A Novel Magnetic Resonance Imaging-Based Radiomics and Clinical Predictive Model for the Regrowth of Postoperative Residual Tumor in Non-Functioning Pituitary Neuroendocrine Tumor

Abstract

Background and Objectives: To develop a novel magnetic resonance imaging (MRI)-based radiomics–clinical risk stratification model to predict the regrowth of postoperative residual tumors in patients with non-functioning pituitary neuroendocrine tumors (NF-PitNETs). Materials and Methods: We retrospectively enrolled 114 patients diagnosed as NF-PitNET with postoperative residual tumors after the first operation, and the diameter of the tumors was greater than 10 mm. Univariate and multivariate analyses were conducted to identify independent clinical risk factors. We identified the optimal sequence to generate an appropriate radiomic score (Rscore) that combined pre- and postoperative radiomic features. Three models were established by logistic regression analysis that combined clinical risk factors and radiomic features (Model 1), single clinical risk factors (Model 2) and single radiomic features (Model 3). The models’ predictive performances were evaluated using receiver operator characteristic (ROC) curve analysis and area under curve (AUC) values. A nomogram was developed and evaluated using decision curve analysis. Results: Knosp classification and preoperative tumor volume doubling time (TVDT) were high-risk factors (p < 0.05) with odds ratios (ORs) of 2.255 and 0.173. T1WI&T1CE had a higher AUC value (0.954) and generated an Rscore. Ultimately, the AUC of Model 1 {0.929 [95% Confidence interval (CI), 0.865–0.993]} was superior to Model 2 [0.811 (95% CI, 0.704–0.918)] and Model 3 [0.844 (95% CI, 0.748–0.941)] in the training set, which were 0.882 (95% CI, 0.735–1.000), 0.834 (95% CI, 0.676–0.992) and 0.763 (95% CI, 0.569–0.958) in the test set, respectively. Conclusions: We trained a novel radiomics–clinical predictive model for identifying patients with NF-PitNETs at increased risk of postoperative residual tumor regrowth. This model may help optimize individualized and stratified clinical treatment decisions.