The Clinical Problem. Radiation pneumonitis (RP) is a potentially life-threatening inflammatory reaction of normal lung tissue that occurs following thoracic radiotherapy. In lung cancer patients, RP significantly reduces quality of life and can force treatment interruptions or dose reductions, limiting therapeutic efficacy.
Incidence and Severity. RP occurs in 15 to 40% of lung cancer patients undergoing radiotherapy, with severe cases (grade 2 or higher) requiring corticosteroid treatment and sometimes leading to respiratory failure. Early identification of high-risk patients could enable prophylactic interventions and individualized dose constraints.
Limitations of Current Approaches. Existing prediction models rely primarily on dosimetric parameters such as mean lung dose and V20 (volume receiving 20 Gy or more). While useful, these parameters do not capture the full biological complexity of individual patient susceptibility, tumor biology, or functional lung status before radiotherapy.
Multi-Omics Opportunity. PET/CT imaging captures both anatomical structure (CT) and metabolic activity (PET) simultaneously. Combining radiomic features from both modalities with dosiomics (dose distribution features) and deep learning representations offers a comprehensive multi-omics framework that may substantially improve RP prediction over dose-based models alone.
Study Cohort. This retrospective multicenter study enrolled 206 lung cancer patients from Hospital One (training and internal validation) and 27 patients from Hospital Two (external validation), all treated with radiotherapy between 2018 and 2023. RP was defined as grade 2 or higher occurring within 6 months of treatment completion.
RP Incidence. Among the internal cohort, 23.3% developed grade 2 or higher RP, while 11.1% developed RP in the external cohort. This class imbalance required careful handling to prevent models from defaulting to majority-class predictions.
Imaging Protocol. All patients underwent pretreatment 18F-FDG PET/CT scans. CT provided anatomical tumor delineation, while PET provided metabolic information including standardized uptake values (SUV). Both modalities were co-registered, enabling direct voxel-level correspondence between functional and structural imaging features.
Dosiomics Feature Extraction. Beyond standard radiomics and deep learning features, dosiomics features were extracted from the three-dimensional dose distribution maps calculated during treatment planning. These features quantify spatial dose heterogeneity within the treatment volume, extending prediction beyond mean dose metrics to capture complex dose-volume relationships.
Eight Model Types. To systematically evaluate information sources, eight model configurations were tested: CT radiomics only (R_CT), PET radiomics only (R_PET), image-fused PET/CT radiomics (R_iFU), feature-fused PET/CT radiomics (R_fFU), dosiomics only (D), CT deep learning (DL_CT), PET deep learning (DL_PET), and fused PET/CT deep learning (DL_FU). This comprehensive comparison identified which data sources contribute independent predictive value.
Four Machine Learning Algorithms. Logistic regression (LR), decision tree (DT), support vector machine (SVM), and XGBoost were each trained and compared across all eight model configurations. This factorial design allowed identification of both the best data modality and the best algorithm, rather than assuming either in advance.
Deep Learning with ResNet-18. For deep learning feature extraction, ResNet-18 was applied to CT and PET images separately and in fusion. The model was pretrained and fine-tuned on the lung cancer cohort, extracting high-level spatial features from the tumor region that complement hand-crafted radiomic features.
Feature Fusion Strategy. For combined PET/CT models, image-level fusion merged the two modalities before feature extraction (R_iFU), while feature-level fusion concatenated separately extracted features (R_fFU). The optimal CT to PET weighting ratio of 6:4 was determined empirically, reflecting the complementary but unequal contributions of anatomical and metabolic information.
Best Radiomics Performance. Among all single-modality and fused radiomics models, XGBoost with image-fused PET/CT (R_iFU) achieved the highest internal validation AUC of 0.86, substantially outperforming CT-only (AUC 0.79) and PET-only (AUC 0.77) models. This confirmed that combining metabolic and anatomical imaging improves RP risk stratification beyond either modality alone.
Deep Learning Results. ResNet-18 with fused PET/CT (DL_FU) achieved the best deep learning AUC of 0.81 on internal validation, again outperforming single-modality DL_CT (AUC 0.76) and DL_PET (AUC 0.73). The consistent advantage of PET/CT fusion over single-modality approaches held across both radiomic and deep learning paradigms.
Dosiomics Contribution. The dosiomics-only model with XGBoost achieved an internal validation AUC of 0.81, comparable to the best deep learning results and substantially above dose-based clinical parameters alone. This demonstrates that spatial dose distribution features carry significant independent predictive information beyond mean lung dose and V20.
