MDL based structure selection of union of ellipse models at multiple scale descriptions

摘要

In this paper we investigate refinements to the structure selection method used in our recently developed minimum description length (MDL) method (Hukkanen et al., 2011, submitted) for interpreting clumps of nuclei in histological images. We start from the SNEF method (Hukkanen et al., 2010), which fits elliptical shapes to the clump image based on the extracted contours and on the image gradient information. Introducing some variability in the parameters of the algorithm we obtain a number of competing interpretations and we select the least redundant interpretation based on the MDL principle, where the description codelengths are evaluated by a simple implementable coding scheme. We investigate in this paper two ways for allowing additional variability in the basic SNEF method: first by utilizing a preprocessing stage of smoothing the original image using various degrees of smoothing and second by using re-scaling of the original image at various downsizing scales. Both transformations have the potential to hide artifacts and features of the original image that prevented the proper interpretation of the nuclei shapes, and we show experimentally that the set of candidate segmentations obtained will contain variants with better MDL values than the MDL of the initial SNEF segmentations. We compare the results of the automatic interpretation algorithm against the ground truth defined by annotations of human subjects.