MD3RI a Tool for Computer-Aided Drusens Contour Drawing

André Mora 1,2, Pedro Vieira 2, José Fonseca 1,2 1 Intelligent Robotics Center, Uninova, Portugal

2 Faculty of Sciences and Technologies, New University of Lisbon, Portugal atm@uninova.pt, pmv@fct.unl.pt, jmf@uninova.pt

ABSTRACT Computer aided tools that can help doctors in repetitive analyses are becoming more reliable and faster what contributes for its increasing acceptance in the medical sector. In this paper it is presented a software application designed to facilitate the quantitative evaluation of Drusen on retina images. The target users are ophthalmologists wanting to diagnose Age-related Macular Degeneration (ARMD) or to evaluate therapies effectiveness. Drusen are ARMD risk factors characterized by yellow spots located around the macula. The application includes image enhancement tools for improving or calibrating the visualization conditions, a manual or semi-automatic contour drawing procedures and several file I/O to allow analyses reproduction. The semi-automatic drawing uses an image segmentation algorithm based on the image gradient which reduces noise artifacts. This algorithm is described in detail in this paper. Results and conclusions taken during application tests performed by ophthalmologists are also presented. KEY WORDS Age-Related Macular Degeneration, Drusen Detection, Medical Image Processing 1. Introduction During a retina inspection for detecting pathologies or evaluating treatments effectiveness ophthalmologists usually look for abnormal structures in the retina surface. The most common are microaneurysms, hemorrhages, tumors, exudates, macular holes and Drusens deposits. Most of these are characterized by fuzzy variables like: shape, size, location, etc. and also these features can appear in low contrast areas making its detection very subjective. For this analysis to be accurate, quantifying the abnormalities, it requires a fastidious work from the expert that must be skilled for this task. Ophthalmologists usually analyze these retina images by visual inspection and, in some cases, quantify the extension of the abnormalities manually. Efficient and easy to use tools that can help ophthalmologists to quantify the abnormalities or optometrists working on

eyeglass shops to automatically analyze the images and alert their clients if they need to visit the ophthalmologist are urgent needs in this field. The development of such tools is a difficult task since for many abnormalities there isn’t a clear definition of the detection criteria. This leads to a lack of reproducibility. In some cases this definition differs from doctor to doctor or worst doctors can have different criteria at different times. Visualization conditions are also another important factor. For a manual analysis these should have common settings of color temperature, lightning and contrast, what can be achieved in some cases using a calibration procedure. The use of automatic detection tools can overcome some of these problems. With these tools there is no need to have special monitor calibration tools, since these only for

Fig. 1 Examples of retina images a) Retina with Hard Drusens; b) Retina with Soft Drusens.

visualization purposes and the information seen by the computer is the same. Also, the detection criteria are the same for every image, not suffering from doctors’ criteria variability. There are already computer based tools for quantifying some types of diseases. These had to be evaluated by a panel of doctors to define a gold standard for the particular type of detection. A gold standard can be defined has a detection algorithm that the majority of doctors agreed to produce good detection results. In this paper it is addressed the detection of Drusen deposits on patients with Age-Related Macular Degeneration (ARMD). This is one of the risk factors for ARMD the leading cause of irreversible blindness in developed countries [1]. Due to unknown causes extra-cellular materials are deposited beneath the retina surface building small sized bubbles that are designated by Drusen deposits. These are visible in retina imaging as yellow spots that vary in size depending on their phenotype and disease evolution. The main terminology to distinguish Drusen phenotypes used in grading systems such as the Wisconsin age-related maculopathy grading system [2] or the Alabama Age-Related Maculopathy Grading System [3] is hard and soft. Hard Drusen (figure 1.a) are mainly characterized by small sized spots with less then 50µm diameter and sharp edges. Soft Drusens (figure 1.b) are medium sized (≈ 250µm) spots with smooth edges. A computer based tool to draw Drusens contour was developed as a first step for evaluating an algorithm that can accurately automatically detect Drusen Deposits [4-6] and meanwhile help doctors analyze their images. The application is entitled Manual Drusen Deposits Detection on Retina Images (MD3RI). The contour is draw over digital fundus images in an easy to use, semi-automatic procedure. The software uses an algorithm based on image gradient information [5] to detect Drusens location which gives also the area of influence of each one. This algorithm is described in detail on the section “Drusens Location Algorithm”. The software application developed to draw Drusens contour and all its functionalities will be presented in the third section of this paper.

