Back to all papers

Dose‑dependent image quality behavior of recently introduced clinical deep learning CT reconstruction methods.

November 20, 2026pubmed logopapers

Authors

Meier JG,Somasundaram E,Brady SL

Affiliations (2)

  • Department of Radiology, Cincinnati Children's Hospital Medical Center, Cincinnati, Ohio, USA.
  • Department of Radiology, University of Cincinnati College of Medicine, Cincinnati, Ohio, USA.

Abstract

Deep learning reconstruction (DLR) algorithms are increasingly used in CT to reduce image noise and support radiation dose reduction. However, DLR methods may alter spatial resolution and noise texture depending on dose level, reconstruction kernel, matrix size, and vendor implementation. To characterize how decreasing radiation dose affects noise magnitude, noise texture, and task-based spatial resolution for two recently available clinical CT deep learning reconstruction algorithms: Canon Precise IQ Engine (PIQE) super-resolution DLR and GE HealthCare TrueFidelity DLR. The American College of Radiology CT accreditation phantom was scanned on an Aquilion ONE/INSIGHT Edition scanner (Canon Medical Systems USA) and a Revolution Apex scanner (GE HealthCare) across 12 dose levels. CTDIvol ranged from 0.7 to 20.0 mGy for Canon and from 0.68 to 20.46 mGy for GE. Canon images were reconstructed using filtered back projection (FBP) with a 512 matrix and PIQE Levels 1-3 with Body, Lung, and Cardiac kernels using a 1024 matrix. GE images were reconstructed using FBP and TrueFidelity low, medium, and high strengths with Standard, Lung, Bone, and Bone Plus kernels using a 512 matrix. Task-based spatial resolution was evaluated using the task transfer function (TTF). Noise magnitude and texture were evaluated using the noise power spectrum (NPS), including the Noise Magnitude Ratio (NMR), defined as the square root of the ratio of NPS areas between the DLR and the FBP reference, and the Central Frequency Ratio (CFR), defined as the ratio of the center of gravity of the NPS between DLR and FBP, reflecting noise texture shifts. At a high-dose comparison of 16.0 mGy for Canon and 16.43 mGy for GE, PIQE produced kernel-dependent changes in bone TTF50% relative to matched FBP, ranging from a 21.3% decrease to a 83.6% increase across kernels and DLR strengths. TrueFidelity increased bone TTF50% relative to matched FBP moderately across all evaluated kernels and strengths, with changes ranging from 3.7% to 14.7%. As dose decreased, PIQE TTF50% and TTF10% decreased substantially across evaluated kernels and inserts, whereas TrueFidelity TTF values remained relatively stable. PIQE produced lower NMR values than TrueFidelity. With decreasing dose, PIQE also demonstrated decreasing NMR and CFR values, indicating greater relative noise reduction accompanied by a shift toward lower-frequency noise texture. PIQE and TrueFidelity demonstrated distinct dose-dependent image quality behavior. PIQE provided substantial noise reduction and high task-based spatial resolution at higher dose levels, consistent with its 1024 matrix super-resolution implementation, but showed stronger dose dependence in TTF, NMR, and CFR. TrueFidelity demonstrated more stable spatial resolution and noise texture across dose levels, with more modest noise reduction. Although newer DLR algorithms demonstrate substantial potential for noise reduction, their effects on spatial resolution, noise magnitude, and noise texture must be characterized before aggressive dose reductions are implemented.

Topics

Deep LearningPhantoms, ImagingTomography, X-Ray ComputedImage Processing, Computer-AssistedAlgorithmsJournal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.