Abstract
Tomographic reconstructions are often segmented to quantify structure parameters. However, to our knowledge, no efforts have yet been made to visualize the accuracy of the segmentation result. This paper introduces a way to visualize the segmentation error, based on the residual projection error, which is the difference between the recorded data and the forward projection of the segmented tomogram. From the residual projection error, a segmentation error tomogram is reconstructed. This error tomogram allows to detect errors in the gray levels of the segmented tomogram, or to discriminate between reconstruction artifacts and actual features of the scanned object. The proposed technique is independent of the algorithms that were used to create and segment the tomogram.
Introduction
In computed tomography (CT), segmentation is a critical step for classifying materials and structures within a reconstructed volume. However, standard segmentation tools typically operate in isolation from the original raw projection data (sinograms). This disconnect can lead to inaccuracies, as reconstruction artifacts—such as streaks or noise—may be misinterpreted as physical features. Currently, there is no established standard for visualizing these errors or validating the segmented output against the physical "ground truth" represented by the original measurements. This research addresses this gap by proposing a method to visualize segmentation errors using the residual projection error.
Methodology
The proposed workflow bridges the gap between the segmented image and the raw data through a four-step process:
- Forward Projection: The segmented image ($s$) is used to simulate a synthetic X-ray dataset ($p_s$) via forward projection.
- Residual Calculation: The synthetic dataset is subtracted from the original recorded projection data ($p$), yielding the residual projection error ($e = p - p_s$).
- Error Reconstruction: A standard reconstruction algorithm (e.g., SIRT) is applied to this residual error to create an "error tomogram" ($\hat{y}$).
- Analysis: The error tomogram serves as a spatial map of segmentation inaccuracies. A perfect segmentation results in a near-zero error tomogram, while structured features in the map indicate missed edges or incorrect gray-level assignments.
Experimental Setup
The method was validated using three synthetic phantoms designed to test various conditions:
- Phantom 1: A binary phantom for ideal, noise-free conditions.
- Phantom 2: A three-level phantom tested with limited projection angles and Poisson noise.
- Phantom 3: A phantom with continuous grayscale gradients to evaluate performance on non-homogeneous regions.
Reconstructions utilized Filtered Backprojection (FBP) and the Simultaneous Iterative Reconstruction Technique (SIRT), with segmentation performed via Otsu’s method.
Results and Discussion
The error tomogram provides a powerful tool for both qualitative and quantitative validation. It effectively discriminates between mathematical reconstruction artifacts and actual physical features missed during segmentation. For instance, streaks present in an FBP reconstruction that do not appear in the error tomogram are confirmed as artifacts.
Quantitatively, the method allows for gray-level correction. By averaging values within a region of the error tomogram, researchers can adjust the segmented class values to match the true physical density. In tests with Phantom 2, corrected gray levels were accurate to within 0.1% to 0.4% of the ground truth, even in the presence of noise. Furthermore, the error tomogram acts as a homogeneity detector; internal density variations in objects segmented as a single material appear as structured residuals.
Conclusion
This technique offers a robust, algorithm-independent way to "close the loop" in CT analysis by re-incorporating raw projection data. While highly effective for homogeneous materials, future work is needed to handle continuous gradients more accurately. Additionally, the requirement for raw data and the computational overhead of an extra reconstruction cycle are practical considerations for implementation. The authors suggest that this approach could eventually be integrated into iterative segmentation refinement loops.
Summary updated