Deep CNN based microaneurysm-haemorrhage classification in retinal images considering local neighbourhoods

Abstract

ABSTRACT Retinal image characteristics can be utilised for early diagnosis of Diabetic Retinopathy (DR). The earliest symptom of DR is the presence of microaneurysm and haemorrhage in retinal fundus images. A computerised classification system can increase the effectiveness of the large volume screening process of retinal images. In this paper, a deep convolutional neural network-based pixel classification approach has been presented where the models, trained on online public datasets, are used for symptom level classification of the fundus images, collected from patient data of a local state hospital. The method achieves average values of sensitivity of 0.4556, specificity of 0.8395 and accuracy of 0.8341 on the local dataset. The CNN did not require exhaustive training with a large number of images as symptom level training was performed on annotated overlapping patches. The average classification time is 0.7275 sec/image. The results in terms of classification metrics and in terms of execution time requirement are very encouraging when compared with different recently developed classification methods and as per the simplicity of the method is concerned.