The effects of spatial resolution on an automated diabetic retinopathy screening system´s performance in detecting microaneurysms for diabetic retinopathy

This paper presents the effects of image quality, given by the number of pixels used to define the image. A microaneurysm (MA) segmentation algorithm that has been shown to achieve about 90% sensitivity and specificity for clinical classification using high resolution images of those diabetic patients who present with MAs was applied to low resolution images to assess the effects of lower resolution of sensitivity and specificity. The low resolution (640 by 480 pixels) 45° field of view (FOV) was provided by a non-mydriatic camera. High resolutions images from a mydriatic fundus camera were acquired by digitizing 35 mm color film slides to 1400 × 1200 pixels for a 30° FOV. The image quality of the digitized 35 mm images was considerably better than those from the non-mydriatic camera. Segmentation of microaneurysms (MAs) was performed using the green channel image of the two modalities. The images were contrast enhanced and corrected for uneven illumination. The images were then filtered using a tophat morphological filter and a threshold applied to segment the candidate MAs. Because the retinal vessels are similar in intensity and contrast to MAs, the retinal vasculature was segmented using matched filters to remove the vessel artifacts from the image. Ground truth, which was provided by an ophthalmic analyst, was used in the tuning step to find the bounds of the different intensity and shape features that characterize the MAs. These features were used to distinguish between MAs from other artifactual objects on the image. The best result that could be achieved with the lower resolution images was 70% sensitivity and specificity. We have found that the effects of pixel resolution on an automated segmentation routine to be of significant in obtaining higher sensitivity and specificity.