CAMP-Net: Consistency-Aware Multi-Prior Network for Accelerated MRI Reconstruction

TL;DR

CAMP-Net addresses the challenge of reconstructing high-fidelity MRI images from undersampled k-space by integrating image-domain priors, k-space priors, and calibration priors within an unrolled optimization framework. It introduces a consistency-aware, multi-prior architecture featuring the Spatial Encoding Module, Image Enhancement, k-space Restoration, Calibration Consistency, and Frequency Fusion, augmented by a Surface Data Fidelity layer and adjacent-slice context. The approach yields state-of-the-art PSNR/SSIM and improved $T_2$ mapping on Calgary-Campinas, SKM-TEA, and fastMRI Knee, particularly at high accelerations, demonstrating robust cross-dataset generalization. The work advances practical accelerated MRI by enabling more accurate restoration of fine anatomical details and has potential extensions to dynamic MRI, super-resolution, denoising, and motion artifact correction.

Abstract

Undersampling k-space data in MRI reduces scan time but pose challenges in image reconstruction. Considerable progress has been made in reconstructing accelerated MRI. However, restoration of high-frequency image details in highly undersampled data remains challenging. To address this issue, we propose CAMP-Net, an unrolling-based Consistency-Aware Multi-Prior Network for accelerated MRI reconstruction. CAMP-Net leverages complementary multi-prior knowledge and multi-slice information from various domains to enhance reconstruction quality. Specifically, CAMP-Net comprises three interleaved modules for image enhancement, k-space restoration, and calibration consistency, respectively. These modules jointly learn priors from data in image domain, k-domain, and calibration region, respectively, in data-driven manner during each unrolled iteration. Notably, the encoded calibration prior knowledge extracted from auto-calibrating signals implicitly guides the learning of consistency-aware k-space correlation for reliable interpolation of missing k-space data. To maximize the benefits of image domain and k-domain prior knowledge, the reconstructions are aggregated in a frequency fusion module, exploiting their complementary properties to optimize the trade-off between artifact removal and fine detail preservation. Additionally, we incorporate a surface data fidelity layer during the learning of k-domain and calibration domain priors to prevent degradation of the reconstruction caused by padding-induced data imperfections. We evaluate the generalizability and robustness of our method on three large public datasets with varying acceleration factors and sampling patterns. The experimental results demonstrate that our method outperforms state-of-the-art approaches in terms of both reconstruction quality and $T_2$ mapping estimation, particularly in scenarios with high acceleration factors.

Figures (13)

• Figure 1: Convergence trajectories of reconstruction under different conditions.
• Figure 2: Architecture of CAMP-Net for accelerated MRI reconstruction. CAMP-Net takes as input under-sampled k-space data and sampling masks of consecutive slices. With coil sensitivity maps from the SEM module, the IEM, KRM, CCM, and FFM modules iteratively reconstruct the slices. Only the center slice undergoes direct supervision via its reconstruction loss; adjacent slices of the reference appear transparent, indicating indirect supervision. Solid and dashed lines represent information flow, with dashed lines depicting inputs external to iterations interacting internally within the modules.
• Figure 3: Details of KRM and CCM. The KRM takes multi-slice multi-coil $k$-space data as input and utilizes a complex-valued network to refine the data, along with a surface data fidelity layer to handle data imperfections. The CCM takes multi-slice multi-coil ACS data as input and utilizes another complex-valued network and surface data fidelity layer to encode calibration information from the input data to guide the KRM in learning consistency-aware correlations.
• Figure 4: Sampling patterns: (a)-(b) 2D Poisson Disc sampling mask at acceleration factors of 5X and 10X for the Calgary-Campinas dataset, with an ACS region within a radius of 16. (c)-(d) 2D Poisson Disc undersampling at acceleration factors of 6X and 8X for the SKM-TEA dataset, with a kernel width of 6 and an ACS region of size 24$\times$24. (e)-(f) 1D random sampling mask at acceleration factors of 4X and 8X for the fastMRI Knee dataset, with ACS lines covering 8% and 4% of the total $k$-space phase lines, respectively.
• Figure 5: Performance of reconstruction models (colored legend) for the Calgary-Campinas dataset at 5X (left) and 10X (right) acceleration.