Super-resolution MRI using finite rate of innovation curves

Greg Ongie, Mathews Jacob

Abstract

Magnetic Resonance Imaging (MRI) often involves a trade-off between image resolution and acquisition time. This paper presents a novel two-stage super-resolution algorithm that leverages the theory of Finite Rate of Innovation (FRI) to recover sharp edges from limited k-space samples. The method first estimates a continuous-domain edge mask by identifying the "innovation" parameters of the image geometry and then utilizes this mask to guide a weighted Total Variation (TV) reconstruction. Experimental results on analytical phantoms demonstrate that the proposed approach significantly outperforms conventional techniques like standard TV and Cadzow denoising, providing higher signal-to-noise ratios (SNR) and superior visual fidelity.

1. Introduction

Magnetic Resonance Imaging (MRI) faces a fundamental trade-off: high-resolution images require long acquisition times and often suffer from a poor signal-to-noise ratio (SNR). Consequently, clinicians frequently settle for low-resolution data, which can lead to blurring and "leakage" artifacts—phenomena where signals from different tissues, such as fat and water, bleed into adjacent regions.

While various super-resolution techniques have been proposed to sharpen these images, most rely on discrete models that struggle to capture the continuous nature of tissue boundaries. This work addresses the challenge of recovering sharp, off-the-grid edges from limited, noisy Fourier (k-space) samples, aiming to provide a robust framework suitable for practical medical imaging applications.

2. Methodology

The proposed reconstruction pipeline consists of two primary stages: edge mask estimation and weighted reconstruction.

2.1 Edge Mask Estimation via FRI

The edges of the image are modeled as a "trigonometric curve," defined as the zero set of a bandlimited polynomial. Leveraging the theory of Finite Rate of Innovation (FRI), the authors demonstrate that the Fourier coefficients of the image's derivatives are annihilated when convolved with the coefficients of this edge mask. To ensure robustness against noise and model order uncertainty, the algorithm employs Singular Value Decomposition (SVD) on a large convolution matrix. By averaging multiple vectors from the SVD null space, a smooth and reliable spatial mask is generated, effectively suppressing spurious edges.

2.2 Weighted Total Variation Reconstruction

Once the edge mask is estimated, it is discretized and incorporated as a weight ($W_\mu$) into a Weighted Total Variation (TV) minimization framework. This weighting scheme guides the reconstruction process by penalizing sharp changes only in regions where the mask is non-zero, thereby preserving sharp boundaries at the identified edge locations while maintaining smoothness elsewhere.

3. Experimental Evaluation

The algorithm was evaluated using two analytical MRI phantoms: the standard Shepp-Logan phantom and a more complex analytical brain phantom. These models allow for the generation of "infinite-resolution" ground truth data. K-space samples were generated directly from these models, and complex white Gaussian noise was added to simulate realistic scanner conditions. The performance was compared against standard Total Variation (TV) regularization and Cadzow-based denoising methods.

4. Results and Discussion

The proposed method consistently outperformed standard techniques in both visual clarity and quantitative metrics. For the Shepp-Logan phantom in a noiseless setting with 20-fold undersampling, the proposed method achieved an SNR of 21.3 dB, significantly higher than the 16.6 dB achieved by standard TV.

Comparison of reconstruction methods: (a, d) Least-Squares, (b, e) Cadzow, and (c, f) the Proposed FRI-based method. The top row (a-c) corresponds to the Shepp-Logan phantom, while the bottom row (d-f) corresponds to the Analytical Brain phantom.

As shown in the figure above, the proposed approach produces much cleaner edges and fewer aliasing artifacts compared to Least-Squares and Cadzow methods, even under significant noise conditions (SNR=20 dB). The visual fidelity of the reconstructions is notably higher, with better preservation of fine details, such as small simulated lesions, which are typically obscured by standard smoothing filters.

5. Conclusion and Future Work

This paper introduces a robust two-stage algorithm for super-resolution MRI using FRI curves. By identifying continuous-domain edges and using them to guide reconstruction, the method achieves superior performance over discrete-model-based techniques. Future work will focus on reducing the computational complexity associated with SVD on large matrices for 3D data and automating parameter selection for varying noise levels. Additionally, validating the method on raw data from clinical MRI scanners is a key next step.

Summary updated

BibTeX
@inproceedings{ongie2015super,
  author = {Greg Ongie and Mathews Jacob},
  title = {{Super-resolution MRI using finite rate of innovation curves}},
  booktitle = {ISBI},
  year = {2015},
  doi = {10.1109/isbi.2015.7164100},
}
Figure