A segmentation framework towards automatic generation of boost subvolumes for FDG-PET tumors: a digital phantom study

Potential benefits of administering nonuniform radiation dose to heterogeneous tumors imaged with FDG-PET have been widely demonstrated; whereas the number of discrete dose levels to be utilized and corresponding locations for prescription inside tumors vary significantly with current existing methods. In this paper, an automated and unsupervised segmentation framework constituted mainly by an image restoration mechanism based on variational decomposition and a voxel clustering scheme based on spectral clustering was presented towards partitioning FDG-PET imaged tumors into subvolumes characterized with the total intra-subvolume activity similarity and the total inter-subvolume activity dissimilarity being simultaneously maximized. Experiments to evaluate the proposed system were carried out with using FDG-PET data generated from a digital phantom that employed SimSET (Simulation System for Emission Tomography) to simulate PET acquisition of tumors. The obtained results show the feasibility of the proposed system in dividing FDG-PET imaged tumor volumes into subvolumes with intratumoral heterogeneity being properly characterized, irrespective of variation in tumor morphology as well as diversity in intratumoral heterogeneity pattern.