Objective To develop an automatic segmentation model of acute stroke lesions based on DWI images,and build a 1-year recurrence prediction model of acute stroke based on this model and machine learning technology.Methods The patients with acute ischemic stroke who received intravascular therapy in Nanjing First Hospital from January 2019 to September 2021 were retrospectively included.The patients were divided into recurrence group and non-recurrence group according to clinical and imaging data within 1 year.The developed EfficientNet-B0 network was applied to segment acute stroke lesions on DWI images and its segmentation efficiency was evaluated.Based on automatic segmentation and manual delineation of tags respectively,the radiomics were extracted and the support vector machine classifier was used to construct the prediction model of acute stroke recurrence.Delong test was used to compare the differences between the two models.Results A total of 268 patients were included in the study,161 in the non-recurrence group and 107 in the recurrence group.The sensitivity,specificity,accuracy and Dice similarity coefficient of DWI automatic lesion segmentation model was 0.791,0.999,0.817 and 0.803,respectively.The area under the curve(AUC)of the prediction model of acute stroke recurrence based on the radiomics of the automatic segmentation lesions was 0.878(95%CI:0.834-0.923)(sensitivity:0.879,specificity:0.851).The AUC of the prediction model of acute stroke recurrence based on the radiomics of of manually outlined tags was 0.865(95%CI:0.819-0.911)(sensitivity:0.860,specificity:0.832).There was no significant statistical difference between the two models(Z=0.526,P=0.599).Conclusion The network proposed in this study can well segment acute stroke lesions on DWI,and the prediction model based on the radiomics of this model can predict the recurrence of acute stroke very well.