3D+t segmentation of PET images using spectral clustering

Segmentation of dynamic PET images is often needed to extract the time activity curve (TAC) of regions. While clustering methods have been proposed to segment the PET sequence, they are generally either sensitive to initial conditions or favor convex shaped clusters. Recently, we have proposed a deterministic and automatic spectral clustering method (AD-KSC) of PET images. It has the advantage of handling clusters with arbitrary shape in the space in which they are identified. While improved results were obtained with AD-KSC compared to other methods, its use for clinical applications is constrained to 2D+t PET data due to its computational complexity. In this paper, we propose an extension of AD-KSC to make it applicable to 3D+t PET data. First, a preprocessing step based on a recursive principle component analysis and a Global K-means approach is used to generate many small seed clusters. AD-KSC is then applied on the generated clusters to obtain the final partition of the data. We validated the method with GATE Monte Carlo simulations of Zubal head phantom. The proposed approach improved the region of interest (ROI) definition and outperformed the K-means algorithm.