Abstract
A model based noise correction algorithm was developed to improve the accuracy of CT perfusion (CTp) blood flow (BF) measurements affected by image noise. The algorithm used tissue attenuation curves (TACs), generated by convolving an impulse response function (IRF) with an arterial input function (AIF) averaged from 59 patient datasets. Gaussian noise was introduced to simulate noise, and BF was measured using deconvolution. The algorithm iteratively compared BF without added noise against noise-impacted BF to estimate ground-truth BF (GTBF). Performance was evaluated with digital perfusion phantoms (DPPs) for GTBF values of 5-420 ml/100 ml/min and added noise (standard deviation 25 HU), measuring absolute difference from GTBF and contrast-to-noise ratio (CNR). For clinical evaluation, CTp data from 14 pancreatic ductal adenocarcinoma (PDAC) patients was used. For DPPs, noise-impacted and noise-corrected BF were 140 ± 111 ml/100 ml/min and 131 ± 125 ml/100 ml/min, compared to GTBF of 131 ± 127 ml/100 ml/min. Post-correction, the absolute difference reduced from 18.8 to 3.6 ml/100 ml/min, with CNR improving from 2.52 to 2.66. In clinical datasets, BF for parenchyma shifted from 148 ± 50.8 to 84.1 ± 96.9 ml/100 ml/min, and for PDAC, from 45.8 ± 20.3 to 13.3 ± 18.7 ml/100 ml/min. The algorithm reduced noise impact, improving BF accuracy and CNR, with potential for lower-dose CT without compromising diagnostic quality.