Deep Residual Learning for Accelerated MRI using Magnitude and Phase Networks
The research paper presents an investigation into accelerating magnetic resonance imaging (MRI) scan acquisition by employing deep residual learning networks. These networks specifically address aliasing artifacts in MRI data, which are common issues in accelerated MRI procedures that rely on compressed sensing (CS) and parallel imaging. The paper focuses on leveraging the strengths of convolutional neural networks (CNNs) in reconstructing MRI images by separating magnitude and phase data.
Deep Residual Network Architecture
The proposed architecture utilizes a dual approach by implementing separate CNNs for magnitude and phase components, forming two distinct networks. The magnitude network focuses on image domain post-processing when only magnitude data is available, whereas the phase network can be exploited iteratively in conjunction with available phase data for k-space interpolation. This innovative strategy is an extension of the U-Net architecture, augmented with the concept of framelet representation, enabling efficient encoding and decoding of features across various scales of image resolutions.
The use of framelet-based techniques—specifically the dual frame U-net—plays a pivotal role in this work. The architecture considers the globalized artifact patterns, especially aliasing, which are complex and not localized; this is achieved by having convolutional layers capable of processing large receptive fields.
Methodology and Numerical Evaluation
The research addresses aliasing artifact removal using networks that were pre-trained and further trained iteratively using the Krasnoselskii-Mann (KM) iteration method to reconstruct MR images effectively. The method showed strong performance, providing accurate and artifact-free images significantly faster than competitive methods like GRAPPA and ALOHA, which are well-regarded in the parallel imaging and CS frameworks, respectively.
The application of this network framework on single and multi-channel MR data showcases a reduction in computational overhead by orders of magnitude, making it a practical option for MRI applications requiring rapid reconstructions. The experiments demonstrated enhancement for several sampling patterns with variation in Auto-Calibration Signal (ACS) lines configurations.
Implications and Future Directions
The implications of this paper are twofold: practically, the framework provides a means to drastically reduce the scan and reconstruction time, enhancing the clinical workflow. Theoretically, the integration of separate magnitude and phase networks suggests potential areas for further exploration in image reconstruction, particularly in domains where the complex nature of data might be split similarly.
This work opens up avenues for future research focusing on complex signal processing in convolutional architectures. Given that MRI procedures produce inherently multidimensional and complex-valued data, further research could focus on the subtleties of loss functions to improve visual quality beyond mere error metrics. Additionally, more sophisticated deep learning strategies like generative adversarial networks (GANs) or fully unsupervised learning mechanisms might provide further improvements in both the efficiency and quality of image reconstructions.
In conclusion, this paper offers a robust contribution to the field of MRI reconstruction. By addressing the key problem of aliasing artifacts with a scalable and efficient solution, it establishes a forward-looking framework for future explorations not only in the context of MRI but potentially for other imaging technologies dealing with similar challenges.