Click here to


Are you sure ?

Yes, do it No, cancel

A Deep Learning Method for Xerostomia Prediction in Head-And-Neck Radiotherapy

K Men*, H Geng , H Zhong , Y Fan , A Lin , Y Xiao , University of Pennsylvania, Philadelphia, PA 19104, USA


(Tuesday, 7/16/2019) 7:30 AM - 9:30 AM

Room: Stars at Night Ballroom 2-3

Purpose: Xerostomia is common sequelae in head-and-neck patients after radiotherapy, which can seriously affect the patient quality of life. In this study, we proposed a xerostomia prediction model with radiation treatment data using a 3D residual convolutional neural network (3D rCNN). The model can be utilized to guide radiotherapy for toxicity reduction.

Methods: 838 squamous cell carcinoma (HNSCC) patients that enrolled to RTOG 0522 were included in the study. We modeled moderate-to-severe xerostomia defined as grade 2 or higher at any time after radiotherapy.377 patients with Grade ⩾2 xerostomia toxicity were categorized as cases with toxicity and 461 patients with Grade 0-1 toxicity as non-toxicity.The inputs of the 3D rCNN included the planning CT images, 3D dose distributions, and the contours of the parotid glands and submandibular glands. A 20-fold cross-validation and a train-validation-test experiment were applied to evaluate the performance of the prediction model. For the cross-validation, the dataset was randomly divided into 20 equal-sized subsets. For each of 20 iterations, 19 subsets were used as the training set and the remaining subset as the test set. For the train-validation-test experiment, 80% of the dataset were randomly selected as the training set, 10% as the validation set, and the remaining 10% as the test set. The evaluation metrics included the accuracy, sensitivity, specificity, and the area under the receiving operator characteristic curve (AUC).

Results: The proposed method achieves promising prediction results. For the 20-fold cross-validation, it had accuracy of 72.1%, sensitivity of 70.3%, specificity of 73.5%, and AUC of 0.71, while for the separated test set, they were 70.2%, 70.3%, 70.2%, and 0.74, respectively.

Conclusion: The proposed model could extract low- and high-level spatial features using 3D CNN filters and achieve promising performance. This potentially can be an effective method that provides objective toxicity prediction for clinical decision support.

Funding Support, Disclosures, and Conflict of Interest: This project was supported by NCI grants U24CA180803(IROC), U10CA180868(NRG), and PA CURE grants. The authors report no conflicts of interest with this study.


Dose Response, Modeling, Image Analysis


TH- response assessment : Machine learning

Contact Email