Mathematical Morphological Processing for Retinal Image Analysis

Abstract

Diabetic retinopathy is the leading cause of the blindness in the western world. Digital retinal imaging with remote image evaluations is a promising new solution to accurately and precisely stage patients conveniently. The spot lesion detection is the primary step. Based on the mathematical morphology, we discussed two lesion extraction algorithms. To avoid over-segmentation, inner and outer markers are introduced into the marker controlled watershed segmentation method. Gradient image is generated by multi-color channels. Marked lesions can be successfully extracted with clear boundaries. The second method, the adaptive multiscale morphological processing, is a novel procedure to efficiently extract spot lesions in the fundus image. The relative contrast of lesions with the surrounding background is used as criteria, which are similar to the human vision property. Entropy-based thresholding can well distinguish lesions. Post processing removes misclassified areas and produces vascular tree. Both algorithms have been tested in the Clemson University's database.