Classification of Intracranial Hemorrhage Using Deep Learning

Abstract

Intracranial hemorrhage (ICH) is a lethal brain state having internal bleeding caused by ruptured blood arteries within the skull. This can be caused by physical damage to the head or disease-related to structural weakening of veins. If ICH is not appropriately detected and treated in a timely manner, it can result in disability or death of the concerned person. Based on which region it is found in the brain, ICH is subdivided into five kinds: intra-ventricular bleed hemorrhage (IVH), intra-parenchymal bleed hemorrhage (IPH), subarachnoid bleed hemorrhage (SAH), epidural bleed hemorrhage (EDH), and subdural bleed hemorrhage (SDH). Computed Tomography (CT) scan is preferred modality over other imaging modalities such as Magnetic Resonance Imaging, X-ray and other diagnostic tools, due to its availability and rapid collection time, for the initial examination of ICH. A skilled radiologist examines these CT scans to find the occurrence of ICH and its location. However, this method of diagnosis is reliant on the availability of a subspecialty-trained neuro-radiologist. It is also subjective and takes huge time to examine manually the CT scan. Hence, there is a need for the development of faster and accurate computer automated methods for the diagnosis of ICH and its type. Advancements in machine learning models along with better computational power allows learning of appropriate features from the data, which was earlier explored from well-known feature extraction methods for better representation of the data. However, deep learning models such as Long Short-Term Memory (LSTM)and Convolutional Neural Network (CNN) requires huge number of samples for training the model from scratch, which will take huge computation time. In literature, the researchers have used pretrained CNN models such as VGG16, VGG19, DenseNet, Alexnet, Resnet etc. However, these models need to be fine-tuned on their data to achieve good performance. Thereby, fewer number of trainable parameters will be considered to build the decision model to overcome the problem of overfitting. Here, I have used DenseNet121, VGG16 and VGG19. The performance of the model is evaluated in terms of F1 score and Classification accuracy on a publicly available dataset.