A Change Detection with Machine Learning Approach for Medical Image Analysis

The research activity focuses on evaluating data from medical images by applying clustering techniques on extracted components for an ex-ante/ex-post evaluation, commonly identified as "change detection", with respect to evolution times and/or comparison with other analyzed subjects. The methodological approach examines an unsupervised automatic method through the implementation of an algorithm for the extraction of meaningful features related to the properties of the medical image data analyzed. It then goes on to normalize the data contained in the matrix to evaluate through multivariate analysis the notion that similar objects produce similar responses without knowing their entity, type, and class descriptions, which are inferred by making observations on the clusters. The main specificity of this algorithm is that classes are identified from compact, well-distinguishable clusters without knowing the extent of their nature; the entire feature space is divided into classes using proximity or similarity criteria. After finishing the process of class identification, properties will be associated with them in relation to the known descriptions. At the end of the procedure, factorial analysis using the principal component method is applied. The clusters extracted from the data are described by their properties, which make it possible to identify, on each of the new factorial axes, homogeneous classes of clusters characterized predominantly by the only variables that have a high correlation value between variable and factor. The automatic identification of classes or "phenomena" that exhibit very different mean and/or variance allows easy reading of the results for domain experts.