An Unsupervised Machine Learning Approach for Medical Image Analysis

The research activity presented in this work focuses the results on data analysis of three-dimensional biomedical PET images by providing clusters that synthesize the extracted components for an objective analysis of the data with respect to the evolution of the diagnosis. The methodological approach is based on the creation of an unsupervised automatic method, through the definition of a feature selection algorithm and extraction of significant numerical attributes relating to the properties of the study elements belonging to the selected clusters. After the normalization of the data contained in the matrix, multivariate analysis was carried out using the concept that similar objects produce similar responses without knowing the entity, type, class descriptions, which are deduced by making observations on the clusters. The factorial analysis with the main components method, applied to the clusters extracted from the data and described by their properties, will allow the identification of homogeneous classes of clusters characterized mainly by only the variables that have a high correlation value between variable and factor, obtaining the automatic identification of classes or "phenomena" that have greater or lesser variance.