A novel deep learning based method for retinal lesion detection

Abstract

Diabetic retinopathy has become one of the most severe complications associated with diabetic retinopathy. Early detection can go a long way to prevent total blindness in the patient. Diabetic retinopathy is characterized by lesions of which exudates and hemorrhages appear prominently. We utilize a region based CNN approach to automatically mark the exudates and hemorrhages in a fundus image. The approach extracts the characteristic region proposals characterizing the disease and successfully mark the lesions with a recall of 90%. We also examine the effects of various color based and affine transform based techniques on the image and obtain a significant improvement in the detection across various sizes and shapes of lesions. We have extensively tested our algorithm both on public databases as well as images captured using the Bosch handheld camera.