Differentiating pancreatic adenosquamous carcinoma from pancreatic ductal adenocarcinoma by CT radiomic and deep learning features / 中华胰腺病杂志

ABSTRACT

Objective:To develop and validate the models based on mixed enhanced computed tomography (CT) radiomics and deep learning features, and evaluate the efficacy for differentiating pancreatic adenosquamous carcinoma (PASC) from pancreatic ductal adenocarcinoma (PDAC) before surgery.Methods:The clinical data of 201 patients with surgically resected and histopathologically confirmed PASC (PASC group) and 332 patients with surgically resected histopathologically confirmed PDAC (PDAC group) who underwent enhanced CT within 1 month before surgery in the First Affiliated Hospital of Naval Medical University from January 2011 to December 2020 were retrospectively collected. The patients were chronologically divided into a training set (treated between January 2011 and January 2018, 156 patients with PASC and 241 patients with PDAC) and a validation set (treated between February 2018 and December 2020, 45 patients with PASC and 91 patients with PDAC) according to the international consensus on the predictive model. The nnU-Net model was used for pancreatic tumor automatic segmentation, the clinical and CT images were evaluated, and radiomics features and deep learning features during portal vein phase were extracted; then the features were dimensionally reduced and screened. Binary logistic analysis was performed to develop the clinical, radiomics and deep learning models in the training set. The models' performances were determined by area under the ROC curve (AUC), sensitivity, specificity, accuracy, and decision curve analysis (DCA).Results:Significant differences were observed in tumor size, ring-enhancement, upstream pancreatic parenchymal atrophy and cystic degeneration of tumor both in PASC and PDAC group in the training and validation set (all P value <0.05). The multivariable logistic regression analysis showed the tumor size, ring-enhancement, dilation of the common bile duct and upstream pancreatic parenchymal atrophy were associated with PASC significantly in the clinical model. The ring-enhancement, dilation of the common bile duct, upstream pancreatic parenchymal atrophy and radiomics score were associated with PASC significantly in the radiomics model. The ring-enhancement, upstream pancreatic parenchymal atrophy and deep learning score were associated with PASC significantly in the deep learning model. The diagnostic efficacy of the deep learning model was highest, and the AUC, sensitivity, specificity, and accuracy of the deep learning model was 0.86 (95% CI 0.82-0.90), 75.00%, 84.23%, and 80.60% and those of clinical and radiomics models were 0.81 (95% CI 0.76-0.85), 62.18%, 85.89%, 76.57% and 0.84 (95% CI 0.80-0.88), 73.08%, 82.16%, 78.59% in the training set. In the validation set, the area AUC, sensitivity, specificity, and accuracy of deep learning model were 0.78 (95% CI 0.67-0.84), 68.89%, 78.02% and 75.00%, those of clinical and radiomics were 0.72 (95% CI 0.63-0.81), 77.78%, 59.34%, 65.44% and 0.75 (95% CI 0.66-0.84), 86.67%, 56.04%, 66.18%. The DCA in the training and validation sets showed that if the threshold probabilities were >0.05 and >0.1, respectively, using the deep learning model to distinguish PASC from PDAC was more beneficial for the patients than the treat-all-patients as having PDAC scheme or the treat-all-patients as having PASC scheme. Conclusions:The deep learning model based on CT automatic image segmentation of pancreatic neoplasm could effectively differentiate PASC from PDAC, and provide a new non-invasive method for confirming PASC before surgery.