Author: Junghoon Lee, Todd R. McNutt, Harry Quon, Bohua Wan 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Johns Hopkins University, Department of Computer Science, Johns Hopkins University 🌍
Purpose: Xerostomia is a common toxicity in head and neck cancer (HNC) radiotherapy (RT). A few deep learning (DL) models have been proposed to predict the chance of xerostomia 12 months after RT with limited explainability. We propose to use input channel dropout to locate important regions in individual input channel and provide decoupled interpretation of the model’s prediction.
Methods: CT, dose, and contours of the salivary glands of 533 HNC patients treated with intensity-modulated RT from 2009 to 2018 were collected, and used to train a 3D residual DL model with either CT or RT dose being randomly dropped out. The model is thus capable of predicting xerostomia given gland contours, CT and/or dose. We further apply CAMERAS (Enhanced Resolution And Sanity preserving Class Activation Mapping for image saliency) to compute high resolution CAM and visualize important regions in CT or dose.
Results: The trained model achieved area under the curve (AUC) scores of 0.75±0.14, 0.77±0.13, and 0.48±0.11 (average±SD) with all inputs, CT dropped out, and dose dropped out, respectively. CAM highlights similar regions when dose is present in input with/without CT dropout. Oral cavity, high dose region and parotid glands are highlighted when dose is present in a true positive case. When the patient receives unilateral radiation dose, the contralateral parotid is highlighted for true negative cases. It is also observed that parotid and submandibular glands are highlighted when dose is dropped out.
Conclusion: We propose to use input channel dropout to train a DL model such that it is able to predict xerostomia 12 months after RT with gland contours, CT and/or dose. We found that CT alone is insufficient to predict xerostomia. High-resolution CAM is generated to highlight important regions in CT and dose separately allowing for fine-grained interpretation of the model’s prediction.