A life-saving combination with a serious side effect. For patients with unresectable locally advanced non-small cell lung cancer (LA-NSCLC), concurrent chemoradiotherapy (CCRT) followed by consolidative immunotherapy with anti-PD-L1 drugs like durvalumab is now standard of care. However, this regimen carries a significant risk of radiation pneumonitis (RP), a lung inflammation caused by radiation injury.
How common is the problem. In the landmark PACIFIC trial, any-grade pneumonitis occurred in 33.9% of patients receiving durvalumab after CCRT, compared to 24.8% in the placebo group. RP can force treatment interruptions, reduce the benefits of immunotherapy, and cause lasting lung damage. In this study's cohort, the RP incidence was 27.45%, closely mirroring the PACIFIC trial.
Why better prediction is urgently needed. Current predictive models rely on clinical factors such as age, smoking history, and tumor location, plus dosimetric parameters like mean lung dose and the percentage of lung receiving certain radiation doses. These models are inadequate for patients on the modern immunotherapy-plus-radiotherapy regimen, leaving clinicians without reliable early warning tools.
The radiomics opportunity. Radiomics extracts hundreds of quantitative texture and shape features from CT images that human eyes cannot reliably detect. By combining radiomics derived from specific dose regions with clinical and blood-based inflammatory markers, this study aimed to build a more accurate RP prediction model tailored to immunotherapy-era patients.
Retrospective cohort from a dedicated cancer center. The study enrolled 161 LA-NSCLC patients treated at Shandong Cancer Hospital between April 2019 and April 2023. All patients had pathologically confirmed stage IIIA to IIIC NSCLC and received curative-intent radiotherapy (50 to 70 Gy) followed by at least two cycles of consolidative immunotherapy with sintilimab or durvalumab.
Defining the outcome. RP of grade 2 or higher was identified using three criteria: a history of radiotherapy, symptoms including shortness of breath, low-grade fever, and dry cough, and CT imaging findings within the high-dose radiation field showing patches of solid lesions. Follow-up was conducted at 1, 3, and 6 months after radiotherapy, then every three months during the first year.
Training and validation split. Patients were divided in a 7:3 ratio into a training cohort of 112 patients and a validation cohort of 49 patients. Within the training cohort, 34 developed RP (30%) and 78 did not. In the validation cohort, 16 developed RP (33%) and 33 did not. The groups were well balanced on most baseline characteristics.
Key clinical predictors identified at baseline. Among all clinical variables, tumor location was the strongest differentiator: 85% of RP cases had tumors in the middle or lower lung, versus only 23% in non-RP patients. Pulmonary comorbidities such as interstitial lung disease and COPD were also significantly more common in RP patients (53% versus 22%).
Mapping radiation dose to lung tissue. The radiation treatment plan divided the lung into nine dose zones ranging from 0 to 60 Gy in 5-Gy increments. CT images from two time points were analyzed: an initial positioning scan before radiotherapy (CT1) and a resetting scan taken after the patient had received a cumulative dose of 40 to 50 Gy (CT2). This two-scan approach captures how lung tissue changes in response to radiation.
Comprehensive feature extraction. Using the MedMind Technology Imaging System, 1,810 radiomics features were extracted from each of the nine dose-zone regions of interest (ROIs) per CT scan. Features spanned seven radiomics classes including shape, intensity, and texture descriptors. A z-score normalization standardized CT values across all images before extraction.
Quality control and reliability testing. Radiologists semi-automatically contoured ROIs in 40 randomly selected patients, and inter-rater reliability was measured using the intraclass correlation coefficient (ICC). Only features with ICC greater than 0.8 were retained, ensuring that selected features can be reliably reproduced across different observers.
Delta radiomics to capture change over time. Beyond static features, the study computed delta radiomics features (DRF), measuring the percentage change in each feature between CT1 and CT2. The formula used was DRF = (RFCT2 minus RFCT1) divided by RFCT1, multiplied by 100%. This longitudinal approach captures how tissue texture evolves under radiation, which is more predictive than a single time-point measurement alone.
LASSO regression for feature reduction. With 1,810 features per dose zone across two time points, there are far more predictors than patients. The LASSO (Least Absolute Shrinkage and Selection Operator) algorithm was applied to each of the nine dose-zone regions separately, using 10-fold cross-validation to select the optimal regularization strength. LASSO forces most feature coefficients to zero, retaining only the most informative predictors.
Building radiomics signatures for each dose zone. Each dose zone's selected features were combined into a linear weighted sum called a radiomics signature (RS), producing nine signatures: RS1 through RS9 corresponding to the nine dose ranges. These signatures served as single numeric scores summarizing the radiomics information from each dose zone.
