Structural Uncertainty Quantification in Low-Dose CT Image Reconstruction

This work presents an evaluation method for assessing the significance of a structure in a reconstructed image for low-dose CT. Due to severe noise, the reconstructed images often contain artifacts. It is important to distinguish real structures from such artifacts. For this purpose, quantifying structural uncertainty is required. One way to evaluate the structural uncertainty is to estimate the covariance matrix of the pixel values. However, it is computationally intractable especially for high-dimensional images, i.e., large-format images. Previous studies suggested the use of the highest posterior density (HPD) to define a credible region. Using the credible region, significance of the structure in the reconstructed images can be evaluated through hypothesis testing: If images without the structure fall outside the credible region, the hypothesis that the true image does not contain the structure is rejected, confirming its significance. The present work expands this method and adapts it more appropriately for low-dose CT. To determine the threshold with which the significance of the structure is evaluated, the chi-square distribution was adopted. In addition, the chi-square was calculated using only the area related to the structure under scrutiny. These modifications provide a more stringent threshold, allowing for more appropriate evaluation of image structures. The feasibility of this method was evaluated and confirmed through image reconstruction simulations.