Machine Learning Methods for Image Analysis in Medical Applications, from Alzheimer's Disease, Brain Tumors, to Assisted Living

Abstract: Healthcare has progressed greatly nowadays owing to technological advances, where machine learning plays an important role in processing and analyzing a large amount of medical data. This thesis investigates four healthcare-related issues (Alzheimer's disease detection, glioma classification, human fall detection, and obstacle avoidance in prosthetic vision), where the underlying methodologies are associated with machine learning and computer vision. For Alzheimer’s disease (AD) diagnosis, apart from symptoms of patients, Magnetic Resonance Images (MRIs) also play an important role. Inspired by the success of deep learning, a new multi-stream multi-scale Convolutional Neural Network (CNN) architecture is proposed for AD detection from MRIs, where AD features are characterized in both the tissue level and the scale level for improved feature learning. Good classification performance is obtained for AD/NC (normal control) classification with test accuracy 94.74%. In glioma subtype classification, biopsies are usually needed for determining different molecular-based glioma subtypes. We investigate non-invasive glioma subtype prediction from MRIs by using deep learning. A 2D multi-stream CNN architecture is used to learn the features of gliomas from multi-modal MRIs, where the training dataset is enlarged with synthetic brain MRIs generated by pairwise Generative Adversarial Networks (GANs). Test accuracy 88.82% has been achieved for IDH mutation (a molecular-based subtype) prediction. A new deep semi-supervised learning method is also proposed to tackle the problem of missing molecular-related labels in training datasets for improving the performance of glioma classification. In other two applications, we also address video-based human fall detection by using co-saliency-enhanced Recurrent Convolutional Networks (RCNs), as well as obstacle avoidance in prosthetic vision by characterizing obstacle-related video features using a Spiking Neural Network (SNN). These investigations can benefit future research, where artificial intelligence/deep learning may open a new way for real medical applications.