A segmentation algorithm for heterogeneous tumor automatic delineation in PET

Accurate volume of interest (VOI) contouring in PET is crucial in numerous oncology applications such as in radiotherapy or for the derivation of semi-quantitative indices of activity concentration in response to therapy studies. On one hand, the most widely used method to semi-automatically determine VOIs in PET is thresholding. Unfortunately, such approaches lead to variable VOIs determination as shown in multiple clinical studies. On the other hand, numerous works have suggested automatic algorithms for lesion detection in PET. Their performance is usually sensitive to variations of noise, intensity and/or lesion contrast. Finally, none of the aforementioned methodologies have been assessed in the presence of non-spherical and/or inhomogeneous lesions. We have previously developed an unsupervised Bayesian segmentation algorithm for PET taking into account the noise, a voxel\´s intensity and spatial information, in order to classify a voxel as "background" or "functional volume". This method has been shown to perform better with respect to other methodologies for volume determination of homogeneous and spherical objects in PET. As lesions are usually not spherical and characterized by non-homogeneous activity distributions we have extended and improved the previously developed approach by introducing the use of three hard classes and fuzzy transitions, rather than two hard classes and one fuzzy transition used in the previous implementation. The new algorithm was evaluated on synthetic images and its performance was compared to the hard 3-class segmentation as well as the fuzzy C-Means clustering algorithm. The results show that the 3-class fuzzy scheme is able to deal with heterogeneous tumor activity concentrations in an accurate manner. Further work will concentrate on the full validation of the proposed methodology using more realistic images of heterogeneous objects reconstructed from GATE simulations and real images from patients undergoing IMRT.