RESUMO
Objective:To construct of a computed tomography (CT) based radiomics model for predicting the prognosis of patients with gastric neuroendocrine neoplasm (GNEN) and inves-tigate its application value.Methods:The retrospective cohort study was conducted. The clinico-pathological data of 182 patients with GNEN who were admitted to 2 medical centers, including the First Affiliated Hospital of Zhengzhou University of 124 cases and the Affiliated Cancer Hospital of Zhengzhou University of 58 cases, from August 2011 to December 2020 were collected. There were 130 males and 52 females, aged 64(range, 56-70)years. Based on random number table, all 182 patients were divided into the training dataset of 128 cases and the validation dataset of 54 cases with a ratio of 7:3. All patients underwent enhanced CT examination. Observation indicators: (1) construction and validation of the radiomics prediction model; (2) analysis of prognostic factors for patients with GNEN in the training dataset; (3) construction and evaluation of the prediction model for prognosis of patients with GNEN. Measurement data with skewed distribution were represented as M(range), and comparison between groups was conducted using the Mann-Whitney U test. Count data were described as absolute numbers, and the chi-square test, corrected chi-square test or Fisher exact probability were used for comparison between groups. The Kaplan-Meier method was used to calculate survival rate and draw survival curve, and the Log-rank test was used for survival analysis. The COX regression model was used for univariate and multivariate analyses. The R software (version 4.0.3) glmnet software package was used for least absolute shrinkage and selection operator (LASSO)-COX regression analysis. The rms software (version 4.0.3) was used to generate nomogram and calibration curve. The Hmisc software (version 4.0.3) was used to calculate C-index values. The dca.R software (version 4.0.3) was used for decision curve analysis. Results:(1) Construction and valida-tion of the radiomics prediction model. One thousand seven hundred and eighty-one radiomics features were finally extracted from the 182 patients. Based on the feature selection using intra-group correlation coefficient >0.75, and the reduce dimensionality using LASSO-COX regression analysis, 14 non zero coefficient radiomics features were finally selected from the 1 781 radiomics features. The radiomics prediction model was constructed based on the radiomics score (R-score) of these non zero coefficient radiomics features. According to the best cutoff value of the R-score as -0.494, 128 patients in the training dataset were divided into 64 cases with high risk and 64 cases with low risk, 54 patients in the validation dataset were divided into 35 cases with high risk and 19 cases with low risk. The area under curve (AUC) of radiomics prediction model in predicting 18-, 24-, 30-month overall survival rate of patients in the training dataset was 0.83[95% confidence interval ( CI ) as 0.76-0.87, P<0.05], 0.84(95% CI as 0.73-0.91, P<0.05), 0.91(95% CI as 0.78-0.95, P<0.05), respectively. The AUC of radiomics prediction model in predicting 18-, 24-, 30-month overall survival rate of patients in the validation dataset was 0.84(95% CI as 0.75-0.92, P<0.05), 0.84 (95% CI as 0.73-0.91, P<0.05), 0.86(95% CI as 0.82-0.94, P<0.05), respectively. (2) Analysis of prognostic factors for patients with GNEN in the training dataset. Results of multivariate analysis showed gender, age, treatment method, tumor boundary, tumor T staging, tumor N staging, tumor M staging, Ki-67 index, CD56 expression were independent factors influencing prognosis of patients with GNEN in the training dataset ( P<0.05). (3) Construction and evaluation of the prediction model for prognosis of patients with GNEN. The clinical prediction model was constructed based on the independent factors influen-cing prognosis of patients with GNEN including gender, age, treatment method, tumor boundary, tumor T staging, tumor N staging, tumor M staging, Ki-67 index, CD56 expression. The C-index value of clinical prediction model in the training dataset and the validation dataset was 0.86 (95% CI as 0.82-0.90) and 0.80(95% CI as 0.72-0.87), respectively. The C-index value of radiomics prediction model in the training dataset and the validation dataset was 0.80 (95% CI as 0.74-0.86, P<0.05) and 0.75(95% CI as 0.66-0.84, P<0.05), respectively. The C-index value of clinical-radiomics combined prediction model in the training dataset and the validation dataset was 0.88(95% CI as 0.85-0.92) and 0.83 (95% CI as 0.77-0.89), respectively. Results of calibration curve show that clinical prediction model, radiomics prediction model and clinical-radiomics combined prediction model had good predictive ability. Results of decision curve show that the clinical-radiomics combined prediction model is superior to the clinical prediction model, radiomics prediction model in evaluating the prognosis of patients with GNEN. Conclusions:The predection model for predicting the prognosis of patients with GNEN is constructed based on 14 radiomics features after selecting. The prediction model can predict the prognosis of patients with GNEN well, and the clinical-radiomics combined prediction model has a better prediction efficiency.
RESUMO
Objective:To explore the value of multiphasic CT-based radiomics signature in predicting the invasive behavior of pancreatic solid pseudopapillary neoplasm (pSPN).Methods:The multiphasic CT images of patients with pSPN confirmed by postoperative pathology in the First Affiliated Hospital of Zhengzhou University from January 2012 to January 2021 were analyzed retrospectively. There were 23 cases of invasiveness and 59 cases of non-invasiveness. The region of interest(ROI) was artificially delineated layer by layer in the plain scan, arterial-phase and venous-phase images, respectively. The 1 316 image features were extracted from each ROI. The data set was divided into training and validation sets with a ratio of 7∶3 by stratified random sampling, and synthetic minority oversampling technique (SMOTE) algorithm was used for oversampling in the training set to generate invasive and non-invasive balanced data for building the training model. The constructed model was validated in the validation set. The receiver operating characteristic(ROC) analysis was used to evaluate model performance and the Delong′s test was applied to compare the area under the ROC curve (AUC) of different predict models. The improvement for classification efficiency of each independent model or their combinations were also assessed by net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices.Results:After feature extraction, 2, 6 and 3 features were retained to construct plain-scanned model, arterial-phase and venous-phase models, respectively. Seven independent-phase and combined-phase models were established. Except the plain-scanned model, the AUC values of other models were greater than 0.800. The arterial-phase model had the best efficiency for classification among all independent-phase models. The AUC values of arterial-phase model in the SMOTE training and validation sets were 0.913 and 0.873, respectively. By combining the radiomics signature of the arterial-phase and venous-phase models, the AUC values of training and validation sets increased to 0.934 and 0.913 respectively. There were no significant differences of the AUC values between the scan-arterial venous-phase model and arterial venous-phase model in both training and validation sets (both P>0.05). The NRI and IDI indexes showed that the combined form of plain-scan model and arterial-venous-phase model could not significantly improve the classification efficiency in the validation set (both NRI and IDI<0). Conclusions:The arterial-phase CT-based radiomics model has a good predictive performance in the invasive behavior of pSPN, and the combination with a venous-phase radiomics model can further improve the model performance.