HRUNET: Hybrid Residual U - Net for automatic severity prediction of Diabetic Retinopathy

Abstract

ABSTRACT Diabetic Retinopathy (DR) is a vision-threatening illness that affects diabetics and is the common cause of vision loss among working-age people. Many technologies have been developed to classify DR reliably at an early stage. In this paper, a new hybrid deep-learning algorithm for automatic severity detection in DR is proposed. The proposed Hybrid Residual U-Net (HRUNET) is applied for the contraction and extensive paths from the down-sample to the up-sample to resolve the limitation of plain skip connections in the segmentation of DR. Further, every convolution layer is replaced by a residue module with varying sizes of kernels, and the residual connection is employed in each module to make the network broader without gradient disappearing. These residual modules are linked amid the network to increase the depth of the network. To avoid gradient vanishing, a batch normalization layer follows all convolutional blocks except the bottleneck layers. HRUNet achieves an accuracy of 94% and 91% on the Asia Pacific Tele-Ophthalmology Society (APTOS) and KAGGLE Datasets, respectively. The proposed HRUNET model has a total of 6,315,732 trainable parameters, which are comparatively lower than state-of-art methods such as VGG16, VGG19, and ResNet50 and have higher accuracy on both the datasets.