A study on noise reduction for dual-energy CT material decomposition with autoencoder

Abstract

A major challenge for the material decomposition task of the dual-energy computed tomography (DECT) is the algorithm often suffers from heavy noise in the results. The purpose of this study is to propose a scheme to increase the noise performance of material decomposition. The scheme we propose in this paper is to apply an autoencoder-based denoising procedure to the photon-counting DECT images before they are fed into the material decomposition algorithm. We implement the autoencoder (AE) by stacking a series of convolutional and deconvolutional layers. The decomposition technique adopted in our work is an iterative method using least squares estimation with the Huber loss function. The noises of the input and the output of material decomposition are analyzed with both simulated data and real data. Phantom and chicken wing experiments are conducted with a photon-counting-based spectral CT scanner to evaluate the proposed material decomposition scheme. The noise analysis of the input and the output of material decomposition demonstrates a positive correlation between them. Comparative experiment indicates a noise reduction in the output density maps for 26.07% to 35.65% after the autoencoder pre-processing is applied. The resultant contrast-to-noise ratio is largely increased, correspondingly. By utilizing the additional autoencoder denoising step, the material decomposition algorithm achieves an improvement in the noise performance of the resultant density maps.