Detection and classification of brain haemorrhage

Abstract

A brain haemorrhage is a rupture of the blood vessels within the brain, and is very often life threatening. There are a number of types of blood haemorrhages, these include intracerebral, intraventicular, subarachnoid, subdural, and epidural haemorrhages [25]. Diagnosis tools used by medical experts to identify the type of pathology, include the use Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) scans, lumbar puncture, or cerebral angiography [11]. In this study we focus on the use of CT scans as diagnosis tools. The detection and localisation of a brain haemorrhage is highly time critical, as the longer a case goes undiagnosed, the higher the possibility of a fatality [13]. In this study we develop an automated software tool to aid radiographers in classifying the type of bleed present in a series of CT scan slices, and hence localising it. This is accomplished through the design of a Computer Aided Diagnosis (CAD) system based on a deep learning algorithm. Specifically, this study explores the use and implements a three dimensional Convolutional Neural Network (3D CNN). A CNN is a network of layers which reduce an image to its most basic features making classification easier. The convolutional layer is the layer which translates the image to usable data, it scans small sections of the image and assigns them to different filter classes. The same holds for a 3D CNN, except in this case the kernels move through three dimensions of data and produce three dimensional activation maps. When building a CNN architecture it is always best to start as small as possible and gradually expand, increasing layers and units, until the validation error stops improving. Architectural optimisations were also used to improve computational performance. It also explores the use of transfer learning to implement a binary classification, and comparing the performance of DenseNet [19], ResNet [15], InceptionV3 [32], and InceptionResNet [31] network architectures. To train and test the multi-class 3D CNN solution, this study makes use of a dataset of 139 cases, where each case holds multiple slices of CT scan images. In preparation, each image is reduced to a 128_128 pixel image and stacked, so that at the end we obtain 143 stacked 3D images with a shape of 100 _ 128 _ 128 _ 1, where 1 is the greyscale channel and 100 is the stack height. In cases that did not contain precisely 100 images, these were truncated or padded accordingly. During this process image augmentation was also done so as to obtain another 65 images for each training case. The dataset was split 80% training and 20% for testing. The divide was carried per classification rather than as a whole, to make certain of a proportionate distribution. The derived results obtained were relatively poor for the multi-class classification solution. The binary classification tool made use of a dataset obtained from Kaggle [22], which was made up of a hundred normal brain scan slices and a hundred CT image slices diagnosed with brain haemorrhage. Data augmentation was used in order to supplement the small size of the dataset. This tool showed more favourable results with a maximum AUC obtained of 0.9750.