〔Abstract〕 Objective To explore the feasibility of a dilated convolution U-Net model for accurate segmentation of pancreas. Methods A kind of multi-scale dilated convolution U-Net model was improved based on standard U-Net. Radiotherapy positioning CT of 100 patients with pelvic tumors, containing the whole pancreas, were included in this study, which were used to train and test two kinds of models, and finally the results were compared. The quantitative indicators for the result evaluation were Dice Similarity Coefficient(DSC), Jaccard Similarity Coefficient(JSC), Hausdorff Distance(HD), and Average Surface Distance(ASD). Results The average DSC was 0.87, JSC was 0.78, ASD was 1.62 mm, HD was 9.85 mm. Comparing the standard U-Net, the dilated convolution U-Net had a better performance. T test was used in result analysis and the P value was less than 0.05 which meant that there were significant differences between the two results. Conclusion Building the dilated convolution model based on U-Net can accurately segment healthy pancreas, which is of great significance for improving the computer aided diagnosis system and the efficiency of radiotherapy.