Machine Learning Applied to the Detection of Retinal Blood Vessels

The field of ophthalmology (the study of the eye) has increasingly turned to medical imaging to play an important role in diagnosing diseases. Widespread medical conditions can be identified with only pictures of the eye using computer automated processes. Determining the segmentation of the circulatory system in the eye is difficult for doctors as the task of distinguishing blood vessels by simply observing retinal images has proven to be challenging without the aid of technology. Several morphological features of retinal veins and arteries, like diameter, length, branching angle, and tortuosity, have diagnostic relevance and can be used to monitor the progression of diseases [1]. This paper details the process and results of an attempt to improve upon the accuracy of retinal image segmentation to aid doctors in diagnosing diseases of the eye using supervised (support vector machine) and unsupervised (modified k-nearest neighbor) machine learning algorithms. INTRODUCTION: Many common medical conditions associated with the eye can be efficiently diagnosed by doctors through the observation of retinal images. This process has the potential to improve through the application of machine learning techniques; altered images that highlight the blood vessel patterns increase doctors’ ability to correctly diagnose various medical conditions such as diabetes, hypertension, and Arteriosclerosis [2]. Furthermore, it may be possible to entirely automate the detection of eye diseases. One way to alter retinal scans to distinguish the blood vessels from the rest of the image is to establish a classification problem in which each pixel is labeled as representing a blood vessel (positive) or representing any other part of the eye image (negative). Various features are available for an algorithm of this type: RGB values of each pixel, pixel location, overall curvature, shading, contrast, and many more. Existing machine learning methods approach this problem through the use of Support Vector Machines (SVMs) with feature vectors comprised of some function of the RGB values of pre-processed image pixels among other variables. Currently, the industry standard is approximately 75% accuracy of each positive image pixel [3] compared to hand-drawn blood vessels carefully constructed by experts. This paper documents the construction and application of an SVM that nearly matches this industry standard generalization error using publicly available data. In addition, an unsupervised k-Nearest Neighbor (k-NN) algorithm is developed to streamline the production of segmented images with the ultimate goal being complete automation of the eye disease diagnosis process. DATA: Retinal images publicly available through Clemson University [4] were used to train and test machine learning algorithms to detect blood vessels. A variety of different blood vessel patterns, image lighting, and eye size were represented by these images. The lack of consistency displayed by these retinal scans reveals the difficulty in distinguishing blood vessels and non-blood vessels. Figure 1 below shows a fairly typical example of a retinal scan—the blood vessels are slightly darker than the rest of the image. Figure 1. Unprocessed Retinal Scan In addition, this data set generously provided images hand-drawn by expert ophthalmologists. These images contain the expert’s estimation of blood vessel location; training labels were generated for each image using its corresponding expert-drawn image. For the purposes of this study, the expert labels were considered to be the ground-truth locations of blood vessels in the given retinal scans. This data set yielded twenty images and associated labels. Figure 2 below shows the expertdrawn image for the retinal scan shown in figure 1: Figure 2. Expert Drawing used for Training Labels