R-clustering technique for initialization of factor analysis of dynamic PET images

In this work we develop a new clustering algorithm for the initial analysis of a dynamic PET image series. Time-dependent activity for each voxel is treated as an independent time series and clustered. The set of representative time series for each cluster is used to initialize a factor analysis algorithm that determines the blood input function, the tissue-specific time-activity curves, and the spatial distribution of tissues with different tracer dynamics from a noisy reconstructed sequence of 3D PET images. The novelty of the approach is in the combination of a unique distance metric between a pair of time series that is more suitable for our subsequent independent component analysis approach. It uses a concept of representative clustering, when only a small number of the analyzed time series are used for clustering, which are automatically determined. The developed method is applied to analyze
¹¹
C-PIB dynamic brain PET images acquired to study β-amyloid deposition in brains of Alzheimer patients.