Segmentation of rat spinal cord in PET using spatiotemporal information

Segmentation in PET images is fraught with difficulty stemming from variable activity values and low SNR. Standard practice utilizes co-registered images from other modalities to provide anatomical information. Sometimes, this information may be missing or of limited usefulness. We present a method of extracting centerlines of rat spinal cords solely from dynamic PET images. The method relies on the unique temporal information in the voxel time activity curves (TAC) to improve segmentation results. Using techniques previously developed for carotid arteries, centerlines were modeled by B-splines to ensure smooth realistic curves. This method is highly automated, only requiring user definition of a small number of seed points. The method was applied to [11C]AFM studies which measure serotonin transporters in the cord. Initial analysis showed that the method yielded comparable results in standard uptake values (SUV) compared to manual delineation of regions of interest (ROI). It also demonstrated an improved outcome over segmentation based on intensity alone.