Semantic segmentation of artery-venous retinal vessel using simple convolutional neural network

Abstract

© 2019 Published under licence by IOP Publishing Ltd. Semantic segmentation is how to categorize objects in an image based on pixel color intensity. There is an implementation in the medical imaging. This article discusses semantic segmentation in retinal blood vessels. Retinal blood vessels consist of artery and vein. Arteryvenous segmentation is needed to detect diabetic retinopathy, hypertension, and artherosclerosis. The data for the experiment is Retinal Image vessel Tree Extraction (RITE). Data consists of 20 patches with a dimension of 128 × 128 × 3. The process for performing semantic segmentation consists of 3 method, create a Conventional Neural Network (CNN) model, pre-trained network, and training the network. The CNN model consists of 10 layers, 1 input layer image, 3 convolution layers, 2 Rectified Linear Units (ReLU), 1 Max pooling, 1 transposed convolution layer, 1 softmax and 1 pixel classification layer. The pre-trained network uses the optimization algorithm Stochastic Gradient Descent with Momentum (SGDM), Root Mean Square Propagation (RMSProp) and Adaptive Moment optimization (Adam). Various scenarios were tested to get optimal accuracy. The learning rate is 1e-3 and 1e-2. Minibatch size are 4,8,16,32,64, and 128. The maximum value of epoch is set to 100. The results show the highest accuracy of up to 98.35%