1st International Conference on Mathematical and Computational Biomedical Engineering - CMBE2009 June 29 - July 1, 2009, Swansea, UK P.Nithiarasu and R.L¨ ohner (Eds.) Using Shape Entropy as a Feature to Lesion Boundary Segmentation with Level Sets Elizabeth M.Massey University of Lincoln, Brayford Pool, Lincoln LN6 7TS UK, bmassey@lincoln.ac.uk Andrew Hunter University of Lincoln, Brayford Pool, Lincoln LN6 7TS UK, ahunter@lincoln.ac.uk James A. Lowell Foster Findlay Associates Limited, Newcastle Technopole Kings Manor, Newcastle Upon Tyne NE1 6PA UK, james.lowell@gmail.com David Steel Consultant Ophthalmologist, Sunderland Eye Infirmary, Queen Alexandra Road, Sunderland SR2 9HP UK, David.Steel@chs.northy.nhs.uk ABSTRACT Accurate lesion segmentation in retinal imagery is an area of vast research. Of the many segmentation methods available very few are insensitive to topological changes on noisy surfaces. This paper presents an extension to earlier work on a novel stopping mechanism for level sets. The elementary features scheme (ELS) in [5] is extended to include shape entropy as a feature used to ’look back in time’ and find the point at which the curve best fits the real object. We compare the proposed extension against the original algorithm for timing and accuracy using 50 randomly selected images of exudates with a database of clinician demarcated boundaries as ground truth. While this work is presented applied to medical imagery, it can be used for any application involving the segmentation of bright or dark blobs on noisy images. Key Words: Shape Special Session, Exudate Segmentation, Level Sets, Medical Image Processing. 1 INTRODUCTION The diagnosis of diabetic retinopathy is based upon visually recognizing various clinical features. Retinal lesions are among the first visual indicators suggestive of diabetic retinopathy. The threat to visual loss increases with the frequency of retinal lesions combined with their encroachment into the macula. To enable early diagnosis, it is necessary to identify both frequency and position of retinal lesions in relation to the fovea and other major structures (such as the optic nerve). In [5] a lesionness measure was introduced and defined as a combination of perimeter size constancy shp and compactness c = p 2 /a, where p is the perimeter and a is the area [3]. The lesionness measure was the core of the stopping mechanism and upon further analysis we discovered a more direct approach by tracking the entropy of the shape of the region of interest (roi). In this work we introduce the notion of using a multivariate histogram to describe the changing shape of the roi and track the shape entropy to determine the best fit. The correlation between the change in shape entropy and the perimeter size constancy indicates the point where the curve best fits the lesion (or region of interest). 2 BACKGROUND Retinal exudates are an interesting challenge for segmentation algorithms as they vary in appearance, conforming to one of three structures: dot exudates, fluffy exudates and circumscribed plaques of exudate. Dot exudates consist of round yellow spots lying superficially or deep in the sensory retina [9]. Exudates are usually reflective and may appear to have a rigid, multifaceted contour, ranging in color from white to yellow [1]. With varying 230