Y.-H. Wen, A. Bainbridge-Smith, A. B. Morris, Automated Assessment of Diabetic Retinal Image Quality Based on Blood Vessel Detection, Proceedings of Image and Vision Computing New Zealand 2007, pp. 132 136, Hamilton, New Zealand, December 2007. Automated Assessment of Diabetic Retinal Image Quality Based on Blood Vessel Detection Yi-Han Wen 1, A. Bainbridge-Smith 1 and A. B. Morris 2 1 Dept. Electrical and Computer Engineering, University of Canterbury, Christchurch, New Zealand. 2 Christchurch School of Medicine, University of Otago, Christchurch, New Zealand. Email: Andrew.Bainbridge-Smith@canterbury.ac.nz Abstract This paper investigates methods of assessing retinal image quality. This is the first step of a larger research programme for computer assisted screening of diabetic retinopathy images. It is based on automatic detection of blood vessels using two techniques: blood vessel detection and K-means cluster. The algorithms are explained and the results obtained from a sample of 80 images. It was found that both of the techniques did not produce satisfactory results. Keywords: retina, blood vessel, quality assessment 1 Introduction The leading cause of blindness in working age people is due to diabetic complications [1, 2]. Diabetes affects the vascular system and in the eye one such pathology can be observed as localised bulging or pouching of retinal blood vessel walls, called microaneurysms. These aneurysms can easily rupture leading to localised death of surrounding retinal cells, due to the loss of the blood supply. The resulting visual loss brings significant costs, both to the individual and to society. Hence, control of blood sugar is most important, both for controlling the mechanism of the disease, and for mitigating its debilitating consequences. However, if ocular complications occur (termed diabetic retinopathy) specific ophthalmic treatments exist. Early intervention and constant monitoring is essential, with an estimated 90% of visual loss being avoidable if followed [3]. Monitoring takes the form of regular retinal examinations; digital images of the retina are taken and graded by a trained professional, to assess the level of severity. Severity in turn determines the frequency of examinations, between 2 yearly and 3 months, and also when treatment occurs. The Canterbury District Health Board (CDHB) operates such a scheme overseen by a consultant retinal specialist ophthalmologist and ophthalmic registrars working under their supervision. This process is generally followed nationally, with a large quantity of images being obtained and assessed by a small number of professionals. The desire for a national New Zealand Screening Program [4] together with a significant shortage of professionals to read them, has lead to a research programme at the CDHB looking at some form of computer assisted screening. This paper describes work looking at automatic assessment of the quality of retinal images. The purpose of which is to partition, based on quality, the current large CDHB retinal image database. Consisting of over 100,000 full-colour images, acquired using a mydriatic digital camera, this resource is a significant research resource to evaluate image processing algorithms that will be used for computer assisted screening. Images of sufficiently acceptable quality will allow full assessment of the retinal vascular system. This paper continues previous work [5], in which the objective is to discriminate between images of sufficient quality; the basis of this is outlined in section 2. In section 3, two approaches based on blood vessel detection are outlined. The results and our findings on these algorithms are discussion in section 4. Finally, a conclusions are drawn given in section 5. 2 Image quality assessment The most important criteria for determining a good quality image is whether the dark blood vessels can be easily distinguished from the light background, regardless of its background colour. Blood vessel detection is linked to the sharpness or focus of the image. For example, Figure 1 and Figure 1 132
(d) Figure 1: Retinal images. High quality with typical background disc colour, high quality but pale disc, poor quality due to blurring, (d) poor quality due to exposure problems. both show good quality retinal images, where the main vessels from the optic disc and the fine blood vessels near the fovea can be seen easily. Conversely, if the image is of insufficient quality, such as those shown in Figure 1 and Figure 1(d), blood vessel detection near the fovea is difficult due to blurring and other exposure problems. Since blood vessels have the property that the two edges of a vessel always run parallel to each other, matched filter is suitable to use. In [6], the brightness intensity of the cross section of the blood vessel is approximated using a Gaussian function. The model is extended to two dimensions by assuming a vessel