Optimizing Sharpness Measure for Bright Lesion Detection in Retinal Image Analysis

Abstract

Due to the spherical shape nature of retina and the illumination effect, detecting bright lesions in a retinal image is a challenging problem. Existing methods depend heavily on a prior knowledge about lesions, which either a user-defined parameter is employed or a supervised learning technique is adopted to estimate the parameter. In this paper, a novel sharpness measure is proposed, which indicates the degree of sharpness of bright lesions in the whole retinal image. It has a sudden jump at the optimal parameter. A polynomial fitting technique is used to capture this jump. We have tested our method on a public available dataset. Experimental results show that the proposed unsupervised approach is able to detect bright lesions accurately in an unhealthy retinal image and it outperforms existing supervised learning method. Also, the proposed method reports no abnormality for a healthy retinal image.