Localizing DART using the reconstructed residual error

Tom Roelandts, Kees Joost Batenburg, Jan SijbersCentrum 2013

Abstract

The reconstructed residual error was recently introduced as a localized measure of the accuracy of a segmented tomogram. In this paper, we use the reconstructed residual error during the reconstruction, and adapt the recently proposed Discrete Algebraic Reconstruction Technique (DART) to only enforce discreteness in regions where this is appropriate. Simulation experiments show that the reconstructed residual error can detect non-homogeneous regions. Based on this observation, the DART algorithm is modified to allow the reconstruction of objects that are only partially discrete. The experiments show that the reconstruction quality of both homogeneous and non-homogeneous regions is improved.

Introduction

In computed tomography (CT), Discrete Tomography (DT) leverages prior knowledge about an object's composition—specifically that it consists of only a few uniform materials—to produce high-quality reconstructions from a limited number of projections. The Discrete Algebraic Reconstruction Technique (DART) is a leading algorithm in this field. However, DART assumes global discreteness. When an object contains non-homogeneous regions, such as density gradients or foreign inclusions, DART attempts to force these pixels into predefined discrete categories, resulting in significant artifacts. This paper introduces a localized version of DART that identifies and relaxes constraints on regions that do not fit the discrete profile.

Methods

The proposed Localized DART algorithm integrates a feedback loop based on the reconstructed residual error to dynamically identify non-homogeneous regions.

The Localized DART algorithm incorporating the reconstructed residual error to update the set of free pixels M.

As shown in the algorithm above, the process begins with a standard SIRT reconstruction followed by segmentation. The core innovation lies in the calculation of the residual error: the segmented image is "re-projected" to simulate X-ray data, which is then compared against actual measurements. This error is reconstructed back into the image space ($e^t$). Pixels exhibiting a high error (exceeding a threshold $c$) are added to a set of "free pixels" ($M$). In subsequent iterations, pixels in $M$ are allowed to vary continuously, while the rest of the image remains subject to discrete constraints. This allows the algorithm to "discover" non-homogeneous regions on the fly.

Results and Discussion

The method was validated using a synthetic $512 \times 512$ pixel phantom containing complex shapes and a non-homogeneous ellipse with a smooth gradient. Using only 20 to 40 projections, the localized DART successfully identified the non-homogeneous region.

Reconstruction iterations showing segmented images (a, c) and their corresponding reconstructed residual error maps (b, d). Note how the error localizes around the non-homogeneous ellipse.

As illustrated in the figure above, the reconstructed residual error map ($e^t$) provides a clear spatial diagnostic. High error values (red/blue regions in panels b and d) correlate precisely with the non-homogeneous ellipse that standard DART fails to segment correctly. By freeing these pixels, the algorithm achieves a clean, high-contrast reconstruction of the uniform regions while accurately capturing the gradient of the ellipse. This hybrid approach outperforms both standard SIRT (which is noisy at low projection counts) and standard DART (which produces artifacts in non-uniform areas).

Conclusion

By using the reconstruction error as a real-time map, Localized DART provides a robust solution for objects that are only partially discrete. While the algorithm requires careful selection of the error threshold $c$ and involves additional computational overhead, it significantly improves data efficiency and reconstruction quality for complex, real-world samples. Future work will focus on automating threshold selection and testing against experimental data with physical noise models.

Summary updated

Abstract

The reconstructed residual error was recently introduced as a localized measure of the accuracy of a segmented tomogram. In this paper, we use the reconstructed residual error during the reconstruction, and adapt the recently proposed Discrete Algebraic Reconstruction Technique (DART) to only enforce discreteness in regions where this is appropriate. Simulation experiments show that the reconstructed residual error can detect non-homogeneous regions. Based on this observation, the DART algorithm is modified to allow the reconstruction of objects that are only partially discrete. The experiments show that the reconstruction quality of both homogeneous and non-homogeneous regions is improved.
BibTeX
@inproceedings{roelandts2013localizing,
  author = {Tom Roelandts and Kees Joost Batenburg and Jan Sijbers},
  title = {{Localizing DART using the reconstructed residual error}},
  booktitle = {Centrum},
  year = {2013},
}
Figure