AUTOMATIC DRUSEN DETECTION FROM COLOUR RETINAL IMAGES Saurabh Garg, Jayanthi Sivaswamy and Gopal Datt Joshi Center for Visual Information Technology IIIT Hyderabad, Hyderabad, India saurabhgarg@students.iiit.ac.in, jsivaswamy@iiit.net, gopal@research.iiit.ac.in ABSTRACT The health of the retina deteriorates with age in some peo- ple due to the appearance of drusens. Drusens are accumu- lation of lipid and other waste material from different layers of the retina. These are markers of age-related macular de- generation (ARMD) as their increasing number generally indicates risk for ARMD, a leading cause of blindness in people above the age of 50. Morphological information of drusens is also crucial in determining the risk factor for ARMD. Colour retinal images are used presently to visu- ally identify the presence of drusens. Automated detection and analysis can provide vital information about the quan- tity and quality of the drusens. In this paper, we report on two methods that we have developed to reliably detect and count drusens. The methods exploit the morphologi- cal characteristics of the drusens such as texture and their 3D profiles. We compare the results of using these two methods and make recommendations for automated drusen analysis. KEY WORDS Drusen detection, retinal image, ARMD, segmentation, medical image analysis 1 Introduction Drusens are deposits of cellular waste that accumulate be- neath the retina. They are the primary cause of age-related macular degeneration (ARMD), the leading cause of late- age blindness. The common method to screen for drusen is through retinal imaging. This paper focuses on methods to automatically detect and segment drusen in a retinal image without human supervision or interaction. These methods can be used to develop tools for screening, which help re- duce costs by minimising the need for detailed scrutiny of large quantity of images by eyecare professionals. These methods could also be applied for treatment evaluation, by providing a quantified measurement of drusen presence that is objective and repeatable. An accurate count of drusens in a colour retinal im- age, provides ample information about the extent of dis- ease. Obtaining this information requires robust detection of drusen regions. The task of automatic detection poses various challenges. Drusens appear as yellowish, cloudy blobs in a retinal image. They exhibit no specific size or shape. The modification of size in individual drusens and their confluence seem to be an essential risk factor in de- veloping macular degeneration. Drusens are classified as either hard or soft. Hard drusens tend to be smaller, more sharply defined and are generally less harmful than soft drusens. Soft drusens may be accompanied by other symp- toms such as new vessel formation or fluid build-up in mac- ula [1] [3]. The fuzzy boundaries of soft drusens pose a challenge in accurately locating the actual drusen region. A further challenge in segmenting drusens is the presence in the retina of other similar structures, such as optic disk, exudates and cotton wool spots. Some faint drusens can also appear similar to normal features of the retina, such as the background pattern caused by the choroidal vessels [3]. Furthermore, non-uniform illumination and variable con- trast within the image (due to acquisition process of the im- age) make detection and segmentation task difficult. Thus the use of traditional segmentation methods are inadequate due to the nature of images as well as to the various aspects of drusen. These are some of the important factors which are needed to be addressed in order to perform an accurate detection and count of the drusens. Figure 1. Sample colour retinal image. There are very few attempts specifically on automated drusen detection or segmentation in retinal imagery. Sebh et al. [4] use a modified morphological operator to detect the brightest points (peaks) within individual drusens. Ra- pantzikos and Zervakis et al. [2] adopt an adaptive local histogram based method to identify an appropriate local threshold for segmenting each drusen. These methods how- ever require a manual segmentation of the region of inter- est which is the region around the macula and between the 2 major veins (arcades). The manual segmentation elimi-