Numerical observer evaluation of 3D statistical PET reconstruction method with system response modeling

Previously we have developed a 3D statistical method, which incorporates point spread function in the system matrix. Comparing to our standard 3D reconstruction algorithm with simple line integral model, this method resulted in improved resolution and noise behavior of phantom images. Reconstruction of clinical data is usually performed from rebinned projection data, resulting in downsampled patient image. Efficient algorithm implementation allows reconstruction of patient data from rebinned data in time close to clinically acceptable. Then numerical observer study to evaluate algorithms performance is feasible. We use channelized Hotelling observer to establish hot spot detectability in noisy images for each method. Patient-derived phantom was used in computer simulations when a few known location tumors of various size and contrasts were added to patient image. Number of iterations and postsmoothing amount, optimal for hot spot detectability, were also investigated.