Magnetic resonance imaging (MRI) has emerged as a powerful modality for disease diagnosis and treatment planning [1]. It provides noninvasive and radiation-free imaging with exceptional soft-tissue contrast and enables accurate anatomical and functional measurements [2]. Nonetheless, the extended acquisition time required for sampling the entire k-space, particularly in protocols necessitating prolonged Echo Time (TE) and Repetition Time (TR) [3], presents challenges due to potential artifacts induced by the patient's physiological movements, such as cardiac motion and respiration [4]. Specific Absorption Rate (SAR) is a critical parameter in MRI operations that can limit image quality development. These artifacts can degrade image quality in the reconstructed images, thereby affecting clinical interpretation. Furthermore, the extended acquisition time restricts patient access to MRI scanners and may result in delays in patient care within the healthcare system. Consequently, there is an increasing demand for innovative approaches to overcome these limitations, enhance the efficiency and reliability of MRI acquisition and reconstruction methods, and guarantee high-quality images while minimizing acquisition time and motion-related artifacts. These advancements will not only improve clinical workflow and enhance the patient experience but will also contribute to the advancement of the MRI field in delivering accurate and timely diagnostic information.

To address the challenges of prolonged MRI acquisitions, researchers have implemented various strategies to enhance efficiency. Among these, k-space under-sampling has emerged as a widely adopted method [5]. However, selectively acquiring partial k-space data may introduce artifacts. To mitigate this, the principles of compressive sensing are incorporated, allowing advanced reconstruction algorithms to generate high-quality images from sparsely sampled data, thereby improving the practicality and reliability of MRI. Researchers have optimized different under-sampling methods, including random and equidistant under-sampling [6], considering signal and noise characteristics to maximize image detail preservation and minimize reconstruction artifacts. This integrated application of under-sampling methods and advanced reconstruction algorithms not only reduces acquisition time but also enhances the spatial and temporal resolution of MRI, opening new avenues for clinical diagnosis and research.

In our comprehensive analysis of existing reconstruction algorithms, encompassing references [7–8], we typically observe lower performance of reconstruction algorithms in the k-space domain (frequency domain) [9,10] compared to the image space domain [[11], [12], [13], [14], [15], [16], [17], [18]]. However, synergistic dual-domain reconstruction [[19], [20], [21], [22], [23], [24]] has shown potential to further enhance performance. Therefore, leveraging existing dual-domain reconstruction methods, we propose an innovative algorithm based on dual-domain cross-fusion reconstruction. Diverging from previous sequential dual-domain reconstruction approaches, we establish two separate reconstruction paths for bimodal reconstruction. In these paths, we introduce a Transformer-based Cross Attention Block (CAB) to cross-guide features at various hierarchical levels between the two paths. Consequently, during their respective reconstruction processes, features from different dimensional hierarchies receive guidance from the alternate path, enabling alternate guidance reconstruction of the two modalities. This approach maximizes the inherent relationship between them, resulting in improved reconstruction performance. Finally, we introduce a Fusion Attention Block (FAB) to dynamically fuse the two reconstruction results, refining fine details while applying global attention.

Additionally, in the context of processing MR images, various algorithms [25,26,27] often treat different regions uniformly, thereby potentially constraining the modeling and representational potential of neural networks. Within MRI, distinct regions harbor information of varying scales. Traditional neural networks, such as Convolutional Neural Networks (CNNs) or Transformers, typically treat all regions indiscriminately, limiting their modeling and representational capabilities. Consequently, our algorithm introduces a Multi-Scale Feature Extraction with a region-specific learning strategy, incorporating both high-frequency and low-frequency feature extractors. The high-frequency feature extractor is designed to focus on extracting details such as edges and textures, while the low-frequency feature extractor aims to capture low-frequency components containing global information, such as overall structure and background. Through this multi-scale feature extraction approach, we effectively capture both high-frequency and low-frequency information in the data, thereby enhancing the Transformer model's perceptual capabilities across the spectrum.

Building upon the aforementioned innovations to address current challenges, we contribute a novel algorithm for MRI reconstruction, Dual-domain Cross-fusion Transformer Network (DCT-Net). This algorithm is designed to enhance reconstruction performance under low sampling rates. Its key innovations can be summarized as follows:

Proposed MRI reconstruction method using DCT-Net. It involves bimodal parallel reconstruction with cross-guidance, to establish the spatial distribution in the image domain and the spectral representation in the frequency domain for improved reconstruction results.

Transformer-based Cross-Attention Block (CAB) and Fusion Attention Block (FAB) are employed for parallel reconstruction, providing cross-guidance and adaptive fusion.

A Multi-Scale Feature Extractor (MFE) incorporating high-frequency and low-frequency feature extractors is utilized to capture subtle variations and significant structures in MR images.