A new unsupervised method for the segmentation of rodent whole-body dynamic PET images: Comparison to other methods

In a Volume of Interest (VOI) based data analysis of Positron Emission Tomography (PET), the relevance of the extracted Time Activity Curves (TACs) is dependent on VOIs delineation accuracy. Several authors proposed unsupervised methods to draw automatically VOIs in PET studies. Among them, Wong proposed a k-means approach [1] and Brankov proposed Expectation-Maximization methods based on the similarity metrics [2], and compared his methods to others [1; 3; 4]. We have designed an automated segmentation method, the Local Means Analysis (LMA), based on local differences in the pharmacokinetics. The method is validated on simulated rat phantom images that include physiological movements, as well as on a real dataset of dynamic PET images acquired in rats.