has an invariant width over the model length (two-dimensional extension). As the vessels may appear in any orientation, a set of segment profiles, one for each candidate orientation, are used in the matched filtering. In their work, and our s, twelve different set of pixel kernels are used to span the orientation space of the blood vessels. The details for computing the actual values in the kernels are explained in [6]. The algorithm applies the filter set by convolving the input retinal image with all twelve kernels. If the magnitude of the filtered output at a given pixel location is larger than a certain threshold value, the pixel is considered as a part of a vessel. The threshold scheme used to distinguish between enhanced vessel segments and the background is explained in [7]. In this paper we consider two techniques for identifying blood vessels, or improving the contrast of the vessels with respect to the background, ultimately for assessing image sharpness. Image sharpness not the only criteria for assessing image quality, contrast and colour saturation are also important, but not considered in this work. 3 3.1 Assessment algorithms Blood vessel detection The first technique applied is based on the work by Chaudhuri[6]. The technique is based on match filtering and a threshold selection scheme. The matched filter is tuned to detect piecewise linear segments of blood vessels in the retinal images. 133
3.2 K-means colour-space reduction This technique aims to improve the identification of the blood vessels by reducing the number of colours in the image, colour-space reduction. The intention is that a reduced colour space will yield a bigger contrast between the dark blood vessels and the light background. Colour clusters are employed to solve the problem of colour reduction. The captured retinal images are encoded in a 24- bit RGB colour-space, 8-bits per channel. This gives a large theoretical colour-space of 2 2 4, but in practice the image size and red dominance of the images significantly reduces it size. We use here the K-means algorithm to further reduce the colour-space [8]. For example, if the image is to be modelled with just 128 colours, a colour-space of size 2 7 bits, then 128 clusters will be generated in the RGB space to represent the input image. The procedure starts by assigning K numbers of cluster centroids within the RGB colour-space. These centroids, or pixels, are randomly placed, with the maximum spread between them for the best discrimination. The second step is to take each pixel in the input image (input vector) and associate it to the nearest centroid. This simply means that each centroid owns a set of pixels. After all the pixels in the input image have been assigned, the first grouping is completed. The next step is to recalculate the K mean values so that the new position of each centroid will be the average of all vectors that have been assigned to the cluster. Once the new K mean values are determined, a new grouping is defined between the same data set pixels and the newly computed centroids, from which new K mean centroid values are computed. The algorithm iterates in this fashion until the centroid values reach a stead-state. algorithm is summarised below, This simple 1. Choose K number of cluster centroids. 2. Assign each pixel in the input image to the nearest centroid. 3. Define the new positions of all centroids. 4. Go back to step 2 until the positions of centroids no longer change. 4 Results 4.1 Chandhuri s vessel detection A sample of 80 test images were selected from the CDHB retinal image database. The images were manually graded for quality with 40 ranked as acceptable and 40 as unacceptable or poor quality. When a good quality image, such as the one shown in Figure 1, is processed by Chandhuri s algorithm, vessel detection is successful, as shown in Figure 2. However we find there are two different sets of output produced by the algorithm when processing images of poor assessed quality. Consider the two insufficient quality images shown in Figure 1 and Figure 1(d). The performance of the algorithm on these images are shown in Figure 2 and Figure 1. For the blurred image of Figure 1 many of the blood vessels are not detected, as expected. However this is not the case for the poor exposure image of Figure 1(d), yielding the unexpected result in Figure 2. In Figure 2, one can state that all the blood vessels are extracted with success, even the fine capillaries in the fovea. As for the first poor quality image, not many blood vessels were extracted, the expected result, since the blood vessels were hardly recognised from the original image in Figure 1. These two results confirm our working hypothesis. This, in turn, led to the idea that a simple measure of the image energy level of the process image (vessel detected image) was sufficient for discrimination. A scatter plot of the input image signal energy and that of the processed vessel detected image energy is shown in Figure 2(d). However, the result in Figure 2 was not expected. It is obvious that significantly more vessels were apparently detected. Consequently, and as illustrated Figure 2(d), discrimination between acceptable and poor quality images based solely on the energy level of the vessel detected image is not possible. Classification is therefore not possible and other techniques are required. 