2. Drusens Location algorithm We propose a Drusens location algorithm based on labeling the maximum gradient path. In medium resolution images there are more then just one pixel pointing to each intensity maximum. Therefore, it is possible to follow one or more ascending paths that inevitably reach an intensity maximum. The proposed algorithm consists in a labeling procedure similar to connect components labeling or to watershed segmentation (described in [7]), but optimized to maximums search. The algorithm begins by determining the image gradient using a 3x3 Sobel operator which evaluates intensity changes in the horizontal and vertical axis. The pixel gradient is a vector oriented to the ascending pixel. The two-dimensional gradient analyses the 3x3 neighborhood of each pixel what eliminates several noise related artifacts. But when signal/noise ratio is low the algorithm can be significantly improved by applying a mean filter for noise reduction and afterwards computing its gradient. Figure 2.b contains an example of an image after applying a 3x3 gradient mean. The original image (Figure 2.a) is 10x10 pixels detail of two small Drusens. The first stage of the labeling procedure – initial label propagation – consists in examining every pixel in a top-left to bottom-right direction and assigning a new label to all pixels that are not marked yet. Whenever a label is assigned to a pixel it is propagated to the neighbor pixel that is in the gradient azimuth direction. This labeling continues on the propagated pixels until the next pixel is already marked. Figure 2.c presents the result of the labeling first stage. Every time the label propagation finishes on the same label that is being propagated, that pixel is marked as a intensity maximum. In the same situation but when there is a different label, these are defined as being compatibles. This means that they belong to the same intensity maximum, i.e., the same Drusen spot. The second stage of the labeling procedure is to apply the label compatibilities resulting on an image with as many labels as possible Drusens. The result is a segmented

1 2 3 4 5 6 7 8

9 1 4 5 6 7 10 11

1 12 6 6 7 10 11 13

14 15 16 7 17 10 13 18

19 20 21 7 7 10 18 22

23 24 25 25 25 7 22 26

27 28 29 30 7 31 32 33

34 35 36 37 37 38 39 40

1 1 1 1 1 1 7 7

1 1 1 1 1 7 7 7

1 1 1 1 7 7 7 7

1 1 1 7 7 7 7 7

1 1 7 7 7 7 7 7

1 7 7 7 7 7 7 7

1 7 7 7 7 7 7 7

1 7 7 7 7 7 7 7 a b c d

Fig. 2 Example of Drusens center detection algorithm. (a) Original image (detail); (b) Image filtered gradient; (c) Labeling 1st stage – initial label propagation – (with label propagation on the two upper rows); (d) Labeling 2nd stage – apply

compatibilities - (two Drusens centered on the highlighted pixels).

image where all pixels that contribute to a spot have the same label. Each group of pixels with the same label constitutes the spots’ area of influence. It is important to notice that the algorithm is so far a non-parameterized algorithm. But a problem arises when flat valleys or flat hills occur. In these cases not all gradient paths end on the same intensity maximum pixel since there might exist more then one pixel for that intensity maximum. Consequently, the algorithm will generate more intensity maximums then it really should. For solving this problem two strategies can be taken: flooding or merging, both requiring parameterization. The flooding method is applied when the label propagation reaches a pixel in which the gradient modulus is below a predefined value. In this situation, the label will continue being propagated in all directions that have gradient modulus below the predefined value. The merging method is more efficient but more time consuming. For every detected maximum, using a connected components approach, it is explored if there is other intensity maximum that can be connected by a path that doesn’t goes lower than a predefined amplitude. At this stage spots are detected and their area of influence defined. The MD3RI will use this information in unnoticeable way to aid ophthalmologists to define Drusens contour. 3. The MD3RI The MD3RI (Manual Drusen Deposits Detection on Retina Images) is an application developed for Ophthalmologists to analyze the presence of Drusen

