Deep-learning xerostomia prediction model with anatomy normalization and high-resolution class activation map

Xerostomia, a common toxicity induced by radiation, severely reduces patients’ quality of life. We propose a deep learning model that predicts the chance of a patient experiencing xerostomia 12 months after radiation therapy. An atlas computed tomography image is created to normalize patient anatomy to help improve model performance. High-resolution class activation maps are generated to better understand the decision of the prediction model. The interpretation of the model’s behavior suggests a correlation between xerostomia and spatial radiation dose in salivary glands.