Pattern analysis for regions of interest in spatial data with applications to medical images

Spatial data are encountered in several applications and areas of research, such as geographic information systems, multimedia databases, and medical image analysis. Analysis of regions of interest (ROIs) in such data typically requires focusing into specific regions that are of particular interest to an expert. When dealing with medical images, these ROIs are often encountered in the form of tumors, lesions, anatomical structures or functional activity regions. When considering such ROIs in medical images, analysis typically relies on effective descriptors that reflect spatial characteristics and structural properties of the ROIs as well as their topological arrangement in the image. Nevertheless, most of the conventional techniques proposed for image analysis are usually based on descriptors that reflect properties of the entire image space. These generic descriptors do not capture the discriminative properties of particular ROIs that are important to applications such as computer aided diagnosis. In this thesis, we address this problem of describing characteristic discriminative properties of spatial ROIs. Particularly, we propose an array of techniques suitable for pattern analysis of ROIs and introduce methods to detect, characterize, and classify various types of ROIs.; The techniques proposed in this thesis constitute a framework of important steps for ROI pattern analysis. We propose a technique that performs an adaptive partitioning of the image space, guided by statistical tests on groups of voxels, in order to detect highly discriminative ROIs among different groups of images. In order to effectively represent ROIs, we introduce a feature extraction technique which constructs a vector descriptor based on concentric spheres radiating out of the ROI's center of mass. The computed features reflect structural characteristics and internal volume properties of the ROI. The extracted feature vector is then used to perform classification and similarity searches. We also introduce a statistical dimensionality reduction technique based on Markov chain Monte Carlo simulations that improves pattern analysis and classification accuracy by selecting a subset of the most informative features.; The proposed analysis framework is extensively evaluated using various types of artificial and real image ROIs. Our discriminative ROI detection technique achieves up to 95% accuracy in detecting and classifying fMRI ROIs with discriminative functional activity among control subjects and Alzheimer's disease patients. This result outperforms other state of the art techniques by almost 15%, while reducing by two orders of magnitude the number of statistical tests typically required by voxel-based approaches. We also analyze MRI images of the human brain in order to investigate gender-based morphological variability. In this case, our results validate previous findings reported in the literature, while effectively reducing by up to almost 50% the computational cost required for the analysis. Finally, we demonstrate the effectiveness of our feature extraction and dimensionality reduction techniques with classification and similarity searches of ROIs generated by a region data simulator and real ROIs extracted from fMRI images. Our characterization methods achieve up to 85% accuracy in classifying individual ROIs extracted from control and patient images, while being up to two orders of magnitude faster than other commonly used feature extraction techniques.; The pattern analysis techniques proposed in this thesis have general applicability to spatial data of various formats and application domains. Considering the fascinating advancements in medical imaging, our framework provides tools for efficient medical image post-processing and analysis, having a great potential to assist medical decision making.