Automatic Classification of Exudates in Color Fundus Images Using an Augmented Deep Learning Procedure

Automatic classification of hard and soft exudates in color fundus images is very helpful for computer-aided diagnosis of retina related diseases, such as diabetic retinopathy (DR). In this study, we developed a novel method for this purpose based on the emerging deep learning technology known as convolutional neural networks (CNNs) by leveraging its strength of explicitly extracting the underlying image textures. We specifically investigate whether the emphasis of the image characteristic within an exudate spot could improve the classification performance. To verify this, we collected a database of fundus image that contains soft and hard exudates. The exudate regions were cropped from fundus images. There are a total of 550 cropped image patches (275 hard and 275 soft) with a fixed dimension of 128x128 pixels. These patches were further thresholded to exclude image background, resulting in another version of image patches merely containing exudate regions. Each version of image patches was randomly divided into 440 for training and 110 for testing, and then fed into the developed deep learning network in a separate or combinatorial way. Experimental results showed that the classification accuracy of this method was 93.41% when the thresholded version of the dataset was used as an augmented learning procedure, as compared to 90.80% and 87.41% when the original and background excluded datasets were used for training, respectively. This suggests that the augmented CNN can provide more accurate classification performance when the region-of-interest (ROI) and the original images were integrated.