Radiomics deep learning analysis for differentiating between mucin-producing pancreatic cystic neoplasms and serous cystic neoplasms

Objectives To develop and compare radiomics and deep learning models using biphasic contrast-enhanced CT (CECT) images for differentiating between mucin-producing pancreatic cystic neoplasms (PCNs) and serous cystic neoplasms (SCNs). Methods This retrospective study included 295 patients with pathologically confirmed pancreatic cystic neoplasms. Radiomics models were constructed by extracting and analyzing radiomics features from arterial phase (AP) and venous phase (VP) CT images. Additionally, 2D ResNet-50 deep learning models based on CECT images were developed for prediction of pancreatic cystic neoplasms. Fusion models were constructed by combining radiomics and deep learning features based on a multi-layer perceptron classifier. Model performance was evaluated in training, validation, and external test sets using receiver operating characteristic (ROC) analysis and decision curve analysis (DCA). Results For differentiating mucin-producing PCNs from SCNs, fusion and deep learning (DL) models achieved higher AUCs than those of radiomics models in external test set (AUC of AP-based fusion model: 0.853, AUC of VP-based fusion model: 0.855; AUC of AP-based DL model: 0.808, AUC of VP-based DL model: 0.848; AUC of AP-based radiomics model: 0.679, AUC of VP-based radiomics model: 0.634). DeLong's test and DCA showed that the AP-based fusion model had excellent predictive performance and clinical applicability in our study. Conclusion Fusion model combining radiomics and DL features improved the discrimination of mucin-producing PCNs from SCNs using biphasic CECT images, which could further broaden the prospects for personalized decision-making in the management of PCNs.