Automated brain hemorrhage segmentation using a deep neural network

Abstract

In this work, an automated brain hemorrhage segmentation using the deep neural network (DNN) has been proposed. A total of 157 patients, ages 23–69, are collected from different hospitals and laboratories, and the total number of 3038 images is collected. Initially, ground truth images are generated using preprocessing steps and validated through 3 experts having more than 20 years of experience. Data augmentation is performed to create a virtual image for training DNN models, and a convolutional neural network-based U-Net model is used. Extensive experimentations were performed using 6000, 8000, 10,000 samples with 20%, 30%, and 40% dropout, respectively. The same combination of experiments has been performed for ReLU and tanh activation function. After the investigations, it has been found that the maximum training accuracy of the proposed model is 98.8% using the ReLU activation function. Similarly, 93.3% of accuracy has been achieved by a tanh activation function.