A study of information-theoretic metaheuristics applied to functional neuroimaging datasets

This dissertation presents a new metaheuristic related to a two-dimensional ensemble empirical mode decomposition (2DEEMD). It is based on Green’s functions and is called Green’s Function in Tension - Bidimensional Empirical Mode Decomposition (GiT-BEMD). It is employed for decomposing and extracting hidden information of images. A natural image (face image) as well as images with artificial textures have been used to test and validate the proposed approach. Images are selected to demonstrate efficiency and performance of the GiT-BEMD algorithm in extracting textures on various spatial scales from the different images. In addition, a comparison of the performance of the new algorithm GiT-BEMD with a canonical BEEMD is discussed. Then, GiT-BEMD as well as canonical bidimensional EEMD (BEEMD) are applied to an fMRI study of a contour integration task. Thus, it explores the potential of employing GiT-BEMD to extract such textures, so-called bidimensional intrinsic mode functions (BIMFs), of functional biomedical images. Because of the enormous computational load and the artifacts accompanying the extracted textures when using a canonical BEEMD, GiT-BEMD is developed to cope with such challenges. It is seen that the computational cost is decreased dramatically, and the quality of the extracted textures is enhanced considerably. Consequently, GiT-BEMD achieves a higher quality of the estimated BIMFs as can be seen from a direct comparison of the results obtained with different variants of BEEMD and GiT-BEMD. Moreover, results generated by 2DBEEMD, especially in case of GiT-BEMD, distinctly show a superior precision in spatial localization of activity blobs when compared with a canonical general linear model (GLM) analysis employing statistical parametric mapping (SPM). Furthermore, to identify most informative textures, i.e. BIMFs, a support vector machine (SVM) as well as a random forest (RF) classifier is employed. Classification performance demonstrates the potential of the extracted BIMFs in supporting decision making of the classifier. With GiT-BEMD, the classification performance improved significantly which might also be a consequence of a clearer structure for these modes compared to the ones obtained with canonical BEEMD. Altogether, there is strong believe that the newly proposed metaheuristic GiT-BEMD offers a highly competitive alternative to existing BEMD algorithms and represents a promising technique for blindly decomposing images and extracting textures thereof which may be used for further analysis