Abstract: Objective To develop a deep learning model based on arteriovenous dual-phase CT imaging features and evaluate its diagnostic value in differentiating benign from malignant pancreatic cystic lesions (PCLs). Methods Preoperative contrast-enhanced CT images of patients with histopathologically confirmed PCLs at Peking Union Medical College Hospital from June 1, 2014 to May 31, 2023 were retrospectively collected. The data were randomly partitioned at the lesion level into training, validation, and test sets in a 3:1:1 ratio. CT images were preprocessed and regions of interest were delineated. Using postoperative pathological results as the reference standard, five deep learning models (ResNet50, DenseNet121, ResNeXt50, EfficientNet-b5, and MobileNetV2) were constructed to extract arteriovenous dual-phase CT imaging features for binary classification of PCLs. The optimal model was selected based on the area under the curve (AUC), accuracy, sensitivity, and specificity. Further comparative analyses were conducted against single-phase CTbased deep learning models, conventional radiomics models, and radiologist interpretations to comprehensively evaluate the performance of the arteriovenous dual-phase CT-based deep learning model in the differential diagnosis of PCLs. Results A total of 480 patients with 485 lesions (206 malignant and 279 benign) were ultimately enrolled. The training, validation, and test sets comprised 291, 97, and 97 lesions, corresponding to 288, 96, and 96 patients, respectively. In the validation set, the ResNeXt50 model based on arteriovenous dual-phase CT features achieved an AUC of 0' 837 (95% CI:0' 748-0' 915), an accuracy of 77' 32% (95% CI:67' 70%-85' 21%), a sensitivity of 80' 49% (95% CI:65' 13%-91' 18%), and a specificity of 75' 00% (95% CI:61' 63%-85' 61%). In the test set, the corresponding values were 0' 822 (95% CI:0' 737-0' 904), 73' 20% (95% CI:63' 24% -81' 68%), 82' 93% (95% CI:67' 94%-92' 85%), and 66' 07% (95% CI:52' 19%-78' 19%), demonstrating overall superior performance. Calibration curves indicated that the predicted probabilities of the model were generally consistent with observed outcomes, albeit with certain shortcomings in probability calibration. Decision curve analysis revealed that clinical decision-making based on this model conferred net patient benefit. Compared with single-phase CT-based models and conventional radiomics models, the ResNeXt50 model incorporating arteriovenous dual-phase CT features exhibited superior overall performance, and its diagnostic performance was comparable to that of radiologists, with good consistency. Conclusions The deep learning model based on arteriovenous dual-phase CT imaging features demonstrates diagnostic value in differentiating benign from malignant PCLs and may serve as an adjunctive reference for preoperative clinical assessment. However, its stability and generalizability warrant further validation in larger cohorts and with external datasets.