Quantitative Evaluation of AI-Based Denoising (ClariCT.AI): Spatial Resolution and Noise Characteristics
Abstract
Purpose
To evaluate a commercial deep learning denoising algorithm (ClariCT.AI) through quantitative phantom measurements and clinical image assessment.
Methods
An ACR CT accreditation phantom was scanned on a clinical CT scanner (GE Revolution EVO) using a routine abdomen-pelvis protocol radiation at standard radiation dose (CTDIvol: 12 mGy). Additional scans were performed at 50%, 75%, 125%, and 150% of the routine dose level. Images were reconstructed with filtered back-projection (FBP) and the scanner’s iterative reconstruction (IR) algorithm, with ClariCT.AI subsequently applied offline to FBP images using default settings. Quantitative analyses were performed using a web-based CT image quality evaluation platform (CTPro) to compare three approaches (FBP, IR, FBP+ClariCT.AI). Measurements included contrast-dependent modulation transfer function (MTFc) from the CT number accuracy module to assess spatial resolution, and noise magnitude and noise power spectrum (NPS) from the uniformity module to characterize noise properties. Clinical abdominal cases were reviewed by a radiologist to measure liver noise across the methods.
Results
ClariCT.AI substantially reduced noise magnitude across all dose levels compared to FBP (39% - 49% reduction), achieving lower noise than vendor IR (23% - 36% reduction). NPS analysis revealed that noise texture of FBP+ClariCT.AI was only slightly altered compared to FBP and IR, with the mean frequencies of 2.6 cm-1, 3.1 cm-1, and 2.8 cm-1, respectively. Spatial resolution was preserved with no significant MTFc changes across dose levels (50% difference: 0% - 1.4%; 10% difference: 0% - 3.6%). Clinical images demonstrated markedly reduced liver noise (18% - 61%) with ClariCT.AI processing
Conclusion
ClariCT.AI achieves substantial noise reduction (up to 49% in phantom, 61% in clinical images) while preserving spatial resolution and maintaining noise texture characteristics similar to conventional reconstruction methods.