Application of information fusion methods to biomedical data
Classification of Magnetic Resonance (MR) and Infrared (IR) spectra promises to become an effective tool for early medical diagnosis of diseases. The proposed thesis project involves the development and comparison of classification strategies and algorithms for the analysis of spectra of healthy and diseased tissue biopsies of various disease states. Several methods of aggregating outcomes of classifiers are considered in order to improve the classification accuracy, and applied to artificial and real-life spectra. Logistic regression, linear combination of classifiers, fuzzy integration, stacked generalization and some other methods of classifier aggregation, as well as different ways of estimating necessary parameters are considered. The results indicate that in many cases aggregation of classifiers improves the classification performance in comparison to that of the classifiers being aggregated. The results on real-life spectra vary. The methods perform well on some data sets and relatively poorly on others. Strategies are recommended to gain from classifier aggregation.