Objective:Based on diffusion basis spectrum imaging(DBSI)combined with deep learning methods,the purpose of this study was to explore the changes in the topological structure of brain networks in preschool children with autism spectrum disorder(ASD)and their diagnostic value.Methods:A retrospective analysis was conducted on clinical and brain DBSI data from 31 diagnosed ASD patients and 30 healthy controls(HCs)at our hospital from September 2022 to July 2023.The a-ges of all subjects in both groups ranged from 2 to 6 years.The original data of DBSI_20 and DBSI_21 sequences for each subject were imported into DSI Studio software to generate DBSI_combine diffusion images.All data of the three sets of images were analyzed for connectivity matrices and graph theory u-sing DSI Studio.Then,the Kruskal-Wallis method was used to extract the most valuable features showing significant intergroup differences in global and node graph metrics.Pearson correlation coeffi-cient(PCC)method was used to perform dimensionality reduction on features with correlation coeffi-cients greater than 0.99,thus to reduce the number of features.And then,the best features were selec-ted out for constructing three predicting models.ROC curve analysis was performed to evaluate the predictive performance of the models based on the three sets of sequences,with metrics including accu-racy,recall,precision,F1 score,and area under the ROC curve(AUC).Results:In the model based on the DBSI_20 sequence,the most discriminative features were Angular_L_eccentricity and Temporal_Mid_R_eccentricity;in the model based on the DBSI_21 sequence,the most discriminative features were Temporal_Sup_R_eigenvector_centrality and Cerebellum_9_L_pagerank_centrality;and in the model based on the DBSI_combine images,the most discriminative features are Postcentral_L_Degree and Angular_L_eccentricity.The model based on the DBSI_20 sequence achieved a higher AUC than that based on the DBSI_21 sequence(0.963 vs.0.481),while the model constructed from the DBSI_combine images had the highest AUC(0.975).The accuracy,recall,precision and F1 score of the DB SI_20 model were 0.890,1.000,0.784 and 0.879,respectively;for the DBSI_21 model,they were 0.579,0.556,0.861 and 0.675,respectively;and for the DBSI_combine model,they were 0.936,0.889,1.000 and 0.979,respectively.Conclusion:The prediction models based on the DBSI_20 sequence or DBSI_combine images can accurately predict preschool children with ASD,and the DBSI_combine model has superior efficacy.
万方数据
维普
