Exudate segmentation on retinal atlas space

Abstract

Diabetic macular edema is characterized by hard exudates. Presence of such exudates cause vision loss in the affected areas. We present a novel approach of segmenting exudates for screening and follow-ups by building an ethnicity based statistical atlas. The chromatic distribution in such an atlas gives a good measure of probability of the pixels belonging to the healthy retinal pigments or to the abnormalities (like lesions, imaging artifacts etc.) in the retinal fundus image. Post-processing schemes are introduced in this paper for the enhancement of the edges of such exudates for final segmentation and to separate lesion from false positives. A sensitivity(recall) of 82.5 % at 35% of positive predictive value on FROC-curve is achieved. Results are obtained on a publicly available HEI-MED dataset and have been compared to two reference methods on the same dataset showing the competitiveness of the proposed algorithm.