Purpose:
The challenges for obtaining improved retinal image quality includes overcoming such factors as uneven retinal illumination, the noise introduced by the optics and electronics of the camera, and the anatomical variances across patient population, including pupil size, intraocular media clarity and retinal pigmentation. We will demonstrate an innovative technique in image enhancement based on the modeling of human visual mechanisms that can yield significant improvement in retinal images collected by current retinal digital cameras. By improving image quality better diagnosis can be made both manually and using computer–based methods.
Methods:
The method is based on a Multi–Scale Spatial Decomposition (MSSD) algorithm inspired from Human Visual System mechanisms to enhance the visual appearance of color images. In particular, we apply the MSSD on retinal images to improve the manual and automatic screening of such images for retinal abnormalities. The MSSD algorithm is a multi–step process in which each step improves one aspect of the image. Image quality metrics which included entropy, contrast, and spatial frequency were calculated in order to quantify the improvements after applying the MSSD algorithm. Manual assessment by an expert grader was also performed.
Results:
We applied the MSSD algorithm on 100 low resolution 480x640 images provided by Joslin Vision Network. The results obtained by our approach presented a significant improvement of the quality of the studied images. Tables 1 show the increase in the image quality metrics in the green channel after the application of the MSSD algorithm. Usually, this plane carry most of the structure of the image.
Conclusions:
We have demonstrated an automatic computer–based algorithm to improve the quality of retinal images automatically. The increase in quantitative metrics of quality suggests increases in contrast, information content and details. We are currently developing a preference matrix with the help of human graders to determine the importance of each quality metric in the assessment of image improvement.
Keywords: diabetic retinopathy • image processing • contrast sensitivity