.png){kind=icon}
.png){kind=icon}
Return
Combining handcrafted radiomics features and deep learning features within different regions of interest in planning CT images for the prediction of symptomatic radiation pneumonitis in locally advanced non-small cell lung cancer
.png){kind=icon}
OA
AI
X
H
Y
J
M
Z
Y
F
Q
J
J
DOI：10.1186/s13014-026-02896-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To predict symptomatic radiation pneumonitis (RP) in locally advanced non-small cell lung cancer (LA-NSCLC) patients by handcrafted radiomics (HCR) and deep learning (DL) features of three different regions of interest (ROIs) in normal lung tissue. Ninety LA-NSCLC patients from Center A were collected as the development cohort, and an additional 57 patients from Center B were collected as the external validation set. Clinical data and lung dose-volume parameters were collected from the patients, and univariate analysis was conducted. Three ROIs in lung were delineated in pre-radiotherapy CT images: Lung-PTV, Lung-GTV, and PTV-GTV. For each ROI, 1046 HCR features and 512 DL features were respectively extracted, and then correlated. For each ROI, three models (HCR, DLR, and a hybrid model) were developed with feature selection via Spearman correlation, mRMR, the LASSO method, and SVM-based modeling. And finally, their performance was evaluated and compared using AUC values from ROC curves. The optimal model was selected by comparing the best AUC through internal validation results, and then external validation was conducted. Calibration and decision curve analyses were used to evaluate the model. In the development cohort, univariate analysis revealed no statistically significant associations with RP. The correlation analysis revealed a weak correlation between HCR and DL features. In internal validation, for Lung-PTV, Lung-GTV and PTV-GTV, HCR models achieved AUCs of 0.658 (95%CI: 0.645–0.671), 0.659 (95%CI: 0.646–0.672), and 0.890 (95%CI: 0.881–0.899); DLR models obtained AUCs of 0.832 (95%CI: 0.824–0.840), 0.900 (95%CI: 0.893–0.907), and 0.777 (95%CI: 0.766–0.788); and the hybrid model showed AUCs of 0.838 (95%CI: 0.831–0.846), 0.901 (95%CI: 0.894–0.908), and 0.919 (95%CI: 0.912–0.926). The results showed that the DLR model outperformed the HCR model for Lung-PTV and Lung-GTV, and the hybrid model surpassed both for all three ROIs. The hybrid model of PTV-GTV was found to be the best, and its external validation was conducted. The results showed that the AUC for the external validation set was 0.828 (95%CI: 0.696–0.960). The model was well calibrated (p > 0.05), and decision curve analysis indicated positive net benefit across relevant threshold probabilities. Hybrid models combining HCR and DL features showed the highest predictive performance in internal comparisons and the best-performing PTV-GTV hybrid model was externally validated, highlighting the potential of combined features for improved prediction.
Keywords:
Radiation pneumonitis prediction
Radiomics
Deep learning features
Machine learning
Journal
IF:
3.2
Papers:
3.9K
Citations:
9.1K
Organization
Cited Papers