Research On Systematic Fusion Of Cognitive Networks And Its Applications

The cognitive processing of human brain generally includes calculation,social recognition,language,decision,and memory etc.In performing a specific cognitive processing task,the human brain will form a corresponding functional network,namely cognitive network.Functional magnetic resonance imaging(f MRI)is a widely established method to detect and delineate the activation in brain regions under specific cognitive task conditions.Functional connectivity analysis based on f MRI can contribute to reveal the coupling mode of components of large distributed nervous system when performing specific cognitive functions.However,the diversity of design paradigms in traditional f MRI cognitive experiment,together with the variety of brain atlases in result analysis,bring out difficulties for the research on brain’s advanced cognitive function.In particular,the considerable distinctions between various definitions of typical cognitive networks from different sources or different versions of brain atlas may inevitably lead to the disparities in the analyzed results and mechanism interpretations.To address these issues,the present dissertation attempts to adopt a systematic methodology of brain informatics to conduct the f MRI data analyses from three different levels,namely cognitive feature selection,cognitive network construction and cognitive network fusion,through combining the multi-task data and multi-source brain atlases under the framework of systematic brain informatics methodology.The main innovations and conclusions are summarized as follows:Firstly,a cognitive feature selection method based on the topological property is proposed,given that the feature selection of typical machine learning methods in brain decoding have the defects in the full consideration of cognitive model and computational model.With the analysis of the f MRI data set generated from mental arithmetic task and the evaluation of selected cognitive features under the curve of brain cognitive network architecture,it is confirmed that the results generated by the proposed method have better interpretability under the architecture of cognitive networks.Thereafter,the cognitive features selected by the proposed method were utilized to construct the classification model to distinguish the cognitive states in mental arithmetic task and eye-opening resting state.The classification model was constructed based on the support vector machine(SVM),and the experiments were conducted with multiple kernel functions.The experimental results proved that the trained classification model based on cognitive features selected by the proposed method can effectively complete the classification task.Compared with other typical machine learning methods,despite its less optimal performance,there was a quite small gap between its performance and that from other methods,resulting in the generally consistent overall performances.With a comprehensive comparison between the rationality of cognitive mechanism explanation and the accuracy of cognitive state classification,it can be judged that the proposed method can produce more prominent results.Secondly,a cognitive network construction method based on task-sensitive voxels is proposed.This method aims to tackle the deficiency and inaccuracy in the description of cognitive network on anatomical brain atlas and functional brain atlas.After the calculation of voxels sensitivity to cognitive tasks in the relevant areas of brain cognitive function and the construction of cognitive network with well-distributed and decentralized ROIs,the topological property indices of the constructed cognitive network and the preference of brain regions constituting the ROIs were evaluated and discussed.With the experiment on the data generated by facial emotion recognition task,it is confirmed that the fusiform gyrus network generated by the proposed method has a higher network integration,and the ROIs adopted in network construction are more preferentially from the right fusiform face area.Such a proposed method was then used to reconstruct the language network defined in a typical brain atlas.The experimental results show that the reconstructed language network not only has higher network connection density and integration,but also the ROIs adopted are more preferentially from areas associated with language functions such as the middle temporal gyrus.Finally,a fusion method and its implementation process targeted at the multi-source cognitive networks are proposed.Experiments were carried out with negative correlated network(default mode network)and positive correlated network(fronto-parietal network)of cognitive tasks to evaluate the proposed method,and the result discussions were conducted as well.The first step involved the combination of the instances of cognitive network.By comparing the performance of each instance’s topological property on the combined cognitive network for distinguishing the cognitive task states,the instance with optimal performance was selected as the primary cognitive network,while the remaining instances were selected as the supplementary cognitive networks.Next,the primary cognitive network was taken as the initial candidate fused cognitive network,and the ROIs of all supplementary cognitive networks were sorted according to their nodal topology properties.The ROI with the optimal priority was moved into the candidate fused cognitive network iteratively.In the iterative process,the performance of the current candidate fused cognitive network was continuously calculated.Finally,the candidate fused cognitive network with the optimal performance in the entire iterative process was taken as the final fused cognitive network,that is,its topological property had the most outstanding performance in discriminating different cognitive states.Particularly in fusing the multi-source fronto-parietal networks,it is found that the fused fronto-parietal network generated with the proposed method has much better performance in cognitive states classification than that produced by typical machine learning methods.The retention rate of ROIs generated by the proposed method is also the most balanced in the frontal lobe and parietal lobe.In summary,with the focus on the brain cognitive functions,especially in the node selection,network generation and network fusion of advanced cognitive function,this dissertation proposed an approach to select the cognitive features,a method of constructing cognitive network,as well as a cognitive networks fusion method.The final results on f MRI cognitive experimental data corresponding to task positive correlated,task negative correlated and task positive/negative correlated cognitive networks indicate that the proposed methods can produce quite satisfied results.