Quantitative Evaluation of AI Denoising on Abdominal CT Imaging with Small Hepatic Cysts and Lesions: Noise and Contrast-to-Noise Performance
Abstract
Purpose
To evaluate the performance of a commercial deep-learning denoising algorithm (ClariCT.AI) for noise reduction and contrast-to-noise ratio (CNR) improvement in abdominal CT images with small hepatic cysts and lesions.
Methods
An ACR CT accreditation phantom was scanned on a GE Revolution EVO scanner using a routine abdomen-pelvis protocol (CTDIvol: 12 mGy) and reconstructed with filtered back-projection (FBP) and the vendor's iterative reconstruction (IR) algorithm. ClariCT.AI was subsequently applied offline to the FBP images using default settings. Noise power spectrum (NPS) analysis was performed to compare noise magnitude and texture. Ten abdominal CT cases with small hepatic cysts and lesions acquired using the same clinical protocol (AEC on; CTDIvol: 7.4-26.1mGy) were retrospectively processed with ClariCT.AI and reviewed by a radiologist. Three circular ROIs were placed to encompass the hepatic cyst or lesion, adjacent liver parenchyma, and subcutaneous fat. Lesion contrast was defined as the difference in mean CT numbers between the cyst/lesion and liver parenchyma, while image noise was defined as the standard deviation measured in the subcutaneous fat ROI. The CNR was calculated as the ratio of lesion contrast to image noise and compared across the three methods.
Results
In the phantom study, ClariCT.AI substantially reduced noise magnitude compared with FBP (42%) and vendor IR (28%). Modest alterations in noise texture were observed for FBP+AI, with mean spatial frequencies shifts of 16% and 6% relative to FBP and IR, respectively. In clinical images, ClariCT.AI demonstrated consistent improvements in noise reduction and CNR across all cases. Mean image noise decreased from 16.2±3.1HU (FBP) and 14.6±3.2HU (IR) to 8.9±3.1HU (FBP+AI), with no apparent relationship to patient dose. Mean CNR increased from 5.4±2.1 (FBP) and 5.8±2.0 (IR) to 10.6±5.1 (FBP+AI).
Conclusion
ClariCT.AI achieves substantial and consistent noise reduction and CNR improvement, indicating potential benefits for abdominal CT image quality and lesion conspicuity.