https://atlas.rsna.org/schemas/2025-11/model.json

Denoising Multiphase Functional Cardiac CT Angiography Using a 3D U-Net with Synthetic Training Data

https://doi.org/10.1148/ryai.230153

1.0

Address correspondence to V.S. (email: moc.liamg@trofdnas.tiev)

Authors declared no funding for this work. Acknowledgments note support from the Stanford Radiology Department Radcombinator program and computer support from the Stanford 3D and Quantitative Imaging Laboratory.

HIPAA-compliant retrospective study approved by the institutional review board; patient consent waived due to use of anonymized data.

Modified 3D U-Net (2D+t) with residual connections and group normalization; L1 loss; Adam optimizer; 64 initial feature maps; learning rate 1e-4; batch size 4. Training on 2×NVIDIA V100 GPUs for ~14 hours. Also trained a 2D U-Net for comparison.

Reduces noise and artifacts in dose-modulated, retrospectively gated cardiac CTA functional frames, enabling valid semi-automatic LV functional measurements (e.g., LV area, LVEF).

Image reconstruction/processing to facilitate downstream quantitative analysis (functional assessment).

Fully automated denoising of cine CT frames; downstream LV segmentation demonstrated with a simple threshold-based semi-automatic method for validation.

Denoising of high-noise time frames from retrospectively ECG-gated, dose-modulated coronary CT angiography in adult patients to facilitate functional cardiac analysis (e.g., LV function) in a research setting.

Multiphase cardiac CTA cine series per axial slice with 10 temporal frames (preprocessed to 128×128×16 volumes via padding/wraparound).

Select lowest-noise time point (typically diastole); generate synthetic low-noise labels for other time points via nonrigid registration (elastix 5.0.0); train 3D U-Net with L1 loss using specified hyperparameters; apply trained model to denoise high-noise frames across the cardiac cycle.

Single-center, single-vendor data (Siemens); standardized injection protocols; anonymization precluded correlation with patient characteristics; not validated for coronary vessel or valve evaluation; synthetic labels can contain registration-induced deformations; evaluation focused on LV function; no direct comparison to other deep learning denoising methods beyond 2D U-Net and BM3D.

Use 3D U-Net denoised images for quantitative LV functional analysis when dose modulation produces unusably noisy frames in retrospectively gated coronary CTA.

Key hyperparameters and preprocessing steps provided; public registration tool (elastix) and comparison algorithm (BM3D/pybm3d) specified; hardware (2×V100) and training time reported.

Training time ~14 hours on 2×V100 GPUs; runtime performance not reported.