deposits in retina images. This tool was developed with the purpose of evaluating the performance of the automatic Drusen analysis methodology described in [4, 5], by comparing its results with the same images analyzed by Ophthalmologists in different Medical and Research Centers. The contribution of different Centers around Europe will allow a greater diversity of analysis procedures resulting in a statistically more precise comparison. The latest version of the MD3RI software and some sample images are publicly available at [8]. The MD3RI basic functionalities include:

• File I/O; • Region of Interest configuration; • Image enhancement; • Contour drawing tools.

Besides the usual file open and save operations the user can also save the analysis for a future use or load an analysis previously saved. This file is an image with the same size as the original, but containing just the area drawn (see the image on the right side of figure 3). The original image where it is defined the contour is displayed on the left side of the application. The region of interest (ROI) that ophthalmologists consider to be significant for Drusens analysis is a circle with approximately twice the diameter of the optic disk centered on the macula (see analysis area on figure 3). After being defined this circle using the mouse cursor, the software only allows contours to be drawn inside this area. The ROI definition can be loaded from a file or saved to a file, so that the analysis configuration can be easily be reproduced on other computers participating in the same trial.

Image to analyze(working image)

Analysis area(region of interest)

Drusens Contour

Analysis image

Controls

Fig. 3 Screenshot of MD3RI User Interface.

For better image visualization, especially on images with non-uniform illumination, a set of simple image enhancement tools where added (see figure 4). The first option is Red Free. This consists in converting color images to grey scale not considering their red component. It is commonly used in these types of images to improve image contrast. Image brightness, contrast and equalization adjustments were also added to improve users’ visual perception. Mean filtering for image noise reduction is available, contributing to smoother and better defined contours. Basic zooming capabilities were included enabling more detailed contours to be drawn. For drawing contours two modes are available: the manual and semi-automatic. In both modes it is possible to compute the total area of Drusens drawn. This area can be expressed in µm2 whenever the image scale is known; otherwise it returns the area in pixels. In Manual mode the contour is drawn pressing the mouse left button and moving it around the Drusen contour. When the user releases the mouse button, the shape drawn is then filled and drawn on the analysis image (right side image). At this point, the Drusens contours of the working image (left side image) are updated with the contours of the shapes drawn in the analysis image. This two image procedure allows the user to add or remove (if the user uses the mouse right button) areas in the analysis image, what can be used to refine or fix contours. Also, the analysis image can be easily used for comparison of contours drawn by different Ophthalmologists. In semi-automatic mode it is used the Drusen location algorithm described on the previous section. For initializing the structure containing the located Drusens and their area of influence it is needed to press the Detect button. Afterward, the contours can be drawn pressing the mouse left button on top of the each Drusen and move the mouse to the right or to the left to increase or decrease the

contour size. This drawing procedure is achieved by adding to the analysis image the Drusen area of influence that is above a certain threshold which is being adjusted moving the mouse cursor. In this mode the mouse right button simply removes the Drusen area from the analysis image. The two image procedure described previously is also applied in this mode. During the analysis and for better perception of the next contour, users can hide temporarily the current contours and the ROI circle. This can be done using the buttons Hide Contour and Hide ROI or the pressing CTRL key. 4. Experimental Results For evaluating the software it was presented to several ophthalmologists in Portugal and in Scotland which gave us valuable feedback. In the initial tests the contours drawn in the application were compared with the ones drawn on images printed in professional photographic paper (figure 6.a). With this test it has been concluded that computer aided drawing is not only must faster then paper drawing but it also much flexible allowing corrections and refinements to be made. Initially the analyses were done without any restrictions on the Drusens location. This was found to be a problem since they can appear in very low contrast areas and because there is no stopping condition it can turn the analysis into a very fastidious task. Defining a region of interest where the analysis is made proven to be a correct choice (figure 5). When comparing between manual (figure 6.b) and semi-automatic (figures 6.c and 6.d) drawing modes the choice went for second one due to its easer operation and better defined contours. In some cases when the semi-automatic produces a contour that needs to be slightly refined the manual operation can always be used as a complement.

