Multi-modal data integration of dosiomics, radiomics, deep features, and clinical data for radiation-induced lung damage prediction in breast cancer patients

Objective

Radiation-induced lung damage (RILD) is a critical complication in breast cancer patients undergoing radiotherapy. This study proposes a multi-modal predictive framework integrating dosiomics, radiomics, deep learning-based features, and clinical data to enhance early detection and risk stratification of Grade ≥2 RILD, ultimately supporting personalized radiotherapy planning.

Materials and methods

A dataset of 450 breast cancer patients receiving radiotherapy was analyzed, incorporating high-resolution CT scans, 3D spatial dose distributions, and comprehensive clinical parameters such as age, BMI, tumor laterality, chemotherapy regimens, and comorbidities. Imaging data were standardized through voxel resampling and intensity normalization, and features were extracted from both radiomics (215 features) and dosiomics. Mutual Information (MI)-based feature selection was applied to enhance model performance, while a 3D autoencoder with attention mechanisms was utilized to capture spatial and structural patterns linked to RILD. Five-fold cross-validation was performed to ensure robustness.

Results

The Intraclass Correlation Coefficient (ICC) analysis identified the most reproducible radiomics features, leading to significant feature reduction while maintaining predictive stability. Multi-modal data integration significantly improved classification performance, with the Voting Classifier achieving 95.89% accuracy and 96.98% sensitivity when using MI-based feature selection. Deep features demonstrated superior predictive power compared to standalone dosimetric data. The 3D autoencoder model with attention mechanisms further enhanced predictive accuracy, achieving 95% accuracy, 0.96 AUC, and 0.93 sensitivity.

Conclusion

The proposed multi-modal AI-driven approach effectively predicts Grade ≥2 RILD, addressing limitations of traditional dose-volume metrics. The integration of radiomics, dosiomics, deep learning, and clinical data enhances model accuracy and interpretability, paving the way for personalized risk assessment and optimized radiotherapy planning. Future research should focus on external validation and real-time clinical implementation to further refine predictive capabilities.