Multichannel deep learning network for predicting survival in stage I NSCLC patients treated with SBRT.

e20108 Background: This study aims to develop a deep learning network (DLN)-based model to predict overall survival (OS) by integrating radiomic, dosiomic, and clinical features. The objective is to identify the most influential predictors of OS in patients with non-small cell lung cancer (NSCLC) treated with stereotactic body radiotherapy (SBRT) within a DLN-based analytical framework. Methods: Radiomic features, dosimetric parameters, and clinical data were collected from 171 NSCLC patients treated with SBRT. Twenty-two dosiomic and clinical features were obtained for this cohort. In addition, 47 radiomic features were extracted from radiation planning CT images within the planning target volume (PTV), encompassing histogram-based, geometric, and texture features derived from gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level size-zone matrix (GLSZM). Univariate analysis identified 19 significant features for model development: 3 clinical (age, age-adjusted Charlson Comorbidity Index, T-stage), 4 dosiomic (gross tumor volume (GTV) size, PTV size, conformity index, mean lung dose), and 12 radiomic features. Radiomic predictors included GLSZM (gray-level non-uniformity, zone entropy, zone percentage, size-zone non-uniformity), GLRLM (run percentage, run-length non-uniformity, run entropy, low gray-level run emphasis), and GLCM (entropy, homogeneity, energy, angular second moment). OS prediction was performed using a DLN with three fully connected hidden layers (256, 128, and 64 neurons) and a dropout rate of 0.5 to reduce overfitting. Results: Across ten randomized training (70%) and testing (30%) splits, the DLN model using 12 radiomic features achieved an ROC-AUC of 0.66 ± 0.05, with a sensitivity of 0.70 ± 0.18 and a specificity of 0.62 ± 0.16 (see Table 1). Expanding the model to include dosiomic and clinical features (19 features total) substantially enhanced predictive performance, increasing sensitivity to 0.73 ± 0.13, specificity to 0.72 ± 0.16, and ROC-AUC to 0.72 ± 0.03. Conclusions: Integrating dosiomic and clinical information with radiomic features significantly strengthens deep learning–based prediction of overall survival in patients with NSCLC treated with SBRT, yielding an approximately 9% improvement in ROC-AUC over radiomics-only models. These findings underscore the clinical potential of multimodal feature integration, with ongoing multi-institutional studies underway to confirm robustness and generalizability. Performance of the DLN-based OS prediction model. Methods ROC-AUC Sensitivity Specificity F1-score Precision 12 significant radiomic features only 0.66 (±0.05) 0.70 (±0.18) 0.62 (±0.16) 0.74 (±0.12) 0.81 (±0.07) Combing radiomic (12), dosiomic (4) & clinical (3) features 0.72 ( ± 0.03) 0.73 ( ± 0.13) 0.72 ( ± 0.16) 0.78 ( ± 0.07) 0.85 ( ± 0.08)