CT-based radiomics nomogram to differentiate non-functional pancreatic neuroendocrine tumors from solid pseudopapillary tumors
ZHOU Yongjin1, GAO Ruijie1, JIANG Chunyan1, 2, DENG Jingjing3, XIA Shuiwei1, SHEN Shaobo1, WANG Zufei1, JI Jiansong1.
1.Department of Radiology, Key Laboratory of Imaging Diagnosis and Minimally Invasive Intervention Research of Zhejiang Province, Lishui Hospital of Zhejiang University, Lishui 323000, China; 2.Department of Radiology, Songyang County People’s Hospital, Lishui 323400, China; 3.Department of Radiology, Lishui People’s Hospital, Lishui 323000, China
Abstract：Objective: To investigate the value of CT-radiomic model for preoperative differentiating nonfunctional pancreatic neuroendocrine tumors (NF-pNETs) from solid pseudopapillary tumors (SPTs). Methods:A total of 87 cases with pathologically confirmed NF-pNETs and SPTs were included and randomly divided into training set (62 cases) and validation set (25 cases) at a ratio of 7:3. A.K. software was used to extract tumor texture features from plain CT, arterial and delayed CT images. The maximum correlation and minimum redundancy (mRMR), least absolute shrinkage and selection operator (LASSO) and 10-fold cross validation were used to retain features. Logistic regression analysis and Rad-score were used to construct the radiomic model nomogram. Receiver operating characteristic (ROC) curve was used to evaluate the diagnostic efficiency of the radiomic model nomogram. Decision curve analysis (DCA) was used to evaluate the benefits of the radiomic nomogram. Results: Totally 396 features were extracted from plain CT scan, arterial phase and delay phase images respectively. Ultimately, 8, 6 and 7 features were retained after mRMR, LASSO and 10-fold cross validation respectively, and 14 features were retained for the combined radiomic models. Multivariate logistic regression was used to construct a combined radiomic model nomogram including gender, age, maximum tumor diameter and texture extracted from plain CT scan, arterial phase, and delay phase images. The combined radiomics model achieved better prediction performance than radiomics model based on the single CT plain scan, arterial phase and portal venous phase CT images. The area under curve of the combined radiomics model was 0.97 (95%CI=0.94-1.00) for the training set and 0.92 (95%CI=0.81-1.00) for the validation set. DCA revealed that when the risk threshold was greater than 0.45, the use of combined radiomics model to differentiate between the tumor and pancreas to predict SPTs vs. NFNETs was of greater clinical value than a treat-all-patients as SPTs scheme or a treat-none as NF-NET scheme. Conclusion: Radiomics nomogram based on combined plain CT scan, arterial phase and delayed phase images shows good performance to distinguish NF-pNETs from SPTs,
which may be used as a noninvasive imaging tool for preoperative evaluation.
