Grading Quality of Color Retinal Images to Assist Fundus Camera Operators

color retinal image, quality assessment, convolutional neural network

Suitable image quality is a prerequisite to ensure accurate diagnosis or person recognition by color retinal images. Many factors during image acquisition, transferring and storing can result in poor quality retinal images. Poor quality images not only increase the possibility of wrong diagnosis, false acceptance, or incorrect identification but also increase diagnosis or recognition time. Therefore, retinal image quality assessment has become an important research topic. In general, only one color channel (most of the time either green or grayscale) is used to assess the quality of retinal images ignoring the quality of other channels. However, all image channels carry complementary information. In this paper, we propose a quality assessment approach for a colored retinal image to assist a fundus camera operator to judge the image quality. In our approach, we analyze the histogram of pixel intensity and uniformity of illumination, as well as check the presence of two main anatomical structures, optic disc, and central retinal blood vessels, in all color channels (i.e., red, green and blue) as well as in grayscale format.We show the effectiveness of our approach by grading 3090 color retinal images of five publicly available retinal databases.