RIMNet: Image Magnification Network with Residual Block for Retinal Blood Vessel Segmentation

Abstract

To understand minute segments of a soft copy image, we zoom the image or magnify it. The underlying motivation of this work is to enable the machine to learn to magnify an image to capture very small and fine details of it. Retinal Blood Vessel Segmentation is such an area where segmentation of small and fine vessels is needed. In this paper, we have designed RIMNet (Image Magnification Network with Residual Block) to accurately segment retinal blood vessels from fundus images. This network follows the architectural pattern of the convolutional neural network(CNN) based encoder-decoder model for image segmentation but in an unconventional way. The encoder part of the model up-samples or magnifies an image to understand delicate image features and the decoder part down-samples the magnified image to get back into the original resolution for better segmentation. The model is trained, validated, and tested on the DRIVE dataset. It shows a sensitivity of 76.83 %, specificity of 98.08%, test accuracy of 94.78%, F1-score of 77.39% and AUC_ROC of 86.96% on the dataset which is pretty competitive to the state-of-the-art retinal blood vessel segmentation model. Here is the link to the repository of code and data for this work https://github.com/Sufianlab/RIMNet.