Deep-learning denoising for ultrahigh-resolution photon-counting detector CT: phantom and in vivo evaluation of non-calcified coronary plaques.

To assess the value of convolutional neural network (CNN)-based denoising for the evaluation of non-calcified coronary plaques on ultrahigh-resolution (UHR) photon-counting detector (PCD) coronary CT angiography (CCTA). A dynamic phantom containing lipid-rich and fibrotic plaques with 50%-diameter stenosis (PDS) was scanned on PCD-CT under varying conditions. For in-vivo imaging, consecutive patients with non-calcified coronary plaques (NCPs) who underwent CCTA with PCD-CT were included. Image series were reconstructed using a sharp vascular kernel (Bv64) with slice thicknesses of 0.2 mm/0.4 mm, quantum iterative reconstruction (QIR) levels 3/4, and with/without CNN denoising. Phantom-based line-profile analysis yielded edge-width at half maximum (EWHM) and 10-90% rise distance as sharpness metrics. Plaque contrast-to-noise ratio (CNR) and PDS were assessed in the phantom and in patients. Two readers evaluated subjective image quality (noise, diagnostic confidence) using a four-point Likert scale. Pairwise comparisons were performed using a Bonferroni-corrected p < 0.002 for significance. Fifty-five patients (median age 74 [66.5-78] years; 19 women) with 97 NCPs were included. Phantom-based sharpness metrics remained unchanged after denoising (all pairwise p > 0.051). CNN denoising increased plaque CNR consistently in the phantom (fibrotic: 0.80-1.64 to 0.92-2.06; lipid-rich: 0.21-0.34 to 0.35-0.46) and in vivo (0.99-2.07 to 1.37-2.92, all pairwise p < 0.001). Denoising did not alter PDS values in phantom or in vivo plaques (all pairwise p > 0.006). Subjective image noise and diagnostic confidence improved across all reconstructions (all pairwise p < 0.001). CNN-based denoising of UHR PCD-CT improves image quality of non-calcified coronary plaques, while preserving sharpness and quantitative stenosis metrics.