Ultra-low-dose CT for malignant metastasis screening using a deep learning image reconstruction algorithm.

PURPOSE: To evaluate the feasibility and diagnostic performance of ultra-low-dose CT (ULD-CT) for screening malignant metastasis using super-resolution deep learning reconstruction (SR-DLR) compared with normal-resolution DLR (NR-DLR) and hybrid iterative reconstruction (HIR). MATERIALS AND METHODS: Patients with cancer undergoing surveillance were enrolled prospectively and underwent contrast-enhanced whole-body ULD-CT. The images were reconstructed using HIR, NR-DLR, and SR-DLR. Two radiologists independently evaluated image quality, lesion detectability, and diagnostic performance. Radiation dose metrics were compared with prior standard-dose CT. RESULTS: Of the 271 patients (mean age 66.3 ± 12.2 years; 134 men), 56 (20.7%) had disease requiring therapeutic intervention. The ULD-CT images reduced the volume CT dose index by 71.1%, dose-length product by 70.2%, effective dose by 70.6%, and size-specific dose estimate by 70.9%. SR-DLR yielded significant improvement in the overall image quality, sharpness, and image noise reduction compared with HIR and NR-DLR (p < 0.001). The mean attenuation values measured in all organs did not differ significantly among the three reconstruction algorithms (p > 0.05). SR-DLR significantly reduced quantitative image noise by 50% compared with HIR (p ≤ 0.004) and significantly improved SNR across all organs (p < 0.001), without altering attenuation values. SR-DLR demonstrated superior detection rates for malignant lesions for 17 primary cancers and 178 metastatic lesions compared with HIR and NR-DLR. For benign lesions, SR-DLR achieved a nearly 100% detection rate across 9 lesion types (n = 1,264). CONCLUSIONS: SR-DLR enables approximately 70% radiation dose reduction, while maintaining superior image quality and high diagnostic performance for detection of visceral and soft-tissue metastatic disease in oncological patients undergoing cancer surveillance.