Automatic detection of vascular lesions of the retina using a localized adaptive thresholding approach

Early detection of vascular lesions such as exudates, hemorrhages, and tumors in the retina is important in early detection of diabetes, hypertension, and cancer. For example, diabetic retinopathy is the leading cause of blindness in adults in the United States and the presence of exudates in fundus imagery is an early sign of diabetic retinopathy. In this paper we present a novel technique to automatically detect exudates, hemorrhages, and tumors in fundus imagery that is robust against spatial and temporal variations of background noise. The detection threshold is adjusted dynamically, based on the local noise statics around the pixel under test in order to maintain a pre-determined, constant false alarm rate (CFAR). A pre-processing step is introduced to accommodate the detection of bright lesions (exudates) as well as dark lesions (hemorrhages and tumors). The CFAR detector addresses the challenge of detecting bright lesions in RGB and multispectral fundus imagery where the background clutter often exhibits variations in brightness and texture. These variations present a challenge to common, global thresholding detection algorithms and other methods. Performance of the adaptive threshold CFAR algorithm is assessed using a publicly available, annotated, diabetic retinopathy database including the Retina Image Bank. Performance of the CFAR detector proves to be superior to techniques such as Otsu thresholding.