Automatic detection of pathologies in the human brain

Abstract

Identifying and classifying pathologies from brain Computed Tomography (CT) images is a critical, yet time consuming task, performed manually by medical experts. Such a repetitive task leads to tiredness, making the physician prone to human error. Automating the identification and classification of pathological areas will assist radiologists, reducing the total time taken for diagnosing a patient, whilst reducing the possibility of erroneous patient diagnosis. In this project, we will investigate machine learning algorithms and image processing tech- niques in order to create a generic method for the automatic detection of pathological areas in CT images of the brain. A collection of 80 CT brain scans were collected and manually seg- mented. Image processing techniques are used upon a 3-dimensional CT scan, preparing it for eventual extraction of simple, computationally non-intensive features from every hemisphere in each 2-dimensional slice of the 3-D volume. Features extracted from the images provided give rise to a data set, from which a decision tree classification algorithm will learn. Given a balanced dataset, the system was able to record above 90% recall on test data with above 70% accuracy, however with precision of just 30%. We conclude that further and more accurate pre-processing is required for the system to extract consistent features, which, coupled with an increase in training data will boost results, possibly comparing to human experts.