TL;DR
This paper introduces a novel algorithm, pFISTA, for faster and higher-quality image reconstruction in Compressed Sensing Magnetic Resonance Imaging (CS-MRI) using tight frames. By reformulating the analysis model and incorporating an orthogonal projection, pFISTA achieves comparable or better results than existing methods with fewer parameters and improved memory efficiency.
What Problem Does This Paper Address?
Compressed Sensing (CS) has revolutionized MRI by allowing faster image acquisition through undersampling k-space data. However, this speedup comes at the cost of image reconstruction challenges, particularly achieving both speed and high image quality. Redundant representations like tight frames can improve image quality, but efficiently solving the reconstruction problem with these representations has been a hurdle. Existing methods often struggle with parameter tuning, memory usage, or computational complexity when dealing with tight frames in CS-MRI. This paper aims to bridge this gap by developing an efficient and user-friendly algorithm for CS-MRI reconstruction using tight frames.
Methods
The core contribution of this paper is the development of a Projected Iterative Soft-Thresholding Algorithm (pISTA) and its accelerated version, Projected Fast Iterative Soft-Thresholding Algorithm (pFISTA).
The authors tackle the analysis model for CS-MRI, which is known to offer better reconstruction quality than synthesis models but is harder to solve. They achieve this by:
- Introducing the Canonical Dual Frame: This allows for the construction of an orthogonal projection operator onto the range of the analysis sparsity operator.
- Reformulating the Analysis Model: The constrained analysis model is transformed into a form that resembles a synthesis model, making it amenable to existing iterative algorithms.
- Projected Proximal Mapping: Instead of directly solving the constrained problem, pISTA replaces the constrained proximal map with an unconstrained proximal map followed by the orthogonal projection. This is a key innovation that simplifies the algorithm.
- Incorporating FISTA's Acceleration Strategy: pFISTA builds upon pISTA by applying the acceleration technique from the Fast Iterative Soft-Thresholding Algorithm (FISTA), significantly speeding up convergence.
Theoretically, the authors prove that pISTA and pFISTA converge to a "balanced sparse model," which offers a good trade-off between synthesis and analysis models.
Datasets
The experiments were conducted on three typical MRI datasets:
- A water phantom image.
- A T2-weighted brain image.
- A T1-weighted brain image.
These images were acquired using different MRI systems and sequences, and then retrospectively undersampled to emulate single-channel MRI data. The authors also experimented with different tight frames (contourlet, patch-based directional wavelets (PBDW), and framelet) and undersampling patterns (1D along phase encoding, 2D, and radial) to demonstrate the algorithm's versatility.
Key Insights and Ideas
- Bridging Analysis and Synthesis: The proposed pFISTA algorithm effectively bridges the gap between the more accurate but complex analysis models and the simpler but less accurate synthesis models. It achieves this by solving a "balanced sparse model" that inherits benefits from both.
- Orthogonal Projection as a Simplifier: The clever use of an orthogonal projection operator allows for the adaptation of algorithms typically designed for synthesis models to solve analysis models, a significant simplification.
- Parameter Stability is Key: A major advantage highlighted is the algorithm's stability with respect to the step size parameter. This makes it much easier for users to tune and apply in practice, unlike some other methods that are highly sensitive to parameter choices.
- Memory Efficiency for Redundant Frames: By operating in the image domain rather than the coefficient domain, pFISTA offers substantial memory savings, which is crucial when dealing with highly redundant tight frames.
Key Results
- Improved Reconstruction Quality: pFISTA consistently achieved better reconstruction quality (lower Relative L2 Norm Error - RLNE) compared to FISTA (which solves the synthesis model).
- Faster Convergence: pFISTA demonstrated significantly faster convergence than the state-of-the-art smoothing FISTA (SFISTA), which is designed for analysis models.
- Comparable Performance Across Frames: The advantages of pFISTA were observed across various tight frames (SIDWT, contourlet, PBDW, framelet), indicating its general applicability.
- Robustness to Step Size: Experiments showed that the reconstruction error of pFISTA is largely insensitive to the step size parameter, making it user-friendly. A step size of 1 is recommended for a good balance of performance and speed.
- Reduced Memory Footprint: The algorithm's operation in the image domain leads to significant memory savings, especially for highly redundant tight frames.
Limitations and Future Work
- Convergence to Balanced Sparse Model: While the paper proves convergence to a balanced sparse model, it acknowledges that this is an approximation of the exact analysis model. However, experimental results suggest this approximation yields comparable or even better results in practice.
- Future Work: The authors suggest extending the analysis to more general frames and dictionaries. They also plan to explore advanced adaptively sparse representations for CS-MRI.
Summary updated