3352 - Multi-Channel 3D Convolutional Neural Network to Predict Grade 2 Xerostomia Following Definitive Radiation for Head and Neck Squamous Cell Carcinoma

K. C. Chuang¹, O. Trejo², Z. Xia¹, and J. Lee¹; ¹Duke University Medical Center, Durham, NC, ²Colorado Associates In Medical Physics, Colorado Springs, CO

Purpose/Objective(s): Late toxicities following definitive radiation therapy for head and neck cancer profoundly impact patient quality of life. Improving prediction of xerostomia and other late effects may help guide radiation treatment planning, patient counseling, and toxicity management. Here, we integrated ten independently trained multi-channel 3D convolutional neural networks (MC-3DCNNs) to predict grade 2 or higher xerostomia at 12 months post-treatment. This ensemble approach improved robustness by leveraging a diverse and expanded dataset. Materials/

Methods: The ensemble model was constructed by averaging predictions from ten distinct MC-3DCNNs, each trained on real patient data from The Cancer Imaging Archive (n=32) and an in-house dataset (n=37). The dataset incorporated radiographic, dosimetric, and clinical features (e.g., age, stage, primary site, and chemotherapy). Preprocessing involved resampling, normalization, and augmentation (increasing the sample size to 180) to enhance the models generalization. The MC-3DCNNs were designed to process multi-channel images, including CT intensity, dose, and structures. A secondary convolutional branch incorporated clinical parameters. The ensemble model provided a consensus prediction, enhancing the stability and reliability of the individual model outputs.
Results: This model predicted grade 2 or higher xerostomia with accuracy, ROC-AUC, precision, recall, and F1-score of 0.72±0.07, 0.82±0.09, 0.73±0.14, 0.74±0.10, and 0.73±0.08 over ten-fold cross-validation. Gradient-weighted Class Activation (Grad-CAM) mapping revealed input regions near the parotid glands were critical in this predictive model.
Conclusion: The MC-3DCNN model showed promising results in predicting late xerostomia following definitive radiation therapy for head and neck cancer. This model served as a training ground for establishing advanced techniques and showed generalized and consistent performance on a dataset combining publicly available and in-house data.