Robust segmentation and accurate target definition for positron emission tomography images using Affinity Propagation

Distributed inflammation in infectious diseases cause variable uptake regions in positron emission tomography (PET) images. Due to this distributed nature of immuno-pathology and associated PET uptake, intensity based methods are much better suited over region based methods for segmentation. The most commonly used intensity based segmentation is thresholding, but it has a major drawback of a lack of consensus on the selection of the thresholding value. We propose a method to select an optimal thresholding value by utilizing a novel similarity metric between the data points along the gray-level histogram of the image then using Affinity Propagation (AP) to cluster the intensities based on this metric. This method is tested against the PET images of rabbits infected with tuberculosis with distributed uptakes with promising results.