Background: Accurate infarct volume measurement requires manual segmentation in diffusion weighted image (DWI) which is time-consuming and prone to variability. We compared two DWI infarct segmentation programs based on deep learning and the apparent diffusion coefficient threshold (JBS-01K and RAPID DWI, respectively) in a comprehensive stroke center. Method: We included 414 patients whose DWI were evaluated using RAPID DWI and JBS-01K. We used the Bland-Altman plot to compare estimated and manually segmented infarct volumes. We compared R-squared, root mean squared error, Akaike information criterion, and log likelihood after linear regression of manually segmented infarct volumes. Results: The mean age of included patients was 70,0±12.4 years, and 60.9% were male. The median time between the last known well and a DWI was 12.4 hours. JBS-01K segmented infarct volumes were more comparable to manually segmented volumes compared to RAPID DWI. JBS-01K had a lower root mean squared error (6.9 vs. 10.8) and log likelihood (p<0.001) compared to RAPID DWI. In addition, compared to RAPID DWI, JBS-01K more correctly classified patients according to the infarct volume threshold used in endovascular treatment trials (overall accuracy 98.1% vs. 94.0%; p = 0.002). In 35 patients who received DWI prior to endovascular treatment, JBS-01K infarct volume segmentation was more closely related to manual infarct volume segmentation. Conclusion: We demonstrated that a deep learning method segmented infarct on DWI more accurately than one based on the apparent diffusion coefficient threshold.