DPD-DEMD: Denoising Prior Guided Weakly Supervised Image Domain Dual-Energy Material Decomposition

Dual-energy material decomposition is widely used in clinical diagnosis, especially for material characterization. However, conventional image-domain methods suffer from noise amplification, thus reducing signal-to-noise ratio and compromising diagnostic accuracy. Although deep learning approaches have shown significant progress, they often require high-quality paired or unpaired labels, limiting their clinical application. To address these issues, this work explores the feasibility of weakly supervised methods and proposes a denoising prior guided weakly supervised learning framework, DPD-DEMD, to achieve high-accuracy image-domain dual-energy material decomposition. DPD-DEMD utilizes pretrained CT denoising models to construct robust priors for our dual energy material decomposition task. Furthermore, we propose an adaptive confidence mask mechanism for pseudo label generation and a multi-prior fusion strategy, thereby substantially improving the stability and reliability of the weakly supervised learning process. In addition, we fully exploit the correlation between dual energy images and further propose global-local regularization loss to improve the material decomposition accuracy. Extensive experiments conducted on both simulated and clinical datasets verify the superior performance and robustness of the proposed method, thereby demonstrating its potential clinical value in material decomposition. Our code is available at https://github.com/zhongxinyun/DPD-DEMD.git.