Advances in Analysis and Exploration in Medical Imaging

With an ever increasing life expectancy, we see a concomitant increase in diseases capable of disrupting normal cognitive processes. Their diagnoses are difficult, and occur usually after daily living activities have already been compromised. This dissertation proposes machine learning methods for the study of the neurological implications of brain lesions. It addresses the analysis and exploration of medical imaging data, with particular emphasis to (f)MRI. Two main research directions are proposed. In the first, a brain tissue segmentation approach is detailed. In the second, a document mining framework, applied to reports of neuroscientific studies, is described. Both directions are based on retrieving consistent information from multi-modal data. A contribution in this dissertation is the application of a semi-supervised method, discriminative clustering, to identify different brain tissues and their partial volume information. The proposed method relies on variations of tissue distributions in multi-spectral MRI, and reduces the need for a priori information. This methodology was successfully applied to the study of multiple sclerosis and age related white matter diseases. It was also showed that early-stage changes of normal-appearing brain tissue can already predict decline in certain cognitive processes. Another contribution in this dissertation is in neuroscience meta-research. One limitation in neuroimage processing relates to data availability. Through document mining of neuroscientific reports, using images as source of information, one can harvest research results dealing with brain lesions. The context of such results can be extracted from textual information, allowing for an intelligent categorisation of images. This dissertation proposes new principles, and a combination of several techniques to the study of published fMRI reports. These principles are based on a number of distance measures, to compare various brain activity sites. Application to studies of the default mode network validated the proposed approach. The aforementioned methodologies rely on clustering approaches. When dealing with such strategies, most results depend on the choice of initialisation and parameter settings. By defining distance measures that search for clusters of consistent elements, one can estimate a degree of reliability for each data grouping. In this dissertation, it is shown that such principles can be applied to multiple runs of various clustering algorithms, allowing for a more robust estimation of data agglomeration.