Abstract-In this paper we address the difficult problem of parameter-finding in image segmentation. We replace a tedious manual process that is often based on guess-work and luck by a principled approach that systematically explores the parameter space. Our core idea is the following two-stage technique: We start with a sparse sampling of the parameter space and apply a statistical model to estimate the response of the segmentation algorithm. The statistical model incorporates a model of uncertainty of the estimation which we use in conjunction with the actual estimate in (visually) guiding the user towards areas that need refinement by placing additional sample points. In the second stage the user navigates through the parameter space in order to determine areas where the response value (goodness of segmentation) is high. In our exploration we rely on existing ground-truth images in order to evaluate the "goodness" of an image segmentation technique. We evaluate its usefulness by demonstrating this technique on two image segmentation algorithms: a three parameter model to detect microtubules in electron tomograms and an eight parameter model to identify functional regions in dynamic Positron Emission Tomography scans.Index Terms-Parameter exploration, Image segmentation, Gaussian Process Model.
MOTIVATIONFor visual analysis image data often need to be segmented. Segmentation refers to the process of partitioning the image into multiple segments, i.e. sets of pixels or voxels, that form contiguous and semantically meaningful regions. If each of these regions is marked by a unique identifier, image segmentation simply means labelling of pixels or voxels. In biomedical imaging, where images are acquired using some kind of tomography or microscopy, segmented regions might correspond to anatomical structures in the case of nonfunctional imaging, and to regions with specific physiological activity in the case of functional imaging.In recent years a variety of semi-and fully automatic techniques have been developed to address the segmentation problem [32]. However, even the current state-of-the-art approaches fall short of providing a "silver bullet" for image segmentation. This has several reasons. One reason is that given some image, the segmentation problem is not well defined; in fact it depends on the application which regions are semantically meaningful. Another reason is that due to different image degradation factors such as low signal-to-noise ratio, imaging artifacts, partial volume effects and shape variability, different kinds of a priori knowledge need to be included. Additionally, the majority of the existing segmentation methods rely on and are sensitive to setting a number of parameters. For example, most of the algorithms contain weighting parameters between multiple competing image-driven or prior-driven cost terms in an attempt to mimic the cognitive capabilities of expert users (e.g. radiologists for medical images).A good parameter setting is usually found by a manual trial and error procedure. Th...