Linear Models for Visual Data Mining in Medical Images

Modern non-invasive imaging modalities such as Computer-Aided Tomography and Magnetic Resonance Imaging brought a new perspective to anatomic characterization, as they allowed detailed observation of in vivo structures. The amount of provided data became, however, progressively overwhelming. Data may be useless if we are not able to extract relevant information from them and to discard undesirable artifacts and redundancy. Ultimately, we aim to discover new knowledge from collected data, and to be able to statistically represent this knowledge in the form of prior distributions that may be used to validate new hypotheses, in addition to clinical and demographic information. In this chapter we propose an analysis of available methods for data mining in very high-dimensional sets of data obtained from medical imaging modalities. When applied to imaging studies, data reduction methods may be able to minimize data redundancy and reveal subtle or hidden patterns. Our analysis is concentrated on linear transformation models based on unsupervised learning that explores the relationships among morphologic variables, in order to find clusters with strong correlation. This clustering can potentially identify regions that have anatomic significance and thus lend insight to knowledge discovery and hypothesis testing. We illustrate this chapter with successful case studies related to the segmentation of neuroanatomic structures and the characterization of pathologies.