Study on prediction of radiotherapy response in non-small cell lung cancer using machine learning models based on localization CT-based radiomics, dosiomics and clinical features
10.3760/cma.j.cn115355-20250122-00052
- VernacularTitle:基于定位CT影像组学、放疗剂量组学及临床特征的机器学习模型预测非小细胞肺癌放疗反应的研究
- Author:
Shuang GE
1
;
Peijun ZHU
;
Qiang DING
;
Jun MA
;
Aiping ZHANG
;
Jing ZHANG
;
Junli MA
;
Xun WANG
;
Shucheng YE
Author Information
1. 济宁医学院附属医院肿瘤放疗科，济宁 272000
- Publication Type:Journal Article
- Keywords:
Carcinoma, non-small-cell lung;
Radiotherapy;
Treatment outcome;
Radiomics;
Dosiomics
- From:
Cancer Research and Clinic
2025;37(10):743-751
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To construct a machine learning model based on localization CT-based radiomics, dosiomics and clinical features for predicting radiotherapy response in non-small cell lung cancer (NSCLC) and validate its application value.Methods:A retrospective case series study was conducted. A total of 138 NSCLC patients who received radiotherapy at the Affiliated Hospital of Jining Medical University from January 2016 to December 2022 were selected. The efficacy was evaluated according to the Response Evaluation Criteria in Solid Tumors (RECIST) 1.1, and the patients were stratified according to the objective remission (complete remission+partial remission). Random stratified sampling was used to divide the 138 patients into a training group (96 cases) and an internal validation group (42 cases) at a ratio of 7∶3. Additionally, 33 patients who received radiotherapy at Jining Cancer Hospital from January 2019 to December 2022 were included as the external validation group. Based on the pre-radiotherapy data of the radiotherapy planning system, PyRadiomics software package was used to extract 107 radiomics features and 107 dosiomics features for each patient. Pearson correlation analysis and LASSO regression analysis were used for dimensionality reduction screening; the final selected features were weighted and integrated to generate radiomics-dosiomics scores (RDS), which were then input into logistic regression (LR), support vector machine (SVM), extremely randomized forest (Extra Trees), K-nearest neighbor algorithm (KNN), lightweight gradient boosting machine (Light GBM), and multi-layer perceptron (MLP) machine learning algorithms to construct 6 radiomics-dosiomics models (RDM) for predicting the objective remission. RECIST 1.1 standard was used to evaluate objective remission as the gold standard, receiver operating characteristic (ROC) curve of 6 RDM for predicting objective remission was plotted, and the optimal algorithm for RDM was selected. Univariate and multivariate logistic regression were performed on demographic characteristics, hematological indicators and radiotherapy parameters of the training group to screen independent risk factors for NSCLC patients who received radiotherapy but did not achieve objective remission. These factors were input into the optimal machine learning algorithm to construct a clinical model (CM). Combined with features from RDS and CM, the clinical feature-radiomics-dosiomics combined model (CRDM) was established, and the nomogram of the model for predicting objective remission in NSCLC patients with radiotherapy was drawn. ROC curves were used to evaluate the efficacy of CM, RDM and CRDM in predicting the objective remission in NSCLC patients with radiotherapy in the training group, internal validation group and external validation group.Results:Four radiomics features (including grayscale variance, low grayscale long-range operation emphasis, low grayscale area emphasis, and small area low grayscale area emphasis, all of which were texture features) and 6 dosiomics features [including 1 first-order feature (robust mean absolute deviation), 4 texture features (grayscale non-uniformity, large area emphasis, large area high grayscale emphasis, contrast) and 1 shape feature (shortest axis length)] were selected. ROC curve analysis showed that the area under the curve (AUC) of the RDM constructed using SVM algorithm for judging the objective remission in the training group and the internal validation group was 0.907 (95% CI: 0.836-0.977) and 0.822 (95% CI: 0.685-0.959), which were higher than RDM constructed using other algorithms, and the sensitivity (96.2% and 91.7%), specificity (78.6% and 76.7%) and accuracy (83.3% and 81.0%) at the optimal cut-off values were all higher. Considering the stability and generalization ability of the model, SVM algorithm was ultimately used to construct RDM, CM and CRDM uniformly. Based on training group data, univariate and multivariate logistic regression analysis showed that elevated platelet-to-lymphocyte ratio (PLR) ( OR = 1.001, 95% CI: 1.000-1.003, P = 0.035) and increased target volume of radiotherapy plan ( OR = 1.001, 95% CI: 1.000-1.001, P = 0.008) were independent risk factors for failure to achieve objective remission. ROC curve analysis showed that in the training group and the internal validation group, the AUC of CRDM predicting objective remission were 0.914 (95% CI: 0.856-0.972) and 0.864 (95% CI: 0.754-0.974), respectively, which were better than CM [AUC were 0.735 (95% CI: 0.612-0.857) and 0.697 (95% CI: 0.507-0.888)] and RDM, respectively. In the external validation group, the AUC of CRDM, CM and RDM were 0.778 (95% CI: 0.500-1.000), 0.667 (95% CI: 0.434-0.899) and 0.741 (95% CI: 0.463-1.000), respectively. Conclusions:The CRDM constructed by combining radiomics, dosiomics and clinical features can comprehensively and accurately evaluate the radiotherapy response of NSCLC patients, and may have important clinical application value in achieving precision medicine and optimizing treatment strategies.