Identifying the best dose zone signature. The nine radiomics signatures were evaluated by their AUC on both the training and validation cohorts. RS8, covering the 50 to 55 Gy dose range, performed best in both sets, achieving an AUC of 0.854 in both training and validation. Other signatures showed much higher variability, with some performing well in training but poorly in validation.
Combining predictors with machine learning. To build the final combined model, 13 candidate variables with p-values below 0.1 were entered into four machine learning algorithms: Random Forest, Neural Network, Gradient Boosting Machine, and Recursive Partitioning and Regression Trees. The three variables that ranked consistently as most important across all four algorithms were RS8, tumor location, and the neutrophil-to-lymphocyte ratio at week 4 (NLR4w).
Three models compared. The study built three logistic regression models: Model 1 using only RS8, Model 2 adding tumor location to RS8, and Model 3 adding the neutrophil-to-lymphocyte ratio at week 4 (NLR4w) to both. Each model was evaluated on AUC, sensitivity, specificity, accuracy, positive predictive value, and negative predictive value.
Performance on the training cohort. Model 1 (RS8 alone) achieved an AUC of 0.854 in training. Model 2 (RS8 plus tumor location) reached 0.918, and Model 3 (all three predictors) reached 0.938. Sensitivity improved markedly with each added predictor, rising from 70.6% for Model 1 to 87.5% for Model 3 in the training set.
Validation cohort results. On the held-out validation cohort, all three models showed stable performance. Model 3 achieved an AUC of 0.869, sensitivity of 86.7%, and a negative predictive value of 92.6%, meaning that patients predicted to be RP-free had a 92.6% probability of actually remaining free of significant pneumonitis. This high NPV is clinically valuable for safely identifying low-risk patients who may not need aggressive monitoring.
Clinical utility confirmed by decision curve analysis. Decision curve analysis showed that Model 3 delivers substantial net benefit over a wide range of threshold probabilities up to 0.85, indicating that the model remains clinically useful across various risk tolerance levels that different clinicians might apply in practice.
Why the highest-dose zone best predicts pneumonitis. The 50 to 55 Gy dose range (RS8) represents the boundary between the high-dose treatment field and surrounding tissue receiving slightly less radiation. Tissue in this zone undergoes the most dramatic changes, and the delta radiomics features capturing how this tissue evolves during treatment were the most informative for predicting subsequent lung injury.
The biological significance of NLR at week 4. The neutrophil-to-lymphocyte ratio reflects the balance between pro-inflammatory (neutrophil-driven) and adaptive immune (lymphocyte-driven) responses. An elevated NLR at week 4 of radiotherapy indicates systemic inflammation is outpacing immune regulation, which predisposes the lung to excessive radiation-triggered inflammatory injury. This study observed significantly higher NLR at week 4 in patients who developed RP (4.14 versus 3.63 in the training cohort).
Why traditional dosimetric parameters did not predict RP here. Standard dosimetric variables including V5, V20, V30, mean lung dose, and the planning target volume to lung volume ratio showed no statistically significant association with RP in this cohort. This may reflect the confounding effect of immunotherapy-enhanced lung inflammation, which can trigger pneumonitis at dose levels that would not cause injury without ICI treatment.
Study limitations. The retrospective design and single-institution setting limit generalizability, and pulmonary function data was unavailable for approximately one third of participants. The absence of an external validation dataset and the relatively small sample size increase susceptibility to selection bias. Multi-institutional prospective studies are needed to confirm these findings and standardize the radiomics workflow.
A novel tool for the immunotherapy era. This is described as the first study to use CT-based dose-segmentation radiomics specifically in the context of post-CCRT consolidative immunotherapy. Prior radiomics models for RP were developed in patients receiving radiotherapy alone or chemoradiotherapy without immunotherapy, where the risk dynamics differ substantially.
Practical implications for treatment planning. Identifying high-risk patients before or early in treatment allows radiation oncologists to adjust target volumes, apply stricter dose constraints on healthy lung tissue, or intensify monitoring. The model's high negative predictive value means low-risk patients can be safely managed with standard follow-up protocols rather than intensive surveillance.
Building on prior dosiomics research. Previous studies showed that dosiomics models using dose-volume relationships outperformed standard dosimetric models for RP prediction. This study advances that work by using fine-grained 5-Gy dose segmentation within specific lung subvolumes and by incorporating longitudinal delta radiomics, which capture how lung tissue texture evolves during treatment rather than relying on a single snapshot.
Path forward. Future studies should validate the combined model in prospective, multi-center cohorts and explore whether 3D volumetric radiomics and integration of pulmonary function data further improve prediction accuracy. Standardization of the radiomics workflow across imaging platforms and processing software will be essential to enable widespread clinical adoption.