Diagnostic efficacy of a combined diagnostic model based on extreme gradient boosting algorithm in differentiating the pathological grading of gastric neuroendocrine neoplasms

Abstract

Objective: To evaluate the diagnostic efficacy of combined diagnostic model based on extreme gradient boosting (XGBoost) algorithm to determine the pathological grading of gastric neuroendocrine neoplasms (NENs). Methods: A total of 81 gastric NENs patients in the First Affiliated Hospital of Zhengzhou University confirmed by definite pathological grading from August 2012 to December 2019 were enrolled. The data of clinical and CT findings were collected. The number of lesions, tumor location, shape, lymph node metastasis, thickness, longitude of tumor and CT values in arterial and venous phase were analyzed. ITK-SNAP software and Python 2.1.0 PyRadiomics software were used to perform the image preprocessing and radiomics features extraction from segmented images. XGBoost algorithm was used to build the CT findings model, radiomics model in arterial phase, radiomics model in venous phase and combined diagnostic model. The diagnostic efficacy of CT imaging model, radiomics model in arterial phase, radiomics model in venous phase and combined diagnostic model were evaluated by accuracy, mean squared error (MSE) and mean absolute error (MAE). Results: The subjects were 28.0 to 78.0 (58.6+10.7) years old, including 56 males (69.1%). The number of lesions, tumor location, shape, lymph node metastasis, thickness and longitude of tumor between G1/G2 and G3 patients showed statistic significances (all P values<0.05), while there were no differences in CT values in arterial and venous phase (both P values>0.05). Six most important features in the combined diagnostic model were A_logarithm_glcm_Imc1, P_squareroot_glcm_Maximum Probability, thickness, longitude, A_wavelet-HHL_glrlm_GrayLevelNonUniformity and P_wavelet-LLL_ngtdm_Contrast, respectively. The accuracy of CT findings model, radiomics model in arterial phase, radiomics model in venous phase and combined diagnostic model were 81.8%, 86.0%, 87.8% and 91.0%, respectively; with MSE were 539.41, 490.08, 429.99 and 371.92, respectively; and MAE were 16.72, 15.25, 14.23 and 12.33, respectively. The MAE value of the combined diagnostic model was lower than those of CT findings model and radiomics model in arterial phase (P<0.001 and 0.004, respectively), while no statistically difference was detected compared to radiomics model in venous phase (P=0.111). Conclusion: The combined diagnostic model based on XGBoost algorithm have a good diagnostic efficiency for the pathological grading of gastric NENs.