Development of a Nomogram Based on Preoperative Bi-Parametric MRI and Blood Indices for the Differentiation Between Cystic-Solid Pituitary Adenoma and Craniopharyngioma

Abstract

Background: Given the similarities in clinical manifestations of cystic-solid pituitary adenomas (CS-PAs) and craniopharyngiomas (CPs), this study aims to establish and validate a nomogram based on preoperative imaging features and blood indices to differentiate between CS-PAs and CPs.

Methods: A departmental database was searched to identify patients who had undergone tumor resection between January 2012 and December 2020, and those diagnosed with CS-PAs or CPs by histopathology were included. Preoperative magnetic resonance imaging (MRI) features as well as blood indices were retrieved and analyzed. Radiological features were extracted from the tumor on contrast-enhanced T1 (CE-T1) weighted and T2 weighted sequences. The two independent samples t-test and principal component analysis (PCA) were used for feature selection, data dimension reduction, and radiomics signature building. Next, the radiomics signature was put in five classification models for exploring the best classifier with superior identification performance. Multivariate logistic regression analysis was then used to establish a radiomic-clinical model containing radiomics and hematological features, and the model was presented as a nomogram. The performance of the radiomics-clinical model was assessed by calibration curve, clinical effectiveness as well as internal validation.

Results: A total of 272 patients were included in this study: 201 with CS-PAs and 71 with CPs. These patients were randomized into training set (n=182) and test set (n=90). The radiomics signature, which consisted of 18 features after dimensionality reduction, showed superior discrimination performance in 5 different classification models. The area under the curve (AUC) values of the training set and the test set obtained by the radiomics signature are 0.92 and 0.88 in the logistic regression model, 0.90 and 0.85 in the Ridge classifier, 0.88 and 0.82 in the stochastic gradient descent (SGD) classifier, 0.78 and 0.85 in the linear support vector classification (Linear SVC), 0.93 and 0.86 in the multilayers perceptron (MLP) classifier, respectively. The predictive factors of the nomogram included radiomic signature, age, WBC count, and FIB. The nomogram showed good discrimination performance (with an AUC of 0.93 in the training set and 0.90 in the test set) and good calibration. Moreover, decision curve analysis (DCA) demonstrated satisfactory clinical effectiveness of the proposed radiomic-clinical nomogram.

Conclusions: A personalized nomogram containing radiomics signature and blood indices was proposed in this study. This nomogram is simple yet effective in differentiating between CS-PAs and CPs and thus can be used in routine clinical practice.

Keywords: craniopharyngioma; machine learning; nomogram; pituitary adenoma; predictive model; radiomics.

Figure 1

The flowchart of patient selection.

Figure 2

The overall workflow of radiomics processing and nomogram construction.

Figure 3

The predictive performance of distinguishing between CS-PAs and CPs in different classifiers. (A) The receiver operating characteristic curve (ROC) and the area under the curve (AUC) of the five different classifiers are showed in the training set, respectively. (B) The ROC and AUC of the five different classifiers are showed in the test set, respectively. SGD Classifier, stochastic gradient descent classifier; Linear SVC, linear support vector classification; MLP Classifier, multilayers perceptron classifier.

Figure 4

Developed radiomic-clinical nomogram. The nomogram, incorporated radiomics signature, age, WBC count and FIB, was developed in the training set. The risk represents the predictive probability of CS-PAs.

Figure 5

Calibration curve of the radiomic-clinical nomogram in the training and test sets. (A) Calibration curve of the radiomic-clinical nomogram in the training set. (B) Calibration curve of the radiomic-clinical nomogram in the test set. The calibration curve showed the calibration of the models in terms of the consistency between the predictive performance of CS-PAs and the actual results observed for calibration. The Y-axis represents the actual performance, and the X-axis represents the performance predicted by the nomogram. The oblique dashed line represents the perfect prediction by an ideal model. The red and green solid lines represent the performance of the nomogram in the training set and the test set, respectively. In addition, a fit closer to the diagonal dashed line indicates a better prediction. (The Hosmer–Lemeshow test showed p=0.367 and p=0.113 in the training and test set, respectively).

Figure 6

Decision curve analysis for the radiomic-clinical nomogram, radiomics model and clinical model. The decision curve showed that if the threshold probability was higher than 20%, then using the radiomic-clinical nomogram to differentially diagnose CS-PAs and CPs has a greater advantage than using a radiomics model and simple clinical model in terms of clinical application. Clinical data, clinical model; Radiomics, radiomics model; Radiomic+Clinical data, radiomic-clinical nomogram.