Automatic Screening of Fundus Images for Analysis of Retinal Blood Vessels

抄録

Screening of fundus images is an important phase for an automatic analysis of fundus image, where they are classified into “good” or “bad” fundus images. The convolutional neural network (CNN) is a typical classification method, but with use of CNN, we cannot decide which feature contributes to the resulting classification. The purpose of this paper is to propose a screening method, keeping accuracy of CNN, to classify the fundus images into “good” or “bad” images using the features that can explain the classification results. In the proposed method, the gray level co-occurrence matrix (GLCM) together with other color space statistics of RGB and HSV are employed. Accuracy rate of the proposed method was almost the same as that of CNN, but the extracted features by the proposed method could explain the classification results. The features employed in the proposed method could give the reasons of the “bad” fundus images.