XGBoost Consistently Superior. Across all eight model types, XGBoost outperformed logistic regression, decision tree, and SVM. XGBoost's gradient boosting framework better handles the high-dimensional, imbalanced, and partially redundant feature sets typical of multi-omics medical datasets compared to simpler linear or kernel-based classifiers.
Integrated R+D+DL Model. The combined model integrating radiomics, dosiomics, and deep learning features with XGBoost achieved an AUC of 0.93 on the training cohort, 0.92 on internal validation, and 0.89 on the external validation cohort. Accuracy was 0.82, specificity 0.67, and sensitivity 0.83 on external validation, demonstrating robust generalization across institutions.
Incremental Value of Each Modality. Sequential model building confirmed that each data source contributed independently: adding dosiomics to radiomics improved AUC from 0.86 to 0.88, and further adding deep learning features raised AUC from 0.88 to 0.92 on internal validation. No single modality achieved performance comparable to the combined model.
SHAP Feature Importance Analysis. SHapley Additive exPlanations (SHAP) analysis revealed that deep learning features provided the highest overall contribution to predictions. The feature designated deep.7 had the highest individual SHAP importance at 10.0%, followed by additional deep learning and radiomic features. This confirms that ResNet-18 extracted biologically meaningful representations beyond what hand-crafted radiomics could capture.
Nomogram with Clinical Variables. A nomogram incorporating gender, adaptive radiotherapy status, SUVp90 (the 90th percentile SUV from PET), and the XGBoost model score achieved an AUC of 0.94 on the training cohort. SUVp90, reflecting peak metabolic activity in the tumor, emerged as the most predictive PET-derived clinical variable.
Why Multi-Omics Outperforms Single Sources. Each imaging modality captures a distinct aspect of tumor biology: CT reveals anatomical structure and tissue density, PET captures glucose metabolism reflecting tumor activity and hypoxia, and dosiomics quantifies the spatial dose heterogeneity that normal lung tissue must absorb. No single modality can substitute for this complementary information in predicting which patients will develop inflammatory RP.
SUVp90 as a Key Biomarker. The 90th percentile SUV emerged as the most clinically interpretable PET-derived predictor. Unlike mean SUV, which averages metabolic activity across the entire tumor, SUVp90 captures the most metabolically active subregion, likely reflecting hypoxic and radio-resistant tumor areas that demand higher effective doses and consequently expose surrounding lung to greater radiation stress.
Adaptive Radiotherapy as a Protective Factor. The inclusion of adaptive RT as a predictor in the nomogram reflects clinical reality: patients treated with adaptive radiotherapy receive dose adjustments based on tumor changes during treatment, potentially reducing unnecessary dose to normal lung tissue. Its independent protective association with lower RP risk supports expanding adaptive RT protocols for high-risk patients.
Comparison with Prior Work. Previous RP prediction models using radiomics alone typically achieved AUCs of 0.70 to 0.80 in external validation. The multi-omics framework here reached AUC 0.89 externally, representing a clinically meaningful improvement that could alter risk classification for a meaningful proportion of patients.
Key Achievements. This study demonstrated that integrating radiomic, dosiomics, and deep learning features from pretreatment PET/CT with XGBoost achieves external validation AUC of 0.89 for radiation pneumonitis prediction in lung cancer patients. The nomogram incorporating clinical variables further improved performance to AUC 0.94, providing a practical clinical tool.
Study Limitations. The external validation cohort of only 27 patients is small, limiting statistical confidence in generalizability. The retrospective design and concentration in Chinese medical centers may restrict applicability to other populations and treatment environments. Differences in PET/CT scanner protocols across institutions could introduce feature instability not fully addressed by preprocessing.
Future Directions. Prospective multicenter validation with larger and ethnically diverse cohorts is the most urgent next step. Integration of additional clinical variables (pulmonary function tests, pre-existing lung disease), genomic biomarkers, and longitudinal imaging during radiotherapy are identified as promising extensions. Automating tumor segmentation would also reduce inter-reader variability in radiomic feature extraction.
Clinical Impact. By identifying patients at high RP risk before radiotherapy begins, this framework could guide adaptive dose planning, trigger prophylactic anti-inflammatory therapy, or prompt consideration of proton beam therapy for high-risk cases. Ultimately, pretreatment RP risk stratification could meaningfully improve the therapeutic ratio of lung cancer radiotherapy.