Vessels semantic segmentation with gradient descent optimization

Semantic segmentation is a segmentation based on the accuracy of the pixel. Segmentation of retinal vessels aims to differentiate between vessels and non-vessels. There are many conventional methods of retinal vessels segmentation. Conventional methods should perform feature extraction manually. Currently, there has been widely researching on deep learning. Deep learning performs feature extraction automatically during the training phase. One of the methods of deep learning is the Convolutional Neural Network (CNN). The research used three steps i.e creating network layers, training the network, and evaluation of accuracy. We use ten layers of CNN. The CNN architecture consists of Convolution layers, Rectified Linear Unit layers, MaxPooling Layers, Transpose Convolution Layers and Softmax Layer. Furthermore, training the network include the optimization of the training phase. There are three optimization algorithms used for the training phase. The optimization algorithms are Stochastic Gradient Descent with Momentum (SGDM), Root Means Square Propagation (RMSProp) and Adaptive Moment Optimization (Adam). The dataset used for the experiment is Digital Retinal Image for Vessel Extraction (DRIVE). Computation process using the Graphical Processing Unit. Variation scenarios are performed to get the optimal accuracy. The vessels semantic segmentation average accuracy achieved 91.11% - 98.08%.