Predicting the efficacy of chemoradiotherapy in advanced nasopharyngeal carcinoma patients: an MRI radiomics and machine learning approach

Abstract

Background: Machine learning methods play an important role in predicting the efficacy of chemoradiotherapy in patients with nasopharyngeal carcinoma (NPC). This study explored the predictive value of machine learning models based on multimodal magnetic resonance imaging (MRI) radiomic features for the efficacy in patients with advanced NPC after clinical chemoradiotherapy.

Methods: A retrospective analysis was conducted on data from 160 diagnosed patients with NPC confirmed by pathology at the First Affiliated Hospital of Bengbu Medical College. Patients were divided into effective group (n=116) and noneffective group (n=44) according to the Response Evaluation Criteria in Solid Tumors 1.1 (RECIST 1.1). After the overall Synthetic Minority Over-sampling Technique (SMOTE) sample balance, the proportion of effective group and invalid group is 1:1, both 116 cases, the total sample number is 232 cases. The region of interest (ROI) depicting the maximum solid component of the tumor on T2-weighted imaging short time inversion recovery (T2WI-STIR), contrast-enhanced T1-weighted imaging (CE-T1WI), and diffusion-weighted imaging (DWI) images was delineated, and radiomic features were extracted. Feature selection was performed through least absolute shrinkage and selection operator (LASSO) ridge regression, and based on the selected features, six machine learning models including random forest (RF), Extreme Gradient Boosting (XGBoost), support vector machine (SVM), logistic regression (LR), Light Gradient Boosting Machine (LGB) and K-nearest neighbor (KNN) were constructed. The model performance of the training set was verified by using the 5-fold cross-validation method, and the effect evaluation and performance visualization were performed on the test set. After that, the SHAP plot was established based on the feature weights, and finally the benefit degree of patients was analyzed using the DCA curve.

Results: A total of 3375 radiomic features were extracted, and 25 important features were selected after feature extraction to establish six machine learning models. The RF model exhibited the highest performance, achieving an AUC of 0.801, accuracy of 0.800, precision of 0.844, recall of 0.750, and F1 score of 0.794 within the test set. DCA results showed that patients could get good benefits.

Conclusions: The machine learning model based on multimodal MRI radiomic features may serve as a promising tool for predicting the efficacy of chemoradiotherapy in patients with advanced NPC.

Keywords: efficacy prediction; machine learning; magnetic resonance imaging; nasopharyngeal carcinoma; radiomics.

Figure 1

Manual outline of the ROI schematic based on the maximum tumor level. (A–C) Results of manual lesion segmentation on T2WI-STIR, CE-T1WI and DWI images, respectively.

Figure 2

Identification of radiomics features using LASSO screening. (A) Curve of regression coefficient with coefficient α. (B) LASSO regression mean square change curve for each fold.

Figure 3

ROC curves of six machine learning models on the validation set (A), and test set (B).

Figure 4

The ROC curve is validated by the random forest model with a 5-fold cross.

Figure 5

Decision curve of RF model test set.

Figure 6

SHAP summary plot of RF model. The x-axis represents the SHAP values, reflecting the impact of each feature on the model’s predictions.