Dual-state adaptive unrolled network for structure-preserving low-dose CT image restoration.

Abstract

Low-dose computed tomography (LDCT) reduces radiation exposure but introduces spatially heterogeneous noise and artifacts that can obscure subtle anatomical structures. We address post-reconstruction image-domain LDCT restoration, where degradation suppression and structure preservation impose competing demands on a single restoration trajectory. We propose DSAUNet, a Dual-State Adaptive Unrolled Network that maintains persistent anatomy- and degradation-oriented latent states across unrolled stages. Rather than acting as transient feature branches, the evolving states participate in parameterizing subsequent restoration behavior through a degradation-oriented global policy coordinate for stage-level modulation and state-derived confidence variables for spatially selective regulation. Direction-aware prior modeling and multi-scale state guidance further realize this state-parameterized restoration process. Under the reported protocols, DSAUNet achieves 40.3721 dB PSNR and 0.9642 SSIM on 2DeteCT, exceeding the strongest baseline by 1.6320 dB and 0.0108, and achieves 47.4277 dB PSNR and 0.9846 SSIM under the patient-level five-fold Mayo evaluation, with a 1.0920 dB PSNR gain over the strongest PSNR baseline. On LoDoPaB-CT, it further obtains 38.5501 dB PSNR and 0.9141 SSIM, with the best PSNR, SSIM, and VIF among the compared methods. Controlled raw-state, persistent state-swapping, policy-coordinate, and confidence-map analyses support functionally differentiated state roles and active adaptive regulation. Bidirectional 2DeteCT-Mayo zero-shot experiments reveal direction-dependent transfer behavior, while unified profiling characterizes an accuracy-complexity trade-off. Together, these results support persistent dual-state controlled unrolling as a structured approach to image-domain LDCT restoration within the evaluated settings.