Prediction of Radiation Pneumonitis Using Deep Learning Applied to Dose–Function Metrics
Poster
68thAAPM_Poster_YuikoN_fin.pdf Hosted by AAPM · 0.4 MBAbstract
Purpose
Predicting radiation pneumonitis (RP) using machine learning is promising, particularly when functional lung heterogeneity is incorporated via dose-function histogram (DFH) and CT ventilation imaging (CTVI). However, traditional dose–function metrics often fail to capture complex, high-dimensional spatial relationships between radiation dose and regional lung function. This study aims to improve RP prediction accuracy by developing a deep learning model utilizing multichannel 3D inputs, integrating both CTVI and 3D dose distributions.
Methods
We retrospectively analyzed 94 patients with non-small cell lung cancer (Stages IB–IIIC) who underwent definitive radiotherapy; 23 patients (24.5%) developed grade ≥2 RP. Dual-channel 3D input volumes were constructed by mapping CTVI (functional channel) and dose distribution (dosimetric channel) to the separate channels of a 3D ResNet-18 architecture. Single-channel models using only dose distributions were evaluated for comparison. Transfer learning was implemented using two fine-tuning strategies: updating only the fully connected (FC) layer and updating Layer 4 through the FC layer. Robustness was ensured using stratified 5-fold cross-validation, randomly repeated 50 times. Performance was evaluated using the mean area under the receiver operating characteristic curve (AUC) and DeLong test.
Results
The dual-channel model (fine-tuned FC layer) achieved the highest performance, with a mean AUC of 0.997 (95% CI: 0.982–1.00). The alternative dual-channel fine-tuning strategy yielded a comparable mean AUC of 0.996 (95% CI: 0.990–1.00). In contrast, the single-channel models achieved significantly worse results, with a mean AUC of 0.963 (95% CI: 0.941–0.984), regardless of the fine-tuning strategy. The superiority of the dual channel approach was statistically significant (p < 0.001).
Conclusion
Integrating regional lung function and 3D dose distributions into a multichannel deep learning framework yielded exceptional predictive accuracy for RP. Therefore, this approach significantly enhances RP prediction, potentially contributing to personalized radiotherapy.