Segmentation of Microaneurysms for Early Detection of Diabetic Retinopathy using MResUNet

Diabetic retinopathy is a disease caused by complications of chronic diabetes and can lead to blindness. Therefore, automatic detection of microaneurysms on retinal fundus images as an early warning system is very necessary and challenging since the size of the microaneurysm is very small compared to the anatomical structures of the retina such as blood vessels and other elements, and hemorrhages as other red lesions. In this study, we propose a novel deep learning network that modifies UNet using residual units with modified identity mapping (MResUNet) to perform microaneurysm segmentation. The purpose of identity mapping modification using convolutional layers and batch normalization is to enrich features and add Residuals in UNet to overcome feature degradation as the network becomes deeper. The mean weighted loss function is used in training to solve the problem of pixel imbalance between the microaneurysms and the background. The proposed architecture is evaluated using the IDRID and DiaretDB1 datasets. The experimental results show that the proposed architecture (MResUNet) achieves higher sensitivity values of 61.96% and 85.87%