Uncertainty-Aware Material Identification In Spectral Photon-Counting CT: A Gold Nanoparticle Case Study
Abstract
Summary
*]:pointer-events-auto [content-visibility:auto] supports-[content-visibility:auto]:[contain-intrinsic-size:auto_100lvh] scroll-mt-[calc(var(--header-height)+min(200px,max(70px,20svh)))]" dir="auto" data-turn-id="d0e0b41c-b4d9-459a-bc0a-7681e303073e" data-testid="conversation-turn-6" data-scroll-anchor="true" data-turn="assistant" tabindex="-1">
Purpose
Spectral photon-counting computed tomography (SPCCT) enables material-specific characterization but is limited by increased image noise and reduced photon statistics, particularly for small or low-contrast structures. The purpose of this work is to develop an automated, effect-size- and uncertainty-aware material identification framework that uses calibration-derived priors to determine Gaussian kernel sizes and intensity-based decision boundaries for different materials and concentrations.
Methods
During a calibration phase, multi-energy images of a reference phantom containing known materials at known concentrations are acquired. For each energy bin, the mean attenuation, variance, and standardized effect size between neighboring materials in attenuation space are estimated. These statistics used to design energy- and material-specific Gaussian kernels whose spatial support is an explicit function of material separability and measurement uncertainty. Using the same calibration information, the framework derives intensity-based decision boundaries between adjacent materials by modeling their attenuation distributions and incorporating both effect size and variance into the classification rule.
Results
By construction, the calibration-derived statistics yield a family of energy- and material-specific Gaussian kernels that automatically adjust the degree of spatial grouping to the local contrast-to-noise conditions. Materials or concentrations with small effect sizes or high uncertainty are associated with larger kernels, whereas well-separated materials retain smaller kernels that preserve spatial resolution. Similarly, the intensity-based decision boundaries between neighboring materials are shifted according to their estimated means, variances, and effect sizes, providing classification thresholds that explicitly account for finite material separation and measurement noise rather than relying on fixed, heuristic thresholds.
Conclusion
The proposed uncertainty-aware framework provides a principled, calibration-driven approach to material identification in photon-counting and other multi-energy CT systems. By jointly adapting spatial grouping and decision boundaries to material separability and measurement uncertainty, it is expected to improve the robustness and reliability of identifying small or low-contrast features , concentration ranges, and energy bin configurations.