Improved radon based imaging using the shearlet transform

Glenn R. Easley, Flavia Colonna, Demetrio LabateProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2009 · DOI

TL;DR

This paper introduces a novel method for reconstructing images from noisy Radon transform data, a common problem in imaging modalities like SAR and CT. By leveraging the shearlet transform, a powerful multiscale and multidirectional signal processing tool, the authors achieve significantly improved image reconstruction quality compared to existing wavelet and curvelet-based techniques, especially in the presence of noise.

What Problem Does This Paper Address?

Many imaging systems, such as Synthetic Aperture Radar (SAR), Magnetic Resonance Imaging (MRI), and Computed Tomography (CT), collect data in a Radon transform domain. To obtain an image, this data must be inverted. However, this inversion process is notoriously unstable, meaning that even small amounts of noise in the collected data can lead to severely degraded or artifact-ridden reconstructed images. This instability necessitates regularization techniques to suppress noise amplification and recover a meaningful image. While wavelet and curvelet transforms have been used for this purpose, this paper aims to develop a more effective shearlet-based regularization method that offers near-optimal performance for a broad class of images.

Methods

The core of this work lies in using the shearlet transform for regularizing the Radon transform inversion. Shearlets are a type of "composite wavelet" that, unlike traditional wavelets, can effectively capture both multiscale and multidirectional information in an image. This is achieved through a combination of anisotropic scaling and shearing operations.

The authors adapt a general framework for operator-biorthogonal decomposition, originally developed for Radon transform inversion, and apply it to shearlets. They define specific shearlet basis functions and their "companion representations" which are derived using fractional Laplacians. These companion functions, denoted as $s_\mu$ and $s'_\mu$, form frames for $L^2(\mathbb{R}^2)$, meaning they can represent any function in the space and allow for stable reconstruction.

The Radon transform inversion is then formulated as $f = R^*KR$, where $R^*$ is a dual transform and $K$ is an operator. The shearlet-based approach replaces the standard inversion with a regularized version that utilizes the shearlet companion functions. Specifically, they define a "Radon isometry" $R = (\Delta_1 \circ I) \circ R$ and construct frames $\{U_\mu\}$ and $\{V_\mu\}$ from the shearlet companion functions. The inversion formula then becomes $f = \sum_\mu [Rf, U_\mu] 2^{3j/2} s_\mu$, where $[Rf, U_\mu]$ are the noisy Radon transform coefficients projected onto the shearlet frame. For noisy data, a soft thresholding function $T_\epsilon$ is applied to these coefficients to estimate the true image.

Datasets

The paper evaluates its method using the following:

  • Shepp-Logan Phantom: A standard phantom image used for testing tomographic reconstruction algorithms.
  • Star-like image: A synthetic image designed to test performance on "star-like domains," where shearlets and curvelets are known to perform well.
  • ISAR dataset: Real-world Inverse Synthetic Aperture Radar (ISAR) data collected from a SAAB 9000 car.

The paper does not introduce new benchmark datasets but rather uses these existing ones to compare its method against others.

Key Insights and Ideas

  • Shearlets excel at capturing directional information: The paper highlights that shearlets, with their inherent ability to represent directional features at multiple scales, are particularly well-suited for inverting Radon transforms, which often involve data from various angles.
  • Near-optimal regularization: The shearlet-based decomposition provides a regularization that is shown to be nearly optimal for a general class of images, outperforming previous methods.
  • Improved performance on star-like domains: Consistent with theoretical predictions, shearlets (and curvelets) demonstrate superior performance for images with star-like features compared to wavelets.
  • Robustness to noise: The shearlet-based approach significantly improves the Signal-to-Noise Ratio (SNR) of reconstructed images, even when the input data is heavily corrupted by noise.
  • Potential for automatic thresholding: The paper suggests that the shearlet framework is well-suited for using Generalized Cross-Validation (GCV) functions for automatic determination of thresholding parameters, simplifying the practical application of the method.

Key Results

The paper presents quantitative results demonstrating the superiority of the shearlet-based Radon transform inversion:

  • Shepp-Logan Phantom:
    • Noisy reconstruction SNR: 7.48 dB
    • Wavelet-based estimate SNR: 16.64 dB
    • Curvelet-based estimate SNR: 18.72 dB
    • Shearlet-based estimate SNR: 19.75 dB
  • Star-like image:
    • Noisy reconstruction SNR: 8.02 dB
    • Wavelet-based estimate SNR: 14.83 dB
    • Curvelet-based estimate SNR: 15.90 dB
    • Shearlet-based estimate SNR: 16.78 dB
  • ISAR dataset: While specific SNR values are not provided for the ISAR dataset, the paper states that the shearlet-based reconstruction is visually superior, and the curvelet-based estimate is noted as being "very poor" due to issues with its downsampling implementation.

Qualitatively, the reconstructed images using shearlets appear sharper and contain fewer artifacts compared to wavelet and curvelet-based reconstructions, especially in the presence of noise.

Limitations and Future Work

  • Limitations:
    • The paper notes that for the ISAR dataset, the curvelet-based estimate was poor due to downsampling issues in its implementation, which might not be a general limitation of curvelets themselves but rather a specific implementation detail.
  • The theoretical MSE approximation rate for shearlets is $O(\epsilon^{4/3})$ for star-like domains, while wavelets are $O(\epsilon)$, suggesting that while shearlets are better, there might still be room for improvement in approximation rates for certain image classes.
  • Future Work:
    • The authors suggest that the shearlet framework is well-suited for using GCV functions for automatic determination of threshold parameters, implying this as a promising direction for practical implementation.
  • The paper's success with shearlets for Radon transform inversion opens doors for exploring their application in other imaging modalities and inverse problems where directional information is crucial.

Summary updated

Abstract

Many imaging modalities, such as Synthetic Aperture Radar (SAR), can be described mathematically as collecting data in a Radon transform domain. The process of inverting the Radon transform to form an image can be unstable when the data collected contain noise so that the inversion needs to be regularized in some way. In this work, we develop a method for inverting the Radon transform using a shearlet-based decomposition, which provides a regularization that is nearly optimal for a general class of images. We then show through a variety of examples that this technique performs better than similar competitive methods based on the use of the wavelet and the curvelet transforms.
BibTeX
@inproceedings{easley2009improved,
  author = {Glenn R. Easley and Flavia Colonna and Demetrio Labate},
  title = {{Improved radon based imaging using the shearlet transform}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2009},
  doi = {10.1117/12.820066},
}
Figure