Brain hemorrhage detection using hybrid machine learning algorithm

Abstract

Machine learning (ML) helps computers learn and program data without humans’ help. According to data scientists, machine learning can extract 60% high-quality information, reduce the cost up to 46%, and increase operation speed by approximately 48% [1]. Recently, there has been successful implementation of machine learning in data analysis, computer vision, computer-aided diseases (CAD), and many more fields. Machine learning is broadly used in the medical industry because of its processing power for image data and pattern recognition quality. The image processing power of machine learning can be used in medical images to classify the brain images automatically. Segmentation and classification of brain image can provide valuable information and quantitative assessment of lesions which can be used for treatment strategies and predicting patient condition (Kamnitsas et al., 2017). According to research [2], an estimated 64-74 million people in the world are affected by traumatic brain injury every year. It affects the lives of nearly every one out of six persons. In our proposed system, we will use a hybrid approach of multiple machine learning algorithms together for the classification of CT brain images and diagnose brain disorders and diseases like brain hemorrhage. Some ML algorithms such as different 3D Convolutional Neural Networks (CNN) , AlexNet, DenseNet121, GoogleNet and some other models like Multilayer Perceptron Model (MLP), Support Vector Machine (SVM) and Random Forest (RF) have been applied successfully in this field in the past. Modifying previous methods, we want to build a hybrid machine learning algorithm by combining different CNN models like VGG-16, VGG-19, Random forest and Multilayer Perceptron (MLP) classifiers for detecting brain hemorrhage. We have used the VGG-16 and VGG-19 model to derive image features from the CT brain images and Random forest classifier and MLP classifier for testing the accuracy of our model. To test the efficiency of our system, we have used CT brain image datasets from Kaggle. The CT brain imaging data will be the input of our model and our model will detect brain hemorrhage and classify them into one of six classes: Epidural, Intraparenchymal, Intraventricular, Subarachnoid, Subdural and No Hemorrhage. Using our hybrid approach the best accuracy we achieved was around 97.24% using a combined approach of VGG-16 and Multilayer Perceptron classifier. Also we used Explainable AI to explain the prediction of the hemorrhagic classes.