Magnetic Resonance Imaging (MRI), as a non-invasive and high-resolution medical imaging technique, is extensively utilized in disease diagnosis and scientific research. MRI provides detailed visualization of human soft tissue structures and plays a critical role in detecting pathologies in vital regions such as the brain, heart, and joints. However, a primary challenge associated with MRI is its prolonged scanning duration, which not only increases patient discomfort but also introduces motion artifacts that degrade image quality and compromise diagnostic accuracy. Furthermore, lengthy scan times restrict the applicability of MRI in urgent clinical scenarios, such as acute stroke assessment, dynamic cardiac imaging, or intraoperative navigation, where rapid imaging is essential. Consequently, achieving significant reductions in MRI acquisition time while reconstructing high-resolution images that meet clinical standards has become a pivotal issue in the field of MRI research.

Traditional MRI acceleration methods can be broadly classified into two categories. The first category includes Parallel Imaging (PI) techniques [1], [2], [3], [4], which accelerate data acquisition by leveraging multiple receiver coils to simultaneously capture signals from different spatial regions. PI exploits the spatial sensitivity differences between coils to reconstruct complete images through methods such as SMASH [5], SENSE [6], GRAPPA [7], and E-SPIRiT [8]. While PI effectively reduces scan times, its acceleration capability is constrained by the number and arrangement of coils. Under high acceleration rates, PI is prone to signal-to-noise ratio (SNR) degradation, the appearance of artifacts, and increased coil costs, which limit its application in scenarios requiring higher acceleration demands [9].

The second category involves undersampling and reconstruction methods based on Compressed Sensing (CS) theory [10], [11], [12]. By undersampling data in k-space, CS-MRI significantly reduces scan times and employs sparse representation and optimization algorithms to recover missing data. Traditional CS-MRI methods rely on predefined sparse transforms, such as Total Variation (TV) [13], Wavelet Transform (WT) [14], [15], and Discrete Cosine Transform (DCT). These methods reconstruct images by constructing sparsity-regularized models. However, under high acceleration rates, they often result in artifacts and suboptimal reconstruction quality. Furthermore, the selection of sparse transforms depends on prior knowledge, making them less adaptable to diverse imaging data. Additionally, the iterative optimization algorithms used in these methods are computationally intensive and time-consuming, rendering them unsuitable for real-world scenarios that demand efficient, real-time imaging.

With the advancement of deep learning technologies, data-driven approaches have introduced new possibilities for MRI image reconstruction. Convolutional Neural Networks (CNNs) have demonstrated the ability to effectively restore high-quality MRI images by learning the mapping relationship between undersampled data and fully sampled images. Wang et al. [16] were the first to apply CNNs to the MRI reconstruction task, proposing an offline, end-to-end model capable of recovering high-quality images from undersampled data. Schlemper et al. [17] developed a Deep Cascade CNN that employs a multi-layer cascade structure to enhance the reconstruction accuracy of dynamic MRI images. Furthermore, Eo et al. [18] proposed a dual-domain cascaded network, KIKI-Net, for information recovery in both k-space and the image domain.

CNNs have also been utilized to design symmetric encoder–decoder structures, such as the U-Net model [19]. Through skip connections, U-Net not only effectively preserves shallow features but also captures partial long-range feature information. Jin et al. [20] introduced FBPConvNet, which leverages the U-Net structure to address inverse problems in biomedical imaging. Following this, U-Net and its variants have demonstrated robust performance in generative tasks [21], [22], [23]. However, the local receptive field of CNNs limits their ability to capture complex global dependencies, leading to inadequate performance in processing intricate textures and edge details.

Subsequently, Generative Adversarial Networks (GANs) [24] were introduced into MRI reconstruction tasks. By leveraging the adversarial learning mechanism between the generator and discriminator, GANs have demonstrated exceptional capabilities in image generation. For instance, Yang et al. [25] developed DAGAN based on a U-Net architecture, incorporating perceptual loss to enhance the realism of image textures. Quan et al. [26] further advanced this approach by proposing RefineGAN, which comprises two cascaded networks for initial reconstruction and refinement, employing cyclic data consistency loss to improve the recovery of complex structures and high-frequency information. Unlike GANs trained solely in the image domain, Shaul et al. [27] designed a cross-domain framework, KIGAN, to estimate missing k-space data and separately address aliasing artifacts. Liu et al. [28] proposed DBGAN, which introduced cross-stage skip connections and instance normalization into a dual-branch generator to improve information propagation and structural detail preservation. Zhao et al. [29] subsequently introduced SwinGAN, a dual-domain Swin Transformer-based GAN with contextual image relative position encoding, enhancing the modeling of local contextual dependencies. They further proposed DiffGAN [30], which integrates a local transformer-based generator with adversarial reverse diffusion, leveraging the synergy between diffusion models and GANs to boost detail recovery and training stability. Xu et al. [31] designed SepGAN, a lightweight GAN utilizing depthwise separable convolutions and attention modules to reduce parameter complexity while maintaining strong performance. Most recently, Noor et al. [32] presented DLGAN, a hierarchical GAN tailored for brain and knee MRI, which demonstrated notable reconstruction accuracy and robustness against aliasing artifacts across anatomical variations. Despite these advances, GAN-based methods still face two critical challenges: the limited receptive field of convolutional operations restricts their ability to capture global structural information, leading to blurring and loss of fine details at high acceleration rates; and the substantial model complexity and parameter count hinder their practical deployment in resource-constrained settings.

Motivated by the aforementioned challenges, this paper proposes a novel lightweight generative adversarial network based on multi-scale context awareness (MCA-GAN), which can rapidly reconstruct high-resolution images meeting clinical requirements under severely undersampled conditions. Specifically, MCA-GAN employs dual-domain generators to improve reconstruction quality through collaborative optimization of k-space and image domain data. The k-space generator restores undersampled k-space data, while the image-domain generator further refines the reconstruction results in the image domain. These generators are connected via inverse Fourier transform (IFT), enabling precise reconstruction by leveraging cross-domain information. Moreover, the network introduces several innovative modules: the Depthwise Separable Local Attention (DWLA) module and Adaptive Group Rearrangement Block (AGRB) for high-quality feature extraction, as well as the Multi-Scale Spatial Context Modulation Bridge (MSCMB) and Channel-Spatial Multi-Scale Self-Attention (CSMS) module for multi-scale feature fusion and capturing long-range dependencies. These modules, designed under the lightweight principle, enhance the reconstruction performance under undersampled conditions while significantly reducing parameter size and computational complexity, offering an efficient and accurate solution for MRI accelerated imaging.

In summary, the main contributions of this work are as follows:

A novel lightweight generative adversarial network (MCA-GAN) based on multi-scale context awareness is proposed. By leveraging multi-scale feature extraction and fusion, combined with collaborative optimization of k-space and image domain data, the proposed approach achieves high-quality MRI reconstruction.

Four plug-and-play core modules, DWLA, AGRB, CSMS, and MSCMB, are designed to enable efficient feature extraction and information capture through mechanisms such as local attention, group rearrangement, multi-scale self-attention, and contextual feature modulation. The design effectively reduces parameter size and computational complexity while maintaining high performance.

Extensive experiments conducted on multiple public datasets validate the outstanding performance of MCA-GAN under various undersampling rates and undersampling masks. The proposed method surpasses comparative approaches in terms of reconstruction quality (PSNR and SSIM), model efficiency, and robustness, demonstrating excellent generalization ability and application potential.