Splat feature classification: Detection of the presence of large retinal hemorrhages

Abstract

Reliable detection of large retinal hemorrhages is important in the development of automated screening systems which can be translated into practice. In this study, we propose a novel large retinal hemorrhages detection method based on splat feature classification. Fundus photographs are partitioned into a number of splats covering the entire image. Each splat contains pixels with similar color and close spatial location. A set of distinct features is extracted within each splat. By learning properties of splats formed from blood vessels, a classifier was trained so that it can distinguish blood splats from non-blood splats. Once the blood splats, i.e. vasculature and hemorrhages, are separated from the background, the connected vasculature was removed and the remaining objects considered hemorrhage candidates. Our approach had a satisfactory performance on a test set composed of 1200 images compared to a human expert.