Deep learning for biomedical image analysis

Abstract

During recent decades, researches have shown great interest to machine learning techniques in order to extract meaningful information from the large amount of data being collected each day. Especially in the medical field, images play a significant role in the detection of several health issues. Hence, medical image analysis remarkably participates in the diagnosis process and it is considered a suitable environment to interact with the technology of intelligent systems. Deep learning (DL) has recently captured the interest of researchers as it has proven to be efficient in detecting underlying features in the data and outperformed the classical machine learning methods. The main objective of this thesis is to prove the efficiency of deep learning techniques in tackling some of the important health issues we are facing in our society, through medical imaging. The first case study covers a dermatology related health issue associated with increased morbidity and health care costs. Managing pressure injuries appropriately is increasingly important for all the professionals in wound care. Using 2D photographs and 3D meshes of these wounds, collected from collaborating hospitals, our mission is to create intelligent systems for a full non-intrusive assessment of these wounds. Five main tasks have been achieved in this specific study: a literature review of wound imaging methods using machine learning techniques, the classification and segmentation of the tissue types inside the pressure injury, the segmentation of these wounds and the design of an end-to-end system which measures all the necessary quantitative information necessary for an efficient assessment of pressure injuries, and the integration of the assessment imaging techniques in a web-based application.. The second task of this thesis examines a neurobiological learning disability: dyslexia. Our main objective was to use functional magnetic resonance imaging, a brain scanning modality, to classify children with dyslexia. A novel approach based on the creation of brain activation volumes and classifying them using a 3D deep learning architecture has been presented in this part of the study. This dissertation is presented as a Ph.D by publication contribution. It is presented as a collection of four published articles.