TL;DR
This paper introduces K-UNN, a novel method for accelerating MRI by interpolating missing k-space data using an "untrained" neural network (UNN). K-UNN leverages physical priors of MR images (sparsity, coil sensitivity smoothness, and phase smoothness) to achieve accurate reconstruction, even outperforming some supervised deep learning methods.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its slow acquisition speed is a major limitation. While techniques like Compressed Sensing (CS) and deep learning (DL) have been developed to accelerate MRI, they often require extensive fully-sampled training data (impractical in clinics) or fail to fully utilize the inherent physical properties of MR images. Existing UNN-based methods, while avoiding the need for training data, often struggle with common sampling patterns and lack theoretical guarantees for reconstruction accuracy. This paper aims to bridge this gap by developing a UNN that effectively incorporates MR physics priors for robust k-space interpolation.
Methods
The core of this work is the proposed "tripled UNN" architecture. Unlike previous UNN approaches that primarily rely on general image priors, K-UNN explicitly incorporates three key physical priors of MR images:
- Sparsity: The underlying MR image is sparse in a specific transform domain.
- Coil Sensitivity Smoothness: The sensitivity maps of different coils are smooth.
- Phase Smoothness: The phase information of the MR signal is smooth.
The UNN is designed with three modules, each responsible for learning one of these priors. These modules work together to generate k-space data that is consistent with the acquired undersampled measurements and adheres to the physical constraints. The authors also provide theoretical guarantees for the accuracy of the interpolated k-space data, particularly for random and deterministic sampling trajectories.
Datasets
The study evaluated the proposed K-UNN on two types of MR data:
- Knee Data: Acquired from a 3T Siemens scanner using a 2D Cartesian turbo spin echo (TSE) protocol.
- Brain Data: Acquired from a 3T Siemens scanner using a Cartesian 2D gradient echo (GRE) protocol.
The paper mentions using data from seven subjects for the knee data, with 227 slices used as a "training set" for comparison experiments with other methods. However, it's crucial to note that K-UNN itself does not require training data in the traditional sense; this "training set" was likely used to generate undersampled data for evaluating the performance of K-UNN against other methods that do require training.
Key Insights and Ideas
- Physics-Informed UNNs: The paper demonstrates that explicitly incorporating physical priors into UNNs significantly enhances their performance for MR reconstruction, moving beyond generic image priors.
- Tripled Architecture for Comprehensive Priors: The novel "tripled UNN" effectively combines sparsity, coil sensitivity, and phase smoothness, leading to more accurate and robust k-space interpolation.
- Theoretical Guarantees: Providing theoretical bounds on the accuracy of interpolated k-space data is a significant contribution, offering confidence in the method's reliability.
- Flexibility: The modular design of the UNN allows for flexibility; if certain priors are not applicable, modules can be removed, forming a "double UNN" architecture.
Key Results
- Superior Reconstruction Quality: K-UNN consistently outperformed traditional parallel imaging methods (L1-SPIRIT, L1-SENSE-LORAKS) and image-domain UNNs (I-UNN) across various sampling trajectories (random, VD regular, partial Fourier).
- Competitive with Supervised DL: In several cases, K-UNN achieved comparable or even superior performance to state-of-the-art supervised-trained k-space deep learning methods (e.g., H-DSLR).
- Robustness to Undersampling: K-UNN showed remarkable robustness to insufficient calibration (ACS) data, maintaining high reconstruction quality where other methods struggled.
- Accurate Prior Characterization: Ablation studies confirmed that K-UNN more accurately characterizes the physical priors of MR images compared to traditional methods.
Limitations and Future Work
- Priors Mismatches: The authors acknowledge that in reality, there might be cases where the assumed priors (sparsity, phase smoothness, coil sensitivity smoothness) are not perfectly met. They suggest that the modular UNN architecture can be recomposed to adapt to such scenarios.
- ACS Data Utilization: The current K-UNN does not explicitly utilize acquired ACS data. Future work could explore how to incorporate ACS data to further improve performance.
- Inter-frame Similarity: For dynamic imaging, the similarity between consecutive frames is a valuable prior. Future work could involve developing an untrained deformation operator to leverage this inter-frame similarity.
- Computational Speed: While K-UNN offers excellent reconstruction quality, solving the optimization problem for each slice individually is computationally intensive, limiting imaging speed. Future research could focus on meta-learning or other techniques to accelerate the imaging process.
Summary updated