Intelligent Analysis of Cerebral Magnetic Resonance ImagesExtracting Relevant Information from Small Datasets

Résumé

Machine learning methods applied to medical imaging are becoming important tools for the analysis and diagnose of patients. The huge availability of repositories from multimodal imaging has favored the development of systems that learn to extract relevant features and construct predictive models from enormous amounts of data, for example, deep learning methods. However, the analysis of imaging datasets from smaller number of subjects, as normally gathered in preclinical and clinical research environments, has received considerably less attention. In this thesis, we implement a variety of advanced computational tools to overcome this problem, supporting robust analysis of Magnetic Resonance Imaging (MRI) from applications that involve small numbers of subjects. We illustrate these approaches analyzing automatically datasets obtained from functional MR images of the cerebral regulation of appetite in rodents and humans, and from functional and structural MR images from glioma development in animal models and humans. The proposed methods evolved from the idea of considering each voxel from the image dataset as a pattern, rather than from the conventional notion of considering each image as a pattern. Chapter 1 describes the motivations supporting these developments, including the objectives proposed, the general structure of the document and the contributions of this research. Chapter 2 provides an updated introduction to the state of the art in MRI, the conventional image pre-processing methods, and the most useful machine learning algorithms and their MRI applications. Chapter 3 presents the experimental design, and image pre-processing steps as applied to the datasets from appetite regulation and glioma development. Chapter 4 implements new supervised learning algorithms for the analysis of MRI datasets as obtained from a small number of subjects. We illustrate this approach, presenting first the Fisher Maps methodology, which allow for the non-invasive and comprehensive visualization of the integral cerebral appetite circuitry, through the automatic analysis of Diffusion Weighted Image (DWI) datasets. This methodology is expanded to the classification of complete images, using the combined predictions obtained from all their pixels. Chapter 5 proposes new unsupervised learning algorithm for the analysis of MRI datasets and illustrates its performance with synthetic data and datasets of brain tumoral studies and glioma development. Finally, Chapter 6 summarizes the main conclusions, providing ample avenues for the continuation of this work. In summary, the present thesis provides new useful approach for the automatic extraction of relevant information from the analysis of MRI datasets in contexts where small datasets are available, using advanced supervised and unsupervised artificial intelligence algorithms. The proposed methods may be easily extended to other paradigms or pathologies, and even to, alternative imaging modalities.