Lesion segmentation in whole-body images of PET

In PET, identification of lesion boundaries in general is not a trivial problem as whole-body images exhibit inhomogeneity. Manual segmentation methods in current commercial software packages to identify lesion boundaries and to quantify in terms of standard uptake value (SUV) are very laborious and tedious. They discourage physicians from taking advantage of the inherently quantitative data and compel them to use qualitative means in their diagnosis and assessment of patient response to therapy. An automated segmentation method will enable physicians to extract maximum or mean SUV values from a lesion volume. It may also allow them to track changes in lesion size. In this study, a segmentation algorithm based on the Markov random field modeling has been proposed to segment lesions, as well as other structures, in whole-body PET images for mainly the purpose of diagnosis and the assessment of tumor response to therapy. The method was applied to phantom and whole-body images. The preliminary results indicate that the method can successfully segment lesions.