A logistic regression model for prediction of glioma grading based on radiomics

Abstract

Objectives: Glioma is the most common intracranial primary tumor in central nervous system. Glioma grading possesses important guiding significance for the selection of clinical treatment and follow-up plan, and the assessment of prognosis. This study aims to explore the feasibility of logistic regression model based on radiomics to predict glioma grading.

Methods: Retrospective analysis was performed on 146 glioma patients with confirmed pathological diagnosis from January, 2012 to December, 2018. A total of 41 radiomics features were extracted from contrast-enhanced T₁-weighted imaging (T₁WI+C) lesion by manual segmentation. Least absolute shrinkage and selection operator (LASSO) was used to select the most-predictive radiomics features for pathological grading and to calculate radiomics score (Rad-score) of each patient. A logistic regression model was built to explore the correlation between giloma grading and Rad-score. Receiver operating characteristic (ROC) curve was performed to evaluate the model's predictive ability with area under the curve (AUC) for the evaluation index. Hosmer-Lemeshow test was used to measure the model's predictive accuracy.

Results: A total of 5 imaging features selected by LASSO were used to establish a logistic regression model for predicting glioma grading. The model showed good discrimination with AUC value of 0.919. Hosmer-Lemeshow test showed no significant difference between the calibration curve and the ideal curve (P=0.808), indicating high predictive accuracy of the model.

Conclusions: The logistic regression model using radiomics exhibits a relatively high accuracy for predicting glioma grading, which may serve as a complementary tool for preoperative prediction of giloma grading.

Keywords: glioma; grading; least absolute shrinkage and selection operator; logistic regression; radiomics.

图1

LASSO回归变量筛选图 Figure 1 Feature selection using LASSO A: Variation of AUC. The horizontal axis shows logλ and the vertical axis shows the AUC. The numbers above the curve represent the number of feature of nonzero coefficient. The left dotted line represents the feature number corresponding to the maximum AUC, while the right dotted line represents that corresponding to 1 standard error of AUC. B: The shrinkage plot of coefficients. The horizontal axis shows logλ and the vertical axis shows coefficient. The numbers above the curve represent the number of features with nonzero coefficient.

图2

Rad-score瀑布图及ROC曲线 Figure 2 Waterfall plot and ROC curve A: Waterfall plot of Rad-score. The vertical axis shows the value of Rad-score; red bars represent patients with actual high-grade glioma; green bars represent patients with actual low-grade glioma; the transverse line represents the cut-off point in ROC curve. Bars under the transverse line represent the predicted patients with low-grade glioma, bars above the transverse line represent the predicted patients with high-grade glioma. B: ROC curve of Rad-score. The red point represents the sensitivity and specificity at the cut-off point.

图3

预测模型的校准曲线 Figure 3 Calibration curve for the predictive model The P value of Hosmer-Lemeshow test for calibration is 0.808, showing no significant difference. This result indicates good agreement between the predictive probability and the actual probability.