IMAGE SEGMENTATION FOR DRUSEN SPOTS DETECTION AND MODELLING Fernando Moitinho*, André Mora*, Pedro Vieira*, José Fonseca* *Intelligent Robotics Center, Uninova, Portugal, Faculty of Sciences and Technologies, New University of Lisbon Portugal, fmm@uninova.pt, atm@uninova.pt, pmv@fct.unl.pt, jmf@uninova.pt Keywords: Image Modelling, Medical Image, Drusen Detection, Digital Image Processing Abstract Age Related Macula Degeneration (ARMD) is an eye disease characterized by the appearing of yellowish spots in the retina. These spots are usually called Drusen spots. Drusen spots are object of analysis by the ophthalmologists for treatment effectiveness evaluation. In this paper we proposed a methodology to help ophthalmologist community in Drusen spots analysis. The objective of our methodology is to achieve a mathematical model of the original image. With this technique noise is removed, analysis can be reproduced independently of the clinician and Drusen spots measures can be done accurately. In the modelling process are used several algorithms: non-uniform illumination correction, based on gaussian blur filter; Drusen location algorithm based on labelling of the maximum gradient path; Levenberg- Marquardt optimization algorithm for image modelling. It is also used an algorithm to divide each original image in smaller images containing Drusen islands. 1 Introduction Age Related Macula Degeneration (ARMD) is a frequent disease, affecting mainly elderly people. It is nowadays the leading cause of blindness in developed countries [1]. Therefore it is commonly diagnosed by ophthalmologists during patient examination. ARMD is characterized by the accumulation of extra-cellular material beneath the basal lamina of the retinal pigment epithelium that builds small sized bubbles [2]. This phenomenon can be seen in retina images as yellowish spots called Drusen spots. Currently ARMD analyses are done manually based only on qualitative aspects. However this analysis is dependent on the clinician and cannot be reproduced. To improve manual analysis, a software application is proposed in this paper. This software application is able to automatically detect and model Drusen spots on retina images. The benefits of using this application are a more accurate analysis, a reduction of image non-uniform illumination problems, independence of the image model to noise and the achievement of a reproducible process independent of the clinician. Another important feature is that it gives important information about each Drusen spot like area and volume. This information can be useful to evaluate the disease progress in a sequence of images taken during a long-term treatment, helping ophthalmologists on the evaluation therapy effectiveness. There are two phenotypes of Drusen spots: hard and soft. These phenotypes are used in grading systems such as Wisconsin age related maculapathy grading system [3] or the Alabama Age-Related maculopathy grading system [4]. Our application is dedicated to the detection of Soft Drusen spots because hard Drusen have reduced significance for disease evaluation (a) (b) Figure 1: Retina images with Soft and Hard Drusen a) Hard Drusen spots; b) Soft Drusen spots The developed application returns a modelled image with relevant Drusen spots and also Drusen spots information (area, volume, location and gaussian parameters). This paper is organized in five sections, starting with a brief presentation of selected related works on Drusen detection. The following section is dedicated to the modelling process describing the following algorithms: non-uniform illumination correction, Drusen spots isolation, Drusen