Penalized-likelihood reconstruction for sparse data acquisitions with unregistered prior images and compressed sensing penalties

J. Webster Stayman, Wojciech Zbijewski, Yoshito Otake, Ali Uneri, Sebastian Schäfer, J. Lee, Jerry L. Prince, J. H. SiewerdsenProceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE 2011 · DOI

TL;DR

This paper introduces a new method for reconstructing medical images from sparse data by leveraging previously acquired images, even if they are not perfectly aligned. The technique simultaneously refines the image and its alignment, leading to significantly improved image quality with reduced data acquisition, which is crucial for lowering patient radiation dose and scan times.

What Problem Does This Paper Address?

Medical imaging techniques like CT scans often require repeated acquisitions for monitoring patient progress, such as tracking lung nodules. These repeated scans can lead to increased radiation exposure and longer patient discomfort. While using prior images as a "prior" in reconstruction can help reduce the amount of data needed for new scans (e.g., by using fewer angles or shorter scan times), this approach is highly sensitive to how well the prior image is aligned with the current scan. Misalignment can introduce artifacts and degrade image quality. Existing methods either assume perfect registration or use separate registration steps, which can be suboptimal. This paper tackles the challenge of effectively using prior images for sparse data reconstruction when the prior and current images might be misregistered, aiming to improve image quality while minimizing data acquisition.

Methods

The authors propose a Penalized-Likelihood Reconstruction (PLR) framework. Their core innovation is a Prior Image with Registration, Penalized Likelihood Estimation (PIR-PLE) objective function. This function jointly optimizes two key aspects:

  • Image Reconstruction ($\mu$): This part uses a likelihood term that accounts for the statistical properties of the tomographic measurements (e.g., Poisson noise), allowing for space-variant noise levels (heteroscedasticity). This is a departure from some prior methods that assume simpler noise models.
  • Registration ($\delta$): This part incorporates a transformation ($W$) and its parameters ($\delta$) to align the prior image ($\mu_\tau$) with the current reconstruction.

The PIR-PLE objective function includes:

  1. A likelihood term that fits the measured data.
  2. A prior image penalty term that penalizes differences between the current reconstruction and the transformed prior image. This term uses compressed sensing (p-norm) penalties to allow for differences, especially in regions of change.
  3. A regularization penalty term to control image smoothness or noise.

Crucially, the registration parameters ($\delta$) are jointly optimized with the image reconstruction. This is achieved through an alternating minimization algorithm (illustrated in Figure 1), where the image and registration parameters are updated iteratively. To handle the non-convexity introduced by p-norms (especially for $p \le 1$), they employ a graduated nonconvexity approach, starting with $p > 1$ and gradually reducing it. For registration, they use a kernel-based interpolator with analytical gradients, implemented on GPUs for efficiency.

Datasets

The paper evaluates their method on two types of datasets:

  • Simulated Digital Phantom Data: A 2D digital phantom was used to test various acquisition scenarios, including:
  • 360° truncated data (central 25% retained).
  • Angularly undersampled data (360° with 60 angles).
  • Limited-angle data (90°). All data were simulated with Poisson noise (10⁴ photons per detector element).

  • Real Cone-Beam CT (CBCT) Test Bench Data: Data was collected from a CBCT test bench using an anthropomorphic chest phantom. This phantom was modified to simulate lung nodule surveillance studies, involving an initial scan of a phantom without a nodule and a follow-up scan of the same phantom with a simulated lung nodule and a slight in-plane translation.

Key Insights and Ideas

  • Joint Optimization is Key: The most significant insight is that simultaneously optimizing image reconstruction and registration parameters leads to superior results compared to separate steps or assuming perfect registration. This allows the algorithm to "learn" the best alignment for the prior image.
  • Handling Unregistered Priors: The PIR-PLE method effectively utilizes prior images even when they are not perfectly aligned, significantly improving image quality in sparse data scenarios.
  • Statistical Measurement Modeling: Incorporating heteroscedastic Poisson statistics in the likelihood term allows for more accurate reconstruction, especially when noise levels vary across measurements.
  • Compressed Sensing for Priors: Using p-norm penalties for the prior image difference allows for flexibility, enabling the reconstruction to deviate from the prior where necessary (e.g., in regions of change like a growing nodule).
  • Graduated Nonconvexity for Stability: This technique is essential for achieving stable and accurate reconstructions when using p-norms for regularization, especially for $p < 1$.

Key Results

  • Improved Image Quality: PIR-PLE reconstructions consistently showed significantly improved image quality compared to standard Penalized Likelihood Estimation (PLE) and Penalized Likelihood Estimation with an unregistered prior (PI-PLE), especially in challenging sparse data scenarios (truncated, undersampled, limited angle). This was evident in the clearer visualization of fine details like line pairs.
  • Accurate Registration: The method accurately estimated in-plane translations, with estimates being within a small fraction of a voxel. This was demonstrated in both simulated and real data.
  • Effective Use of Priors: Even with intentionally mismatched priors, the PIR-PLE method demonstrated predictable behavior, borrowing from the prior where data was missing and relying on measurements otherwise, leading to a more complete reconstruction.
  • Comparison to Fully Sampled Data: In the real CBCT test bench data, the PIR-PLE reconstruction from sparse data showed excellent agreement with a reconstruction from fully sampled data, highlighting the method's effectiveness.

Limitations and Future Work

  • Nonconvexity and Local Minima: While the graduated nonconvexity approach helps, the optimization for registration parameters ($\delta$) can still be susceptible to local minima, especially for complex transformations or poor initial estimates. The authors note that a good starting estimate is often beneficial.
  • Computational Cost: The joint optimization and GPU implementation suggest a significant computational load, although it is likely more efficient than separate, iterative registration and reconstruction steps.
  • Generalizability of Registration: The current implementation focuses on rigid transformations. Future work could explore more complex, non-rigid registration models to handle more general patient motion or anatomical changes.
  • Parameter Tuning: The regularization parameters ($\beta_1, \beta_2$) and p-norm values require careful selection to balance the influence of the prior and data fidelity. Further research could explore automated or adaptive parameter selection strategies.

Summary updated

Abstract

This paper introduces a general reconstruction technique for using unregistered prior images within model-based penalized- likelihood reconstruction. The resulting estimator is implicitly defined as the maximizer of an objective composed of a likelihood term that enforces a fit to data measurements and that incorporates the heteroscedastic statistics of the tomographic problem; and a penalty term that penalizes differences from prior image. Compressed sensing (p-norm) penalties are used to allow for differences between the reconstruction and the prior. Moreover, the penalty is parameterized with registration terms that are jointly optimized as part of the reconstruction to allow for mismatched images. We apply this novel approach to synthetic data using a digital phantom as well as tomographic data derived from a conebeam CT test bench. The test bench data includes sparse data acquisitions of a custom modifiable anthropomorphic lung phantom that can simulate lung nodule surveillance. Sparse reconstructions using this approach demonstrate the simultaneous incorporation of prior imagery and the necessary registration to utilize those priors.
BibTeX
@inproceedings{stayman2011penalized,
  author = {J. Webster Stayman and Wojciech Zbijewski and Yoshito Otake and Ali Uneri and Sebastian Schäfer and J. Lee and Jerry L. Prince and J. H. Siewerdsen},
  title = {{Penalized-likelihood reconstruction for sparse data acquisitions with unregistered prior images and compressed sensing penalties}},
  booktitle = {Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE},
  year = {2011},
  doi = {10.1117/12.878075},
}
Figure