Classification of brain haemorrhage in head CT scans using deep learning

Abstract

A brain haemorrhage is defined as a bleed in the brain tissue and it is the third leading cause of mortality across all ages and are caused either by a haemorrhagic stroke, or a significant blow to the head. One of the most commonly used diagnostic tools for patients being treated for a brain injury or patients with symptoms of a stroke or rise in the intracranial pressure is non-contrast Computed Tomography (CT) scan. Computer-Aided Diagnosis (CAD) systems have been developed and were introduced to aid radiologists and professionals in their decision making. Deep Learning CAD systems were not highly researched before, but due to recent advancements in technology, deep learning algorithms have become more popular, and are now being researched for their applications in medical imaging. This study utilises deep learning models to develop a computer aided diagnosis (CAD) system to classify the different type of haemorrhages in head CT scans. The system was designed in such a way that it builds up on the work done on the final year projects of Mr John Napier and Ms Kirsty Sant; where Mr Napier’s work focused on developing a system that can detect the presence of a brain haemorrhage in CT scans; and Ms Sant’s work involved using a machine learning technique to classify the brain haemorrhages, based on the intensity, shape and texture features. Deep learning architectures, namely ResNet, DenseNet, and InceptionV3 architectures were analysed, in order to find the best performing architecture to classify the different types of brain haemorrhages from head CT scans. Moreover, a linear Support Vector Machine, was also built in order to be able to compare the performance of these architectures with it. The dataset was obtained from the General Hospital of Malta, totalling to 64 anonymised brain haemorrhage CT scans, 58 of these were used for training the deep learning models, and the remaining 6 cases were used to test the models. Each of the architectures were executed for 100 epochs, and the overall training accuracy was 0.1786 for ResNet, 0.2976 for DenseNet, 0.3690 for InceptionV3 and 0.6083 for the linear multiclass support vector machine.