Automatic detection of epidural hematoma on the brain by using image processing techniques

Abstract

In modern medicine, detection of a hematoma depends mostly on the utilization of imaging techniques such as Computerized Tomography (CT) and Magnetic Resonance Imaging (MRI). Detection of brain injuries on brain images is a convoluted and challenging task for the radiologist. Challenges mostly occur due to the nearness and nested nature of different types of brain tissues. The diversity of cerebrum structures increase the algorithm complexities required to detect and segment the injury region. Traffic accidents and falls are the two most frequent causes of traumatic brain injury (TBI), falls being slightly more prevalent. According to American Speech Language Hearing Association, every year, at least 1.7 million (TBIs) occur in the United States and Epidural Hematoma (EDH) cases constitute more than 45 % of the TBI. Studies show that the overall incidence rate for the TBI is approximately 300 per 100,000. Therefore, TBI detection and management is an important health care problem. In TBI detection and management, the goal of all imaging techniques is to locate the injury site and predict its progression. Each imaging technique has advantages for describing specific types of TBI. However, due to its high detection speed, availability, and high sensitivity, CT is always the primary choice when dealing with TBI. MRI is used less frequently because it requires long image acquisition times, has higher costs, is very sensitive to patient movements ,and is not suitable for patients with claustrophobia. In this thesis, we aim at detecting, i.e., marking the border and measuring the size of, EDH regions in CT scans of the brain. Proposed system contains many image processing operations, including image segmentation and binary morphology. Gaussian Mixture Modeling (GMM) segmentation is used as the primary method and its results are compared with the well-known k-means segmentation. All the codes in this thesis have been developed in MATLAB software environment and the experimental data consisting of 37 CT images of EDH (or bleeding) cases was obtained from a publicly available dataset named as "https://radiopaedia.org/". Professional help is received from an expert radiologist to select these images and build the ground truth (i.e., the actual boundary information of EDH regions) for them. The doctor has visually inspected all of 37 images and marked the boundaries of bleeding regions by red color using an image editing software. Then, the proposed algorithms were tested on these images, the obtained results were compared with the ground truth provided by the expert, and finally error rates were calculated. Obtained results are very promising and encouraging; on average, proposed GMM based segmentation method yields 85 % detection rate compared to 83 % of k-means method.