; third row: Mask and error maps.
• Figure 4: Visual comparison of methods for 4$\times$ acceleration using a 1D random subsampling mask on IXI dataset. First row: Reconstructed images; second row: Zoomed region; third row: Mask and error maps.
• Figure 5: Visual comparison of different methods for 4$\times$ acceleration using a 1D random subsampling mask on the multi-coil knee dataset and 4$\times$ acceleration using a 1D equispaced subsampling mask on the multi-coil brain dataset. For each dataset, results are displayed as follows: the first row shows the reconstructed images, the second row presents the zoomed-in regions, and the third row illustrates the sampling masks and corresponding error maps.