Image Quality Evaluation In Deep-Learning-Based CT Noise Reduction Using Virtual Imaging Trial Methods: Lesion Characterization
Abstract
Purpose
Deep learning-based image reconstruction and noise reduction (DLR) techniques are increasingly adopted in clinical CT to improve image quality at reduced radiation doses. While prior studies have demonstrated benefits for lesion detection, DLR performance in challenging lesion characterization tasks (differentiating malignant from benign lesions) remains unexplored across varying lesion contrasts, dose levels, and denoising strengths. This study applies a patient-data-based virtual imaging trial (VIT) framework to objectively evaluate DLR performance for lesion characterization.
Methods
Based on a patient-data-based VIT framework we developed previously, a lesion characterization task was designed by inserting cylindrical lesions (15 mm diameter, 5 mm height) with or without spiculated boundaries into real patient liver CT projection data to simulate benign and malignant lesions. Ensembles were generated at four contrasts (-20, -30, -50, -100 HU) and three dose levels (12.5%, 25%, 50% of routine dose). Images were reconstructed with filtered backprojection (FBP), iterative reconstruction (IR), and an in-house DLR at weak, medium, and strong denoising strengths. Channelized Hotelling observer (CHO) index of detectability (d') was computed for classifying spiculated vs. round lesions using 3,000 image pairs (5 image slices × 600 noise realizations) per condition.
Results
For lesion characterization, d’ decreased with greater denoising strength at lower contrasts (-20 to -50 HU). DLR-Strong reduced d’ relative to both FBP and IR at lower contrasts (-20 to -50 HU) across all doses and at 12.5% dose for -100 HU contrast, but improved d’ at -100 HU with 25–50% doses. In contrast, for lesion detection, DLR consistently improves d’ over FBP, with gains rising as denoising strength increased.
Conclusion
Patient-data-based virtual imaging trials with CHO evaluation demonstrated that DLR can degrade lesion characterization performance at low contrasts and very low doses, despite improvements in lesion detection, warranting caution when applying strong DLR under such challenging conditions.