Texture and CT-features in differentiation of hypervascular pancreatic neuroendocrine tumors from renal cell carcinoma metastases: diagnostic model

Abstract

Objective: to develop a diagnostic model that includes CT and radiomic features for the differential diagnosis of pancreatic neuroendocrine tumors (PNETs) G1 and G2 and pancreatic renal cell carcinoma (RCC) metastases.Material and Methods. 78 patients with 79 hypervascular PNETs and 17 patients with 24 pancreatic RCC metastases who underwent pancreatic resection and histological verification were selected in the study. All the patients underwent preoperative contrast enhanced CT (CECT). We assessed tumor attenuation, composition (cystic/solid), homogeneity (homogeneous/heterogeneous), calcification and presence of the main pancreatic duct (MPD) dilation. We calculated lesion-to-parenchyma contrast (LPC), relative tumor enhancement ratio (RTE) and extracted 52 texture features for arterial phase of CECT. Qualitative and texture features were compared between PNETs and pancreatic RCC metastasis. The selection of predictors for the logistic model was carried out in 2 successive stages: 1) selection of predictors based on one-factor logistic models, the selection criterion was p < 0.2; 2) selection of predictors using L2 regularization (LASSO regression after standardization of independent variables). The selected predictors were included in a logistic regression model without interactions, the coefficients of which were estimated using the maximum likelihood method with a penalty of 0.8.Results. There was no difference in composition, homogeneity (homogeneous/heterogeneous) and presence of the MPD dilation between groups. We did not find calcification in pancreatic RCC metastasis, in contrast to the PNETs (9% contained calcifications). After selection, the LCR, CONVENTIONAL_HUmin, GLCM_Correlation, NGLDM_Coarseness were included in the final diagnostic model, which showed a sensitivity and specificity of 95.8%; 62% in the prediction of pancreatic RCC metastases.Conclusion. The diagnostic model developed on the basis of texture and CT-features has high sensitivity (95.8%) with moderate specificity (62%), which allows it to be used in complex diagnostic cases to determine the patient's treatment tactics.