Fig 1.
The overall workflow of the present study designed to achieve two major objectives: Construction of the BN-based signature and exploration of synergistic combination with known predictive factors.
This study consisted of two main sections. First, the most feasible set of image features (signature) was determined by evaluating their performance in the prediction of RP+. Then, the combination of the image signature and well-known predictive factors including KL-6 and DVI was explored for further improvement of the predictive performance of the model. pCT: planning computed tomography, BN: Betti number, WD: wavelet decomposition, GLSZM: gray-level size zone matrix, GLCM: gray-level co-occurrence matrix, GLDM, gray-level dependence matrix, GLRLM: gray-level run-length matrix, NGTDM: neighborhood gray-tone difference matrix, RP: radiation-induced pneumonitis. LASSO: least absolute shrinkage and selection operator, SVM: support vector machine, KL-6: Krebs von den Lungen-6, DVI: dose volume indices.
Fig 2.
Scheme of the imbalance data adjustment of patients with symptomatic (grade ≥2) radiation pneumonitis positive (RP+) and negative (RP−) for a voting-based feature selection and an ensemble support vector machine (SVM) model construction.
LASSO-LR: least absolute shrinkage and selection operator logistic regression; SVM: support vector machine; AUC: area under the receiver operating characteristics curve; RI: robustness index.
Table 1.
Demographic distributions of the patients and significant differences between the training and test datasets.
Table 2.
Demographic distributions of the patients, whose pretreatment data of serum Krebs von den Lungen-6 (KL-6 database) was available, and significant differences between the training and test datasets.
Fig 3.
Selected features obtained from voting with LASSO logistic regression models based on (a) Betti number (BN), (b) planning computed tomography (pCT), (c) wavelet decomposition (WD) features, (d) a combination of all feature types (BN+pCT+WD) and a combination of two conventional feature types (pCT+WD).
Table 3.
Summary of the areas under the receiver operating characteristics curves (AUCs) and robustness indices (RIs) obtained between the validation and test for the image signatures.
Fig 4.
Comparison of the areas under the receiver operating characteristics curves (AUCs) among support vector machine models for (a) the validation and (b) the test with 95% confidence interval indicated by error bars. Bar graphs on the left side represent the results for the exploration of the best imaging biomarker, and those on the right side represent the results for the investigation of complementarity with known predictive factors.
Table 4.
Summary of the areas under receiver operating characteristics curves (AUCs) and robustness indices (RIs) obtained between the validation and test for the Betti number (BN)-based signatures with serum Krebs von den Lungen-6 (KL-6) or dose volume indices (DVI).
Table 5.
Comparison of areas under the receiver operating characteristics curves (AUCs) obtained in the present study to the AUCs obtained in previous studies.
Table 6.
Relationships between the predictions using the BN+KL-6 model and the pre-existing pulmonary diseases in the training dataset.
Fig 5.
Distributions of the pretreatment serum Krebs von den Lungen-6 (KL-6) and support vector machine (SVM) outputs based on BN+KL-6.
Representative features and Betti number (BN, b0) maps were shown for the cases where the SVM model could successfully predict the radiation pneumonitis (RP) status [positive (+) or negative (–)]. Labels of the BN maps (b0) are followed by the threshold value for obtaining the BN maps. Hist: histogram, NGTDM: neighborhood gray-tone difference matrix, GLSZM: gray-level size zone matrix.