Image-Domain GAN Denoising for Sn100 kVp Ultra-Low-Dose Chest CT: A Retrospective Paired Image-Quality Study

Background: Ultra-low-dose (ULD) chest CT can reduce radiation exposure but may compromise image quality. We evaluated image-domain generative adversarial network (GAN)-based denoising at low-dose (LD) and ULD levels, focusing on ULD-AiR versus LD-ADMIRE S3. Methods: In this single-center retrospective paired study, 262 participants underwent LD and ULD chest CT on the same scanner. Images reconstructed with Advanced Modeled Iterative Reconstruction at strength 3 (ADMIRE S3) were post-processed with AiR Denoising v1.0, yielding four series. Objective metrics, including signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), and 5-point subjective ratings were compared within dose levels and between ULD-AiR and LD-ADMIRE S3. Exploratory regression assessed associations of ULD-to-LD SNR log-ratios with the volume CT dose index (CTDI_vol) log-ratio and anthropometric variables. Results: The median paired reduction in estimated effective dose from LD to ULD CT was 46.5%. Within each dose level, AiR reduced image noise and increased SNR, CNR, and subjective scores. Compared with LD-ADMIRE S3, ULD-AiR showed lower image noise and higher SNR/CNR in the lung, aorta, and muscle; liver findings were less consistent, whereas vertebral metrics were less favorable. Lung-parenchyma scores did not differ significantly, mediastinal soft-tissue scores favored ULD-AiR, and overall image-noise scores favored LD-ADMIRE S3. Exploratory regression identified CTDI_vol log-ratio associations with aortic and muscle SNR log-ratios; some anthropometric associations were sensitive to an influential observation. Conclusions: Image-domain GAN denoising improved several objective and subjective image-quality metrics at both dose levels. With a median paired reduction of 46.5% in estimated effective dose, ULD-AiR showed tissue- and endpoint-specific image-quality differences relative to LD-ADMIRE S3. These findings do not establish diagnostic or screening equivalence.