PREDICTING CHANCE OF SURVIVAL FOR PATIENTS WITH TRAUMATIC BRAIN INJURIES

Abstract

Traumatic Brain Injuries (TBI) are considered one of the leading causes of death worldwide. TBI is defined as a severe head trauma that disrupts the brain's regular activity. Penetrating Traumatic Brain Injuries (pTBI) are extreme cases of TBI in which an object penetrates the outer membrane of the brain. Even though these pTBI injuries are less prevalent than closed head injuries, they often result in a worse prognosis. Thus, the severity of pTBI makes the accurate prediction of the survival rate of pTBI patients a crucial step in applying adequate medical treatment and allocating the needed resources for these patients. In clinical practice, experts predict the outcome of affected patients based on their medical experience and the standardized Glasgow Outcome Scale (GOS). On the other hand, conventional automated survival rate prediction methods that use traditional machine learning (ML) approaches have been limited to estimating a survival score for TBI patients based on a set of clinical input features such as age, GOS, and laboratory variables. Those previous methods rely on handcrafted features only. They do not consider the analysis of medical images that provide critical information for assessing the patients' survival risk. In this thesis, we propose a new problem formulation to introduce the use of deep learning frameworks with brain computed tomography (CT) scans for pTBI prognosis. Our aim is to automatically extract feature representations directly from the scans and improve pTBI patients' prognosis. We design a data-processing pipeline combining brain extraction and CT scan registration methods to reduce the variability between patients. To address the challenge of very limited training data, we explore two new transfer learning approaches: A 3D deep learning (DL) architecture (V-PBI) based on 3D convolutions with a DL model called V-Net, and a 2D hybrid architecture combining Convolutions Neural Networks with Bidirectional Long Short Term Memory networks (CNN-BiLSTM). Both architectures aim at sharing domain knowledge to improve the prognostic task from small training data sets. The 3D V-PBI uses previous knowledge from segmentation tasks to stabilize the model's performance on the survival rate prediction tasks. The 2D CNN-BiLSTM model decomposes the 3D CT scan into 2D slices, and uses pre-trained CNN networks to extract the spatial information from these slices. The BiLSTM is designed to capture the sequential relation between the different slices. Multiple experiments were conducted on a data set of 125 patients provided by the University of Chicago. The data set consisted mainly of patients suffering from gunshot injuries in the head. The two proposed architectures are compared to a baseline 3D CNN. The results showed that the proposed V-PBI model outperformed the other architectures, and achieved an average accuracy of 78.3% and an average area-under-the-curve (AUC) of 0.816, while the 2D CNN-Bi-LSTM achieved an average accuracy of 76.7% and an average AUC of 0.777, and the 3D baseline model achieved an accuracy of 73.3% and an average AUC of 0.737.