其他摘要 | In recent years, the incidence of Autism Spectrum Disorder (ASD) has increased significantly worldwide, but diagnosing ASD is still difficult, and the analysis of ASD's pathological causes always is a big problem in this filed. Therefore, it is very meaningful to research autism classification method based on machine learning algorithms, which will provide more objective indicators to assist doctors in diagnosis and to improve the accuracy and efficiency of autism diagnosis. Based on the open dataset ABIDE-I, this thesis researches the autism classification algorithm by using temporal information and spatial information of resting state functional magnetic resonance imaging data (rs-fMRI) by means of the depth learning theory. In addition, this thesis also analyses the possible pathological causes of ASD by using functional connectivity's information between possible brain regions.
In time domain, the autism classification model built with the long short term memory (LSTM) structure is trained by the time sequence information of Anatomical Automatic Labeling atlas (AAL) and Craddock 200 atlas (CC200) respectively. In this process, in order to reduce the impact of the noise from different acquisition devices, this study performs data cleaning and data enhancement on ABIDE-I data. Finally, the classification methods have got the accuracy of 59.23% (AAL), and 68.65% (CC200) for classifying ASD and typically developing (TD).
In spatial domain, the autism classification model built with multi-layer perceptron (MLP) is trained by CC200's functional connectivity information, and the autism classification model built with three-dimension convolutional neural network is trained by Regional Homogeneity (ReHo) information. On ABIDE-I database, the classifier trained by CC200 information has achieved 70.3% classification accuracy, and the classifier trained by ReHo has achieved 69.3% classification accuracy. In addition, it further studies the fusion strategy of features which are from CC200 and ReHo, and finally has obtained 79.2% classification accuracy on ABIDE-I database.
Through the analysis method of weight parameters in deep learning, from the perspective of brain functional regions, this thesis researches ASD's possible abnormal brain functional regions based on the pre-trained LSTM network's model parameters in time domain, and has found that the main abnormal brain functional regions of ASD include Frontal, Cuneus, Cingulum, Angular, Calcarine, ParaHippocampal, Amygdala, etc. From the perspective of brain functional connectivity, this thesis researches possible abnormal brain functional connectivity regions based on the pre-trained MLP network's model parameters of CC200's connectivity matrix in space domain, and has found that some brain functional connectivity of ASD between left and right brain regions are weaker than that of typically developing (TD). Furthermore, based on the triple network structure, this thesis also analyses the functional connectivity of "brain network" and has found that the functional connectivity of ASD between anterior cingulate cortex (ACC) and prefrontal cortex (PFC) is weaker than that of TD.
The experimental result proves that ASD classifiers learnt by deep learning technology based on resting state functional magnetic imaging, can diagnose ASD patients effectively. In addition, the experiment result also proves that it is useful and feasible to perform ASD's pathological exploration in brain functional regions and brain functional connectivity based on ASD classifiers' weight parameters, which provides a new research idea for ASD's pathological cause exploration. |
修改评论