Dose‑Dependent Image Quality Behavior of Two Clinical Deep‑Learning CT Reconstruction Methods
Abstract
Purpose
To characterize how decreasing radiation dose affects noise magnitude, noise texture, and task‑based spatial resolution in deep learning(DL) CT reconstruction algorithms.
Methods
An ACR CT phantom was scanned across decreasing CTDIvol values between 0.7 and 16.4 mGy on two systems. On the Aquilion One Insight(Canon Medical Systems USA), images were reconstructed using PIQE Level 2 DL with Cardiac, Body, and Lung kernels on a 1024² matrix. On the Revolution Apex(GE Healthcare), images were reconstructed using TrueFidelity Medium DL with Standard, Lung, Bone, and BonePlus kernels on a 512² matrix. Images were also reconstructed with filtered back projection(FBP). Noise Power Spectrum (NPS) based metrics were computed, including the Noise Magnitude Ratio (NMR), defined as the square root of the ratio of NPS areas between the DL reconstruction and the FBP reference, and the Central Frequency Ratio (CFR), defined as the ratio of NPS peak frequencies between DL and FBP, reflecting noise‑texture shifts. TTF10% was measured from the bone insert to assess resolution. Microsoft Copilot assisted in analysis of study data and preparation of the abstract.
Results
For the Lung kernel, at a mid‑dose comparison (PIQE 6.2 mGy vs TrueFidelity 6.14 mGy), PIQE/TrueFidelity ratios were 0.234 for NMR, 0.537 for CFR, and 1.449 for TTF10%, indicating lower noise magnitude, lower central‑frequency ratio, and higher task‑based resolution for PIQE. PIQE demonstrated clear dose‑dependent trends for the Lung kernel, described by NMR=0.039*ln(CTDIvol)+0.077 (R² = 0.997), CFR=0.127*ln(CTDIvol)+0.273 (R² = 0.994), and TTF10%=0.152*ln(CTDIvol)+1.066 (R² = 0.934). In contrast, TrueFidelity metrics for the Lung kernel remained approximately constant across the same CTDIvol range. This overall pattern of strong dose dependence for PIQE and minimal dose‑related change for TrueFidelity, was consistent across all kernels evaluated.
Conclusion
PIQE exhibited a strong reduction in noise magnitude and clear dose‑dependent changes in NMR, CFR, and TTF10%, whereas TrueFidelity showed little to no dose variation.