Objective To explore the value of machine-learning models based on magnetic resonance imaging (MRI) radiomics features for the preoperative prediction of the blood supply in pituitary neuroendocrine tumors.

Methods  A retrospective analysis was performed on the clinical and imaging data of 136 patients with pathologically confirmed pituitary neuroendocrine tumors (diameter >10 mm) from April 2013 to April 2023 at Yi Jishan Hospital of Wannan Medical College. Based on the intraoperative findings, the patients were assigned into richly vascularized (n=50) and normally vascularized (n=86) groups. All patients were allocated randomly in a 7:3 ratio into a training (n=96) or a validation group (n=40). Three machine-learning algorithms, multivariate Logistic regression (LR), random forest (RF), and support vector machine (SVM), were used to establish radiomics prediction models. Receiver operating characteristic (ROC) curves were plotted to evaluate the diagnostic performance of the models; decision curve analysis (DCA) was used to assess the net clinical benefit of the models.

Results  The clinical model achieved areas under the ROC curve (AUC) of 0.74 and 0.82 in the training and validation groups, respectively. The radiomics models using T1-weighted imaging (WI), T2WI, T1WI-enhanced, and combined sequences achieved AUCs of 0.80, 0.84, 0.82, and 0.84 in the training group and 0.82, 0.80, 0.85, and 0.83 in the validation group, respectively. The LR, RF, and SVM models had AUCs of 0.85, 0.87, and 0.84 in the training group and 0.85, 0.85, and 0.83 in the validation group, respectively. All radiomics models demonstrated greater diagnostic efficacy than the clinical model. DCA indicated that the LR, SVM, and combined-sequence models achieved good net clinical benefits; the LR model showed the best results.

Conclusions  Machine-learning models based on MRI radiomics exhibit high predictive value, surpassing the clinical judgment of radiologists based on MRI images alone, and offer a favorable net clinical benefit.