DSPAN: Dual-domain structure-aware adversarial learning with physics-inspired consistency for low-dose CT denoising

Abstract Background Low-dose computed tomography (LDCT) reduces radiation exposure but increases noise and artifacts that can obscure low-contrast anatomy. We developed the Dual-domain Structure-Preserving Adversarial Network (DSPAN) to improve LDCT image fidelity. Methods DSPAN combines a ResNet-based generator with dual-domain structure-aware attention and coupled gradient- and frequency-consistency learning. We used five-fold patient-level cross-validation on all 10 patients in the 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge dataset. Each fold included seven training patients, one validation patient for checkpoint selection, and two test patients. Every patient was tested once. Slice-level measurements were averaged within each patient, and performance was summarized descriptively across the 10 patients. No null-hypothesis tests were performed. Results DSPAN achieved a mean peak signal-to-noise ratio of 40.2105 ± 1.6183 dB, structural similarity of 0.9424 ± 0.0220, and root mean squared error of 2.5359 ± 0.5303 in 8-bit-equivalent units. It had the best mean value for all six evaluated image-quality metrics. Compared with DU-GAN, mean peak signal-to-noise ratio increased by 0.1505 dB and root mean squared error decreased by 0.0410. Module ablation showed a 1.0662 dB mean peak signal-to-noise ratio gain over ResNetGAN alone. Conclusions Combining structure-aware attention with gradient and frequency consistency improved internal image-quality performance in this 10-patient dataset. External validation and diagnostic reader studies are needed to establish generalizability and clinical relevance.