3592 - Quantitative Radiomic and Deep Features to SBRT Toxicity Outcome: Dual-Channel Deep Neural Network-Driven Pneumonitis Prediction Early-Stage NSCLC Patients

Purpose/Objective(s): Pneumonitis is a clinically significant and potentially dose-limiting toxicity in patients with non-small cell lung cancer (NSCLC) undergoing Stereotactic Body Radiation Therapy (SBRT). This study presents the development of a dual-channel Deep Neural Network (DNN) integrating deep imaging features and radiomic features the prediction of SBRT-related pneumonitis.

Materials/Methods: This study included 216 patients with early-stage (I–II) NSCLC treated with SBRT at our institution. The median age was 74 years (range: 49–91 years), and 54% of patients were male. Among the study cohort, 24 patients (11.1%) developed treatment-related pneumonitis. All patients received SBRT with a total prescribed dose ranging from 48 to 60 Gy delivered in 3 to 5 fractions. The pneumonitis prediction model is developed based on dual-channel DNN architecture while integrating deep imaging features and radiomic features, extracted from planning CT images while considering gross tumor volume (GTV) contour. The deep imaging features were extracted using convolution neural network (CNN) followed by dense neural networks. Radiomic features from planning CT images were gleaned from GTV zone and evaluated based on the histogram, gray level co-occurrence matrix, gray level run length matrix, and gray level size zone matrix. Model’s effectiveness was assessed using five-fold cross validation approach, in which 75% of data used for training and remaining 25% for testing in each fold.

Results: The proposed dual-channel DNN model demonstrated superior predictive performance, achieving a mean AUC of 0.81 (±0.05), sensitivity of 0.88 (±0.18), specificity of 0.75 (±0.13), and a weighted average F1-score of 0.80 (±0.08). In comparison, the DNN model utilizing only deep imaging features achieved a mean AUC of 0.73 (±0.07), while the model based solely on radiomic features yielded a mean AUC of 0.74 (±0.05). Overall, the dual-channel framework demonstrated an approximate 9.5–11% improvement in AUC compared to models constructed using individual feature groups, highlighting the added predictive value of integrating complementary feature representations (see in Table below).

Conclusion: This study highlights the value of integrating CT image–derived deep features with tumor structure–based radiomic features for DNN-based prediction of SBRT-related pneumonitis. The proposed dual-channel framework demonstrates the potential benefit of multimodal feature fusion in improving predictive performance. However, prior to clinical implementation in routine practice, validation using larger, multi-institutional retrospective cohorts is necessary to ensure robustness, generalizability, and clinical reliability.