Fig. 4 Image enhancement panel.

Fig. 5 Drawing example

In figures 6.c and 6.d it is presented two different analyses for the same image done by two ophthalmologists. As it can be noticed in these images the analysis criteria are significantly different. Between these two images only 60% of the pixels are equally marked and 40% of the pixels are differently marked. The stopping condition and the contour criteria are different in both Ophthalmologists. This observation reinforced our convincement of the need for tools that can automatically detect Drusens and that can guarantee repeatability. 5. Conclusions In this paper an application for computer aided drawing of Drusen Deposits contour is presented. It was developed mainly for comparing the results of automatic Drusen detection methodologies with images drawn by Ophthalmologists. In addition, with this application they can increase their rate of images analyzed, since it provides an easy semi-automatic drawing procedure. Also, because these are digital results it is possible to process the results in a more easily and flexible way.

The semi-automatic drawing procedure was positively accepted by the ophthalmologists involved in the study. This procedure is based on an algorithm that uses the image gradient to find Drusens spots and additionally defining their area of influence. Each Drusen contour is defined by the contour defined by the area of influence at a configurable intensity value. Basic image enhancement tools where made available for adjusting the interface to the user visualization conditions. As future work it is planned to improve the semi-automatic contour drawing for accurate quantification, by incorporating more semi-automatic and automatic features. Detection of other structures or abnormalities is also considered. Calibration procedures are another issue to be addressed in the future. Acknowledgements This work is funded by the Fundação para a Ciência e Tecnologia, trough the POCTI and POSI Research Programs (project nº POCI/SAU-ESP/57592/2004).

a b

c d

Fig. 6 Drusen contours drawn with MD3RI. (a) drawn on photographic paper (b) MD3RI in manual drawing mode; (c) (d) MD3RI in semi-automatic drawing performed by two different ophthalmologists;

The authors acknowledge the University of Aberdeen and Hospital Santa Maria for supplying the retina images used in this work and for their valuable feedback testing the software. References [1] D. Pauleikhoff, et al., Drusen as risk factors in age-related macular disease, Am J Ophthalmol, vol. 109(1), p. 38-43, 1990. [2] R. Klein, et al., The Wisconsin age-related maculopathy grading system, Ophthalmology, vol. 98(7), p. 1128-1134, 1991. [3] C. A. Curcio, N. E. Medeiros, and C. L. Millican, The Alabama Age-Related Macular Degeneration Grading System for donor eyes, Invest Ophthalmol Vis Sci, vol. 39(7), p. 1085-1096, 1998. [4] A. Mora, J. Fonseca, and P. Vieira. Drusen Deposits Modelling with Illumination Correction, in Biomed-2005, Innsbruck, Austria, 2005. [5] A. Mora, P. Vieira, and J. Fonseca. Drusen Deposits on Retina Images: Detection and Modeling, in MEDSIP-2004, Malta, 2004. [6] K. Rapantzikos, M. Zervakis, and K. Balas, Detection and segmentation of drusen deposits on human retina: Potential in the diagnosis of age-related macular degeneration, Medical Image Analysis (Elsevier), vol. 7, p. 95–108, 2001. [7] R. Gonzalez and R. Woods, Digital Image Processing (Addison-Wesley,1992). [8] Uninova, MD3RI - Manual Drusen Deposits Detection on Retina Images, http://www.uninova.pt/~atm/md3ri.