Segmentation of Hard Exudates And Hemorrhages from Diabetic Retinopathy Images Using Residual U-Net with Squeeze and Excite Blocks

Diabetic Retinopathy (DR) is a common eye disease worldwide and usually found in diabetes patients. Deep learning based algorithms have recently produced promising findings for the automatic detection and segmentation of DR lesions from fundus images. In this paper, we present an approach for segmenting DR using a Residual U-Net model. Here, we have incorporated Squeeze and Excite (SE) blocks into the Residual U-Net architecture to segment various DR diseases from fundus images. SE blocks are attention techniques that dynamically recalibrate the feature maps of a neural network. These enable the network to prioritize useful features while suppressing less important ones, hence boosting the model’s ability to discriminate. We intended to enhance the segmentation performance by enabling the model to better capture spatial and contextual information by integrating SE blocks into the Residual U-Net architecture. The developed model architecture was trained and tested on fundus image DR datasets. The results obtained by the developed model for hard exudates and hemorrhages segmentations were better as compared to U-Net and its variants.