4.2 K-means colour reduction algorithm The K-means algorithm is process and time intensive, therefore only a subsection of an image was trialled for these results. The algorithm was tested with two different cluster sizes, K = 150 and K = 200. Small values of K are desirable, as each cluster is to be associated as either vessel or background. Results are shown in Figure 3. From Figure 3, it is obvious that the darker blood vessels contrast better when it is compared to the background. However, it is not the case in Figure 3. The reason is that the number of clusters used is quite close to the number that is used to represent the original image. However, it is clear in both figures that the blood vessel boundaries are blurred by the processing. This is the opposite to the desired result of enhancing the edge detection via colour-space reduction. Consequently, this algorithm also fails for classification. 134
(d) Figure 2: Blood vessel detection technique based on Chaudhuri[6]. Applied to the image in Figure 1, Applied to the image in Figure 1, Applied to the image in Figure 1(d), (d) Scatter diagram comparing energy of the pre and post processed images using the detection algorithm, 2 for pre-assessed adequate quality images and for pre-assessed poor quality images. Figure 3: K-means recolouring algorithms. Original image, Colour reduced to K = 150 clusters, K = 200 clusters. 135
5 Conclusions Diabetes is now endemic within New Zealand and the western world[9]. Constant monitoring, allowing early intervention, is essential to mitigate against preventable blindness. However, the large amount of images needed for grading could easily exceed the manual resources currently available for grading purpose. This paper describes techniques for assessing image qualities as a pre-cursor for further image processing analysis. Two approaches where investigated both based on blood vessel detection. The first approach applied the technique proposed by Chaudhuri[6]. This approach worked well with a number of images, extracting blood vessels with success. However in a significant subset of pre-assessed low quality images, many erroneous fine blood vessels were identified. Consequently discrimination between acceptable and poor quality images cannot be made with this technique alone. In summary [6] appears only to produce satisfactory result if the input image is of good quality. This result is, however, not completely negative. In subsequent processing stages of an assisted-screening programme blood vessel detection of good quality images is required and this would suggest Chaudhuri s technique is a likely candidate algorithm for this work. The second approach based on K-means clusters is also an unsatisfactory technique for assisting in assessing image quality. However, given the magnitude of the problem to be solved it is important that a technique be developed. We shall return to the drawing board. New Zealand (IVCNZ), December 2006, pp. 281 286. [6] A. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum, Detection of blodd vessels in retinal images using two-dimensional matched filters, IEEE Transactions on Medical Imaging, no. 3, pp. 263 269, 1989. [7] A. Otsu, A threshold selection method from gray-level histograms, IEEE Transactions on Systems, Man and Cybernetics, pp. 62 66, 1979. [8] J. B. Macqueen, Some methods of classification and analysis of multivariate observations, in Proceedings of the Fifth Berkeley Symposium on Mathemtical Statistics and Probability, 1967, pp. 281 297. [9] PricewaterhouseCoopers, Type 2 diabetes: Managing for better health outcomes. economic report for diabetes new zealand inc. PricewaterhouseCoopers Ltd., Tech. Rep., 2006. References [1] USA Today. (2007) Diabetic retinopathy. [Online]. Available: http://www.healthscout.com/ ency/68/268/main.html,accessed20july2007 [2] N. Z. M. of Health, Diabetes in new zealand: Models and forecasts 1996-2011, New Zealand Ministry of Health, Tech. Rep., 2002. [3] H. Li and O. Chutatape, Automatic location of optic disk in retinal images, in IEEE International Conference on Image Processing 2001, August 2001, pp. 837 839. [4] S. S. S. of New Zealand Ltd., National diabetes retinal screening grading system and erferral recommendations 2005, Save Sight Society of New Zealand Ltd., Tech. Rep., 2005. [5] Y. Kwon, A. Bainbridge-Smith, and A. Morris, Quality assessment of retinal images, in Proceedings of the Image and Vision Conference 136