Mixture principal component analysis for distribution volume parametric imaging in brain PET studies

In this paper, we present a mixture principal component analysis (mPCA)-based approach for voxel level quantification of dynamic positron emission tomography (PET) data in brain studies. The parameters of the probabilistic mixture model are determined using an EM algorithm. The problem of interest here requires neither the accurate arterial blood measurements as the input function nor the existence of a reference region. The effects of mPCA are examined in two different ways on the basis of whether the compartmental model for tracer dynamics is considered. First, the mPCA approach itself is used to classify all voxels into the specific binding and non-specific binding groups, and the resulting power is used for revealing the underlying distribution volume (DV) image. Second, the proposed mPCA-based classification approach is incorporated as the clustering preprocessing into our earlier work to simultaneously estimate the DV parametric image and the input function. The efficiency and superiority of the proposed scheme is demonstrated by real brain PET data