3596 - Prediction of Xerostomia after Definitive Radiation for Head and Neck Squamous Cell Carcinoma via Interpretable Deep Learning Model

Purpose/Objective(s): Late toxicities following definitive radiation therapy for head and neck cancer profoundly impact patients' quality of life. An interpretable multi-channel 3D convolutional neural network was trained to predict grade 2 or higher xerostomia at 12 months post-treatment. The model not only improved the prediction of xerostomia but also provided interpretable guidance for radiation treatment planning and toxicity management.

Materials/Methods: The model was trained on real patient data from The Cancer Imaging Archive (n=52) and an in-house dataset (n=83) of patients who received definitive radiation for head and neck cancer. The dataset incorporated radiographic, dosimetric, and clinical features (e.g., age, stage, primary site, and chemotherapy). The model was designed to process multi-channel images, including CT intensity, dose, and structures, and a secondary convolutional branch incorporated clinical parameters. Gradient-weighted Class Activation Mapping (Grad-CAM) was created to reveal the quantitative values associated with the model's decision based on dose distribution and salivary glands.

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, respectively, over ten-fold cross-validation. Grad-CAM mapping revealed that input regions at both parotid glands were critical in the predictive model (values = 36.38 ± 16.10 vs. 23.71 for patients with and without xerostomia, p < 0.001). The maximum, mean, and median doses to both parotids were also statistically significantly different between patients with and without xerostomia (Dmax = 68.65 ± 8.10 vs. 62.01 ± 13.19 Gy, Dmean = 31.85 ± 9.89 vs. 26.44 ± 9.48 Gy, Dmedian = 26.46 ± 12.99 vs. 20.98 ± 10.08 Gy, p-values = 0.0019, 0.0027, 0.0094, respectively).

Conclusion: The proposed model showed promising results in predicting late xerostomia following definitive radiation therapy for head and neck cancer. This model also provided insights into the decision-making process and guidance to improve treatment planning strategies and toxicity mitigation.