U-STAR: Unsupervised Structure-Texture Adaptive Refinement Framework for Material Decomposition

Material decomposition enables spectral CT to distinguish and quantify different materials, serving as the cornerstone of several clinically important applications. However, the material decomposition process inherently amplifies noise, severely degrading the quality of material basis images. Although deep learning methods have shown promising results, most of them rely on high-quality paired training data, which limits practical clinical deployment. To address this challenge, we propose U-STAR, a two-stage unsupervised structure-texture adaptive refinement framework for image-domain material decomposition. In the first stage, we design a structural prior-guided network that exploits low-noise structural priors derived from basis transformation as reliable guidance, thereby facilitating stable unsupervised denoising. To mitigate texture degradation, we further employ adversarial learning with a gradient consistency loss to restore realistic textures and fine anatomical details. In the second stage, we introduce a test-time adaptation strategy to improve generalization. This self-supervised mechanism further suppresses noise by enforcing output consistency under positive and negative noise perturbations, enabling zero-shot, instance-specific adaptation. Simultaneously, material decomposition constraints are incorporated during fine-tuning to maintain physical fidelity. Experiments on clinical dual-energy CT (DECT) data and simulated photon-counting CT (PCCT) data demonstrate the superior performance of the proposed method in noise suppression and structure preservation, as well as its strong robustness, highlighting its potential for clinical material decomposition applications.