An Efficient Automated Corneal Ulcer Detection Method using Convolutional Neural Network

In this paper, we propose an automated system to detect corneal ulcer disease of eye from the facial image that is taken by general-purpose digital camera using convolutional neural network (CNN). Our proposed method first detects and segments the eye part of face using Haar Cascade Classifiers. Then the method detects whether the corneal disease exists or not by CNN. If corneal ulcer disease exists, our method automatically segments the ulcer area. At first, the outer skin area of the eye is removed using the GrabCut method and segment the iris and sclera region by applying Hough gradient and active contour techniques. Finally to localize and segment ulcer area active contour technique is used and measure the ratio of ulcer size. Here erode and dilate operations are applied on iris image to polish the corneal ulcer segmentation results. For training, the available dataset is limited in size. Thus we apply several augmentation techniques such as random rotation, flip, shift, brightness, noise, and random zoom. From the result we see augmented result is superior to the results that is applied in original images. The experimental results of CNN model achieved an accuracy rate of 99.43% with sensitivity 98.78% and specificity 98.60% by augmentation techniques.