CT-based dosiomics and radiomics model predicts radiation-induced lymphopenia in nasopharyngeal carcinoma patients

Abstract

Purpose: This study aims to develop and validate a model predictive for the incidence of grade 4 radiation-induced lymphopenia (G4RIL), based on dosiomics features and radiomics features from the planning CT of nasopharyngeal carcinoma (NPC) treated by radiation therapy.

Methods: The dataset of 125 NPC patients treated with radiotherapy from August 2018 to March 2019 was randomly divided into two sets-an 85-sample training set and a 40-sample test set. Dosiomics features and radiomics features of the CT image within the skull bone and cervical vertebrae were extracted. A feature selection process of multiple steps was employed to identify the features that most accurately forecast the data and eliminate superfluous or insignificant ones. A support vector machine learning classifier with correction for imbalanced data was trained on the patient dataset for prediction of RIL (positive classifier for G4RIL, negative otherwise). The model's predictive capability was gauged by gauging its sensitivity (the likelihood of a positive test being administered to patients with G4RIL) and specificity in the test set. The area beneath the ROC curve (AUC) was utilized to explore the association of characteristics with the occurrence of G4RIL.

Results: Three clinical features, three dosiomics features, and three radiomics features exhibited significant correlations with G4RIL. Those features were then used for model construction. The combination model, based on nine robust features, yielded the most impressive results with an ACC value of 0.88 in the test set, while the dosiomics model, with three dosiomics features, had an ACC value of 0.82, the radiomics model, with three radiomics features, had an ACC value of 0.82, and the clinical model, with its initial features, had an ACC value of 0.6 for prediction performance.

Conclusion: The findings show that radiomics and dosiomics features are correlated with the G4RIL of NPC patients. The model incorporating radiomics features and dosiomics features from planning CT can predict the incidence of G4RIL in NPC patients.

Keywords: dosiomics; machine learning; nasopharyngeal carcinoma; radiation-induced lymphopenia; radiomics.

Figure 1

Study workflow overview. (A) Data acquisition; (B) Segmentation of the region of interest by radiologists; (C) Feature extraction including clinical characteristics, radiomics feature, and dosiomic feature; (D) Model building and validation. Abbreviation: HIS, hospital information system; DVH, dose volume histogram.

Figure 2

Correlation analysis of the features used in the model, there is no correlation between these features. R0- R2 are radiological features; D3- D5 are dosiomic features; C6- C8 are clinical features. R0, wavelet-LHL_glcm_Idn; R1, logarithm_glszm_Gray Level NonUniformity Normalized; R2, wavelet-HHL_glcm_Maximum Probability; D3, original_shape_Major Axis Length; D4, log-sigma-4-0-mm-3D_glszm_Small Area Emphasis; D5, wavelet-LLH_firstorder_Mean; C6, Age; C7, baseline_ALC; C8, Volume of GTVnx.

Figure 3

The features were selected using the LASSO regression model. (A) Selection of the regulation parameter lambda (λ). The vertical black dashed line defines the optimal λ at the minimum MSE. (B) LASSO coefficient curves of features. Vertical black dashed lines are drawn at the best lambda in (A), non-zero features under the best λ are selected. Abbreviation: LASSO, the least absolute shrinkage and selection operator; MSE, Mean Squared Error.

Figure 4

Performance of the ROC curves. (A) ROC curves of four models were compared using 5-fold CV in the training set. (B) ROC curves of four models were compared in the test set. (C) 5-fold CV ROC curves for the combined model in the training set. (D) 10-fold cross-validation ROC curves for the combined model in the training set. (E) ROC curve for the combined model in the test set. Abbreviation: ROC, the receiver operating characteristic; AUC, the area under the curve; CV, cross-validation.