Effects of image denoising on quantitative material decomposition in photon-counting spectral computed tomography

Abstract

Material decomposition algorithms enable discrimination and quantification of multiple contrast agent and tissue compositions in spectral image datasets acquired by photon-counting computed tomography (PCCT). Image denoising has been shown to improve PCCT image reconstruction quality and feature recognition while preserving fine image detail. Reduction of image artifacts and noise could also improve the accuracy of material decomposition but the effects of denoising on material decomposition have not been investigated. In particular, deep learning methods can reduce inherent PCCT image noise without using a system-based or assumed prior noise model. Therefore, the objective of this study was to investigate the effects of image denoising on quantitative material decomposition in the absence of an influence of spatial resolution on feature recognition. Phantoms comprising multiple pure and spatially uniform contrast agent (gadolinium, iodine) and tissue (calcium, water) compositions were imaged by PCCT with four energy thresholds chosen to normalize photon counts and leverage contrast agent k-edges. Image denoising was performed by the established blockmatching and 3D-filtering (BM3D) algorithm or deep learning using convolutional neural networks. Material decomposition was performed on as-acquired, BM3D-denoised, and deep-learning-denoised datasets using constrained maximum likelihood estimation and compared to known material concentrations in the phantom. Image denoising by BM3D and deep learning improved the quantitative accuracy of material concentrations determined by material decomposition compared to ground truth, as measured by the root-mean-squared error. Material classification was not improved by image denoising compared with as-acquired images, suggesting that material decomposition was robust against inherent acquisition noise when feature recognition was not challenged by the system spatial resolution. Deeplearning-denoised images balanced preservation of local detail compared to more aggressive smoothing with BM3D, as measured by line profiles across features.