Quantitative evaluation of noise reduction strategies in dual-energy imaging.

In this paper we describe a quantitative evaluation of the performance of three dual-energy noise reduction algorithms: Kalender's correlated noise reduction (KCNR), noise clipping (NOC), and edge-predictive adaptive smoothing (EPAS). These algorithms were compared to a simple smoothing filter approach, using the variance and noise power spectrum measurements of the residual noise in dual-energy images acquired with an a-Si TFT flat-panel x-ray detector. An estimate of the true noise was made through a new method with subpixel accuracy by subtracting an individual image from an ensemble average image. The results indicate that in the lung regions of the tissue image, all three algorithms reduced the noise by similar percentages at high spatial frequencies (KCNR=88%, NOC=88%, EPAS=84%, NOC/KCNR=88%) and somewhat less at low spatial frequencies (KCNR=45%, NOC=54%, EPAS=52%, NOC/KCNR=55%). At low frequencies, the presence of edge artifacts from KCNR made the performance worse, thus NOC or NOC combined with KCNR performed best. At high frequencies, KCNR performed best in the bone image, yet NOC performed best in the tissue image. Noise reduction strategies in dual-energy imaging can be effective and should focus on blending various algorithms depending on anatomical locations.