Spectral CT reconstruction with an explicit photon-counting detector model: a one-step approach

Pierre‐Antoine Rodesch, V. Rebuffel, Clarisse Fournier, Florence Forbes, L. VergerMedical 2018 · DOI

TL;DR

This paper introduces a novel "one-step" approach for spectral CT reconstruction that simultaneously accounts for the object's material composition and the detector's spectral response. This method significantly reduces artifacts and improves material decomposition accuracy compared to traditional "two-step" methods, paving the way for more quantitative and reliable spectral CT imaging.

What Problem Does This Paper Address?

Traditional CT imaging provides a single image representing the overall attenuation of X-rays through an object. Spectral CT, on the other hand, aims to differentiate materials based on their unique X-ray attenuation properties at different energy levels. This is achieved using photon-counting detectors (PCDs) that can distinguish the energy of incoming photons. However, PCDs introduce their own complexities, such as charge-sharing, fluorescence escape, and pile-up effects, which distort the measured X-ray spectrum.

Reconstructing spectral information from PCD data typically involves two main challenges:

  1. Geometric inversion: Reconstructing the spatial distribution of attenuation coefficients from the measured X-ray projections (similar to standard CT).
  2. Detector spectral degradation: Correcting for the distortions introduced by the PCD's response to the X-ray spectrum.

Existing methods often tackle these problems sequentially in a "two-step" approach: first, they attempt to correct for detector distortions, and then they perform spectral reconstruction. This can lead to error propagation and the introduction of artifacts. This paper addresses the need for a more integrated approach that can simultaneously handle both the geometric inversion and the detector's spectral response, aiming for more accurate and artifact-free material decomposition.

Methods

The core innovation of this paper is a "one-step" reconstruction method that integrates the detector's spectral response directly into the reconstruction process. Key aspects of their method include:

  • Explicit Detector Model: They model the detector's spectral response using a Detector Response Matrix (DRM). This matrix quantifies the probability that a photon of a given energy entering the detector is registered in a specific energy bin.
  • Forward Model: The forward model describes how the true material distribution and the detector response combine to produce the measured photon counts in each energy bin for each projection ray. This model is expressed as: $m_{i,c} = \sum_{E} \sum_{j} \sum_{m} DRM(c; E) N^0(E) e^{-\sum_{m} \mu_m(E) A_{ij} f_{mj}} \Phi_m(E) A_{ij}$ where $m_{i,c}$ is the measured photon number in bin $c$ for ray $i$, $DRM(c; E)$ is the detector response, $N^0(E)$ is the raw flux, $\mu_m(E)$ is the attenuation coefficient of basis material $m$, $A_{ij}$ is the projection matrix, and $f_{mj}$ is the basis material map for material $m$ at voxel $j$.

  • Basis Material Decomposition: The attenuation coefficient of the object is decomposed into a linear combination of two known basis material attenuation coefficients. The goal is to reconstruct maps of these two basis materials.

  • Poisson Likelihood: The measured photon counts are assumed to follow a Poisson distribution, which is a standard assumption for photon counting statistics.
  • Spatial Penalization: To improve reconstruction quality and reduce artifacts, a spatial regularization term is applied to the material basis maps. They use a Tukey function for robust penalization, which is less sensitive to outliers than L1 or L2 penalties and helps preserve edges while smoothing.
  • Bayesian Framework: The reconstruction problem is formulated within a Bayesian framework, seeking to maximize the posterior probability of the material maps given the measurements and prior information. This leads to the optimization problem: $f^{\{m\}} = \arg \max \{PL(m; \hat{m}(f^{\{m\}}) + \beta R(f^{\{m\}})\}$ where $PL$ is the Poisson log-likelihood and $R$ is the spatial regularization term.

  • ML-TR Algorithm: They adapt the Maximum Likelihood-Transmission (ML-TR) algorithm to solve this optimization problem.

  • Initialization: For faster convergence, the reconstruction is initialized using a standard polynomial "two-step" method.

Datasets

The authors tested their method on simulated slice-CT data. They used the Sindbad-SFFD software to generate measurements for a phantom consisting of a 30 cm diameter cylinder with six 5 cm diameter inserts made of various materials (cortical bone, compact bone, aluminum, adipose tissue, polyethylene, and Plexiglas). The X-ray source was set to 140 kV with 0.1 cm aluminum filtering, and Poisson noise was added to the simulated data. The spectral convolution was performed using a realistic DRM. Two measurement bins were used: [15 keV - 65 keV] and [66 keV - 115 keV].

Key Insights and Ideas

  • "One-step" is better than "two-step": The simultaneous inversion of the geometric projection and detector response is crucial for reducing artifacts and improving the accuracy of material decomposition. Separating these steps can lead to error propagation.
  • DRM is key: Explicitly modeling the detector's spectral response with a DRM is essential for accurate spectral CT reconstruction with PCDs.
  • Robust penalization matters: The use of a Tukey function for spatial regularization provides smooth reconstructions while effectively preserving edges and correcting artifacts that plague simpler methods.
  • Quantitative accuracy: The "one-step" method achieves significantly higher quantitative accuracy in material density estimation compared to the "two-step" approach, with errors around 1% versus 5-10%.
  • Improved material differentiation: The method can better distinguish between softer inserts in scatter plots, indicating a more robust material decomposition.

Key Results

  • The "one-step" method successfully reconstructs two basis material maps (water and cortical bone) from simulated spectral CT data.
  • Qualitative results show that the "one-step" method significantly reduces streaking artifacts observed in the "two-step" method, particularly in the more attenuating inserts and the center of the object.
  • Quantitative analysis (Table 1) shows that the "one-step" method provides mean values for material basis maps that are much closer to the true values for various materials compared to the "two-step" method. For example, for water, the "one-step" method yields a water map value of 1.00, while the "two-step" method gives 0.97. For cortical bone, the "one-step" method achieves 0.99, versus 0.91 for the "two-step" method.
  • Figure 3 demonstrates that the "one-step" method provides better separation and representation of different material inserts in a Cortical Bone/Water density map, with accuracy close to the ground truth.

Limitations and Future Work

  • Parameterization of regularization: The authors acknowledge that the efficiency of the spatial prior term depends on the specific case and application. Further experiments are needed to optimize this parameterization.
  • Phantom complexity: The current study used a relatively simple phantom. Future work will investigate adapting the prior to more complex geometries and materials, especially those found in medical applications (e.g., soft tissues).
  • Detector non-linearity: The current model assumes linear detector behavior. Future work will explore implementing non-spatial priors based on tissue-type identification or correlations between viewing angles, and will also investigate the impact of detector non-linearity due to pile-up effects.
  • DRM confidence: The confidence in the accuracy of the DRM itself is an area for further study.
  • Optimal bin number: Determining the optimal number of energy bins for reconstruction is another aspect for future investigation.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{rodesch2018spectral,
  author = {Pierre‐Antoine Rodesch and V. Rebuffel and Clarisse Fournier and Florence Forbes and L. Verger},
  title = {{Spectral CT reconstruction with an explicit photon-counting detector model: a one-step approach}},
  booktitle = {Medical},
  year = {2018},
  doi = {10.1117/12.2285792},
}
Figure