Incremental value of deep learning denoising in low-dose coronary CT angiography in predominantly obese patients

To determine the added value of post-hoc convolutional neural network (CNN)-based denoising for low-radiation coronary CT angiography (CCTA).

In this IRB-approved, post-hoc study, consecutive patients underwent clinically indicated CCTA on a third-generation dual-source CT scanner using a fast-pitch protocol and fully patient-tailored contrast injection. Images were reconstructed with a vascular kernel (Bv40) using iterative reconstruction (ADMIRE 4), with and without CNN-based denoising. Two readers assessed signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), attenuation, and subjective image quality across coronary segments using a 4-point scale. Linear mixed-effects models evaluated the relationship between body-mass index (BMI) and denoising efficiency.

Fifty-six participants (mean age 51.7 ± 12.8 years; 57.1% women), including 35 obese individuals (62.5%, BMI ≥ 30.0 kg/m²), were analyzed. The dose-length product was 77.15 [50.10-133.38] mGy·cm. CNN-denoising reduced image noise (35.1 ± 7.8 vs 21.0 ± 4.7 HU, p < 0.001) and improved SNR (12.0 ± 3.3 vs 22.0 ± 5.7, p < 0.001) and CNR (11.0 ± 3.0 vs 19.0 ± 5.2, p < 0.001), with stable attenuation (all p > 0.1). Subjective quality improved (2 [2,3] vs 3 [2,3], p < 0.001). Poor-quality vessels decreased from 23/156 (14.7%) to 14/156 (9.0%), and good-to-excellent quality increased from 82/156 (52.6%) to 117/156 (75.0%). Mixed-effects modeling showed 31% noise reduction (-13.9 HU, p < 0.001) and SNR (+8.5, p < 0.001) and CNR (+7.4, p < 0.001) gains, irrespective of BMI (p > 0.46).

CNN-based denoising improved objective and subjective image quality in low-dose CCTA independently of BMI in a predominantly obese cohort.