Feature extraction by subspace fitting of time activity curve in PET dynamic studies

In computer-aided tumor detection, it is important to exploit features which discriminate lesions from normal tissue. Signal subspace is a relatively robust feature and has been used to identify signals in many applications. In this paper, the authors demonstrate that the time activity curves (TACs) of lesions and normal tissue in multi-frame dynamic positron emission tomography (PET) images can be characterized by two distinct subspaces. The subspace fitting techniques used in classical array sensor processing are applied to extract the subspaces of time activity curves associated with lesions and normal tissues, respectively. The MUltiple SIgnal Classification (MUSIC) algorithm and the least squares based subspace fitting method are both investigated. Based on a physiologic compartmental model of tracer kinetics, the authors show that the TACs can be represented by a linear combination of exponential functions. The problem of fitting TACs into a subspace spanned by these exponential functions then turns out to be the well-known problem of finding direction of arrival (DOA) in array signal processing. The two problems differ only in the search range. The results of applying the MUSIC method and least squares based fitting to the clinical dynamic PET data are also shown and compared in this paper