Turn down the noise--a blinded evaluation of iterative image reconstruction in radiation therapy computed tomography...

The current standard reconstruction algorithm for computed tomography (CT) scans is filtered back projection. Alternative algorithms using iterative reconstruction (IR)-in our case, "sinogram affirmed iterative reconstruction"-have been increasingly implemented in diagnostic CT imaging. We studied its potential in improving radiation therapy planning images.

Raw planning CT data sets of patients from varied disease sites were reconstructed using filtered back projection and IR levels 1, 3, and 5 with equal radiation dose. For each site, 2-7 patient scans were selected; 2-3 physicians blindly evaluated the 4 3-dimensional image sets. Using a visual analogue scale, they rated the sharpness, noise, perceived ease in delineating gross tumor/clinical target volume and organs at risk, and overall appreciation of the images. Interobserver correlation was calculated with the Spearman correlation coefficient (ρ). Generalized estimating equations assessed the differences in the mean score for each criterion between reconstructions. When significant differences existed, pairwise comparisons compared the least-squares means. The preference for each reconstruction was rank ordered for each observer. The rank occurrences were computed; generalized estimating equations and pairwise comparisons were again calculated.

Sixteen physicians reviewed 548 image sets. The sharpness had a ρ = -0.22-0.53, noise ρ = -0.34-0.38, delineating gross tumor/clinical target volume ρ = -0.28-0.53, delineating organs at risk ρ = -0.47-0.42, and overall appreciation ρ = -0.17-0.38, suggesting a low level of agreement among observers. IR 3 and 5 had consistently higher scores and ranks than filtered back projection (P = .02 and P = .015, respectively). Paradoxically, IR 5 scored both highest and worst the most frequently. IR 3 was more consistently well-ranked for all criteria.

This report is the first to clinically evaluate IR in radiation therapy planning. When used to reduce noise in current CT simulation protocols, IR images were generally preferred. Although highly processed images polarized observers, the use of moderate IR was appreciated for most disease sites.