Detection of partial volume effected lesions in dynamic positron emission tomograph images

Most existing partial volume tissue classification techniques estimate the constituents (or proportions) of mixture from multiple-channel observations under one hypothesized mixel model, assuming that each pixel (or voxel) is composed of all types of tissue in the field of view including lesions presented in the image. In this paper, the authors propose to test a pixel in a multi-frame positron emission tomography (PET) dynamic image sequence by two hypothesized mixel models: H₀ and H₁. It is assumed that the proportion of a lesion contributed to the mean of an observed time activity curve in the pixel is zero under H₀, but large than zero under H₁. A generalized maximum likelihood ratio GLR test is derived to distinguish which hypothesis holds for the given pixel. In order to estimate the constituents of mixture under the convexity constraint for each hypothesis, a recursive search algorithm is designed to find a global optimum, which yields a closed-form expression for the GLR test function. The performance of the GLR test is analyzed and tested using clinical PET dynamic data