In this study, deep learning and radiomics methods were used to construct classification prediction models for SCA, MCA and IPMN. Then, more than 1300 radiomics features were extracted, including 256 deep learning features, which were used to construct the feature set. The fused model was constructed by logistic regression. The model established in this study has achieved good results in the classification prediction of SCA, MCA, and IPMN, which proves that deep learning radiomics models have clinical value in the classification of pancreatic cystic tumors.
Deep learning methods have been widely used in the disease classification. In pre-training, a transfer learning step was added to increase dataset samples in the early stage, and some network parameters were frozen to reduce overfitting [30]. The expansion of local image data capacity can better realize the prediction efficiency of deep learning itself [31]. The fusion of deep learning features and radiomics features has been shown to be superior to deep learning or radiomics alone [32–35], which is consistent with our experimental conclusions, suggesting that there may be complementarity between deep learning features and radiomics features.
Our findings suggest that specific morphological features may improve the prediction efficiency for classification of pancreatic cystic lesions. Among them, important features such as location, number of cysts (≥ 6) and wall calcification were used in the fused model to construct a SCA differential diagnosis model. Unlike previous studies, in our study, tumor size was not involved in the prediction model to distinguish between serous and non-serous pancreatic cystic tumors. Previous research results also confirmed that image features can be included in the classification prediction models of SCA and MCA, and even play a role in the classification models of atypical SCA and MCA [24, 27]. Obviously, location serves as an important feature to improve classification prediction models in radiomics. We believe the reason is that radiomic features do not include location information, and the location of tumor as an imaging feature can improve the prediction efficiency of models.
Communication with the pancreatic duct as an important morphological feature is involved in constructing the MCA and IPMN differential diagnosis model. The current consensus is that communication with the pancreatic duct is an important imaging feature of IPMN to distinguish it from other pancreatic cystic tumors. In previous radiomics studies, Yang and Shen's studies did not incorporate imaging features into the feature set to construct the classification model of pancreatic cystic lesions [22, 28]. Whereas in our study, only communication with the pancreatic duct played a role in our MCA and IPMN differential diagnosis model. We believe this is due to the inability of radiomics features derived from IPMN tumors to reflect the tumor-pancreatic duct relationship. Therefore, in radiomics studies, the evaluation of morphological features is necessary.
As a retrospective radiomics study, this study has certain limitations. First of all, despite the large number of cases included in our study, this is still a single-center study and the performance of our models on other datasets is uncertain. Secondly, we used 2D imaging data for the convenience of clinical use, while 3D data may contain more information about tumor heterogeneity. However, our study shows that the 2D-based radiomics model also has high predictive performance for the classification of pancreatic cystic lesions. Thirdly, we only used CT image data for model building. The classification models of pancreatic cystic lesions still need to be explored based on different imaging modalities. In the future, we will further explore the classification model of pancreatic cystic lesions based on other modal images to meet the needs of clinical individualized treatment.