Inception V4 Convolution Neural Network towards Detection and Classification of Brain Hemorrhage using CT Images

Brain Hemorrhage is occurs due to obstruction of blood vessels in the brain. It causes high mortality and disability across the human population in the world. Due to advancement of the technologies in recent years, numerous researchers employs computer vision approaches and artificial intelligence approaches in form of machine learning and deep learning to the detect and classifies brain hemorrhages. Despite of many advantages, those approaches leads to several approaches in terms of computation complexity and class imbalance issues. To mitigate those challenges, new inception v4 convolution neural network architecture is designed in this article to detect and classify the brain hemorrhages using CT images from RSNA Brain CT hemorrhage benchmark dataset. Initially inception V4 convolution neural network composed of three inception modules process the CT images towards feature extraction, segmentation and classification. Inception module is capable of extracting the spatial and temporal features simultaneously in the convolution layer using multiple kernel size in parallel which reduce size of the network. Further it uses factorization to reduce the parameter of the convolution which is composed of the disease and non- disease feature to represent segmentation. Extracted features are employed to max pooling layer to extract the high level features of the image which is considered as disease segment of the image. Finally high level feature map is projected to fully connected layer composed of the activation function through ReLu and softmax function through support vector machine classifier. Extensive experimental analysis of the proposed inception v4 convolution neural network is performed using RSNA 2019 Brain CT Hemorrhage Challenge dataset on python environment. Dataset is partitioned into training and testing dataset. Experimental analysis is performed using training data and performance analysis of the model is evaluated using cross fold validation of the test data to determine the accuracy of the model against state of art approaches. On evaluation, it is proved that proposed model achieves 98.7% accuracy in detecting and classifying brain hemorrhages with increased processing speed and reduced classification error.