A material decomposition method for dual-energy CT via dual interactive Wasserstein generative adversarial networks

Dual-energy computed tomography (DECT) enables material-specific imaging but suffers from noise and beam-hardening artifacts. Prior research has shown that conventional methods struggle with preserving image edges and suppressing noise. While various material decomposition techniques have been proposed, none fully resolve the issue of decomposed image quality. This gap motivated the development of data-driven approaches to enhance DECT performance. Existing methods often fail to model spatial and spectral correlations effectively. Additionally, no prior work had resolved how to integrate edge-preserving strategies into deep learning models for DECT. The challenge lies in balancing noise reduction with accurate material separation. This study addresses these limitations through a novel generative adversarial network framework.

This study aims to improve the accuracy of DECT material decomposition while preserving image edges. The specific problem is the poor quality of decomposed images, particularly at edges, caused by noise and beam-hardening artifacts. The motivation is to develop a data-driven solution that outperforms existing methods. The authors propose using dual interactive Wasserstein generative adversarial networks (DIWGAN) to model spatial and spectral correlations. This approach is designed to ensure material specificity and enhance decomposition accuracy. The study also seeks to incorporate edge-preserving strategies into the network architecture. The goal is to train a model that can generate high-quality, noise-suppressed material-specific images. The ultimate aim is to validate this method using both simulated and real-world data.

The study employs a dual interactive Wasserstein generative adversarial network (DIWGAN) for material decomposition in DECT. Two generators are trained to synthesize images of two basis materials from input DECT data. Each generator processes images from a specific energy bin, ensuring material specificity. Discriminators are used to distinguish generated images from ground-truth labels. A hybrid loss function combines L1 loss, edge loss, and adversarial loss to preserve textures and edges. Feature sharing between generators allows for cross-energy bin information exploitation. A selector mechanism is introduced to alternate training between generators in each iteration. The model is trained on digital phantom, XCAT phantom, and mouse data to evaluate decomposition accuracy.

The DIWGAN method achieved superior material decomposition accuracy compared to existing methods. On digital phantoms, bone and soft-tissue regions were separated with an error of less than 3 mg/ml. XCAT phantom results showed that DIWGAN outperformed direct matrix inversion and iterative decomposition in noise and artifact suppression. Compared to Butterfly-Net, DIWGAN reduced RMSE by 0.01 g/ml in soft-tissue images. PSNR and SSIM values reached 31.43 dB and 0.9987, respectively, indicating high image quality. Noise standard deviation decreased by 69% compared to direct matrix inversion. The method also demonstrated improved performance over FCN and Butterfly-Net by 60% and 33%, respectively. Mouse data confirmed the method's potential in real-world applications. These results suggest that DIWGAN effectively suppresses noise and beam-hardening artifacts.

The proposed DIWGAN method improves DECT material decomposition accuracy and edge preservation. The authors state that the method outperforms existing techniques in suppressing noise and beam-hardening artifacts. The use of dual generators and a hybrid loss function is suggested to enhance decomposition quality. The results from digital and XCAT phantoms indicate that DIWGAN achieves material separation with high accuracy. The authors propose that the method's success is due to the integration of spatial and spectral correlations. The performance on mouse data suggests potential for real-world clinical applications. The study concludes that DIWGAN is a promising approach for material decomposition in DECT. The authors suggest that further validation with additional datasets may confirm these findings.