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

Xerostomia is the most common toxicity in radiation therapy (RT) of head and neck cancer (HNC), impacting patient’s quality of life. In this study, we propose a deep learning model that predicts the chance of a patient experiencing xerostomia 12 months after RT. We first created an atlas computed tomography (CT) image via group-wise registration of 25 randomly selected HNC patients and normalized all patients’ CT and RT dose to the atlas space by deformable registration. We hypothesize that normalizing patient’s anatomy and associated RT dose would allow the model to better focus on the spatial dose impact on xerostomia and minimize noises from anatomical variations. Next, a 3D residual network is trained to predict xerostomia at 12 months post-RT. High-resolution class activation maps are generated to interpret the prediction of the trained model and suggest that salivary glands have high impact on xerostomia. The xerostomia prediction model was trained and tested on 533 HNC patient data, and achieved mean±SD AUC, sensitivity, specificity, negative predictive value, precision and accuracy of 0.78±0.12, 0.69±0.12, 0.59±0.18, 0.90±0.09, 0.32±0.09, 0.62±0.11, respectively. Our results suggest that the developed model is a promising approach for predicting the occurrence of xerostomia at 12 months post-RT.