TL;DR
This paper proposes a novel approach to accelerating MRI scans by improving the reconstruction of undersampled data. They introduce domain-specific modules for image and k-space reconstruction, arguing that these domains have distinct properties that require tailored solutions, leading to significant performance gains over existing methods.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but acquiring high-quality images can be time-consuming. To speed up scans, researchers often undersample the data collected in k-space, which is the raw data space where MRI signals are recorded. However, this undersampling leads to aliasing and blurring in the reconstructed image. Reconstructing high-quality images from this undersampled data is a critical challenge. While deep learning methods, particularly those that operate in both the image and k-space domains (dual-domain reconstruction), have shown promise, existing approaches often use the same network architecture for both domains. This paper argues that this is suboptimal because image and k-space data have fundamentally different characteristics and require domain-specific processing for effective reconstruction.
Methods
The authors propose a dual-domain reconstruction framework that incorporates two novel domain-specific modules:
- K-space Global Initialization (K-GLIM): This module is designed to improve the initial reconstruction in the k-space domain. It utilizes a Channel-wise Multi-head Self-attention (C-MSA) mechanism to capture global information and interactions within the k-space data, which is crucial for accurate interpolation.
- Image Parallel Local Detail Enhancement (I-PLDE): This module focuses on enhancing local details in the image domain. It employs a parallel branch that emphasizes local information, inspired by "divide-and-conquer" strategies. This approach aims to refine image details without being hindered by the large receptive fields often needed for artifact removal.
These modules are integrated into a state-of-the-art dual-domain network, DuDoRNet, resulting in their proposed model, DuDoRNet+. The framework also includes domain-specific Shallow Feature Extraction (X-SFE) and Global Feature Refinement (GFR) components, along with four stages of domain-specific building blocks (X-Si).
Datasets
The primary dataset used for training and evaluation is the Multi-Contrast IXI dataset. The authors uniformly sampled 14 slices from T2-PD paired volumes. The dataset was split patient-wise into training (70%), validation (10%), and testing (20%) sets. Images were center-cropped to 224x224.
Key Insights and Ideas
- Domain-Specific Processing is Crucial: The core insight is that image and k-space domains have distinct properties that necessitate tailored reconstruction strategies. Treating them identically, as many prior methods do, limits performance.
- Receptive Field Matters for K-space Interpolation: The paper highlights that the feasibility of interpolating k-space data is directly related to the size of the receptive field. As acceleration rates increase, the probability of successful interpolation decreases, underscoring the need for effective k-space processing.
- Parallel Local Detail Enhancement: The I-PLDE module demonstrates that a parallel approach to capturing local details in the image domain can effectively complement global artifact removal, leading to sharper and more refined reconstructions.
Key Results
The proposed DuDoRNet+ model achieved state-of-the-art performance across various MRI reconstruction settings:
- Image-domain reconstruction: Achieved the best PSNR and SSIM values.
- Dual-domain reconstruction: Outperformed existing methods under both 4x and 8x acceleration rates.
- Reference-protocol-guided dual-domain reconstruction: Demonstrated superior results, achieving the highest PSNR (40.59 dB) and SSIM (99.29%) at 8x acceleration.
The paper also includes an ablation study (Table 2) showing that applying the proposed K-GLIM and I-PLDE modules to their respective domains (k-space and image) yields the best results, confirming the importance of domain-specific design.
Limitations and Future Work
The authors acknowledge that their current work focuses on single-coil MRI reconstruction. Future work will involve extending their domain-specific design principles to multi-coil reconstruction and exploring their utility under different sampling patterns. They also plan to validate their approach by transforming other state-of-the-art models.
Summary updated