Comparison of radiomics and deep learning signature for the lymph node metastasis detection in pancreatic neuroendocrine tumor

#3621

Introduction: PNETs is rare pancreatic tumors and the accuracy of diagnostic for lymph node metastasis (LNM) is low. Therefore, the quantification of the radiomics and deep learning features (DLF) may help to elevate the accuracy of detection the LNM.

Aim(s): To develop and compare the novel signatures which based on the radiomics and deep learning features for prediction of the lymph node metastasis (LNM) in pancreatic neuroendocrine tumor (pNET).

Materials and methods: 220 patients from Fudan University Shanghai Cancer center were enrolled. All the patients have enhanced multislice computer tomography (MSCT) imaging, surgical and pathological information. The segmentation of tumors was in the arterial phase of CT imaging by two senior radiologists. The DLF was extracted from the DensenNet201 architecture based on deep convolutional neural networks. The weighted gene co-expression network analysis was used to select the features which related to LNM. Then the least absolute shrinkage and selection operator method was used to determine the final features. Finally, the novel signature was constructed by the lightGMB. The clinical factors selection and model construction were based on the univariate and multivariate logistic regression. The performance evaluation based on the ROC curve.

Conference:

Presenting Author: Tang W

Authors: Tang W, Gu W, Chen J,

Keywords: Pancreatic neuroendocrine tumor, lymph node metastasis, radiomics, deep learning features, Multislice computer tomography (MSCT),