Click here to


Are you sure ?

Yes, do it No, cancel

A Recurrent Neural Network for Xerostomia Prediction in Head and Neck Cancer From Daily CBCT Images

H.H. Tseng1*, B Rosen1 , K Brock2 , A Eisbruch1 , M Mierzwa1 , J.T. Chien3 , R Ten Haken1 , I El Naqa1 , (1) University of Michigan, Ann Arbor, Ann Arbor, MI, (2) UT MD Anderson Cancer Center, Houston, TX, (3) National Chiao Tung University, Hsinchu, Taiwan


(Sunday, 7/29/2018) 3:00 PM - 6:00 PM

Room: Exhibit Hall

Purpose: To build a classification model for xerostomia that incorporates temporal information from high dimensional data taken during head & neck (H/N) cancer radiotherapy.

Methods: After deformable image registration, 52 intensity histogram and shape features were first calculated from daily CBCT images of the parotid glands of 91 H/N patients (over a course of 6-7 weeks). Two deep-learning methods, variational autoencoder (VAE) & recurrent neural network (RNN), were combined to produce a predictive model capable of processing that temporal information. This approach combines VAE's ability to faithfully map original information into lower-dimensional latents, with RNN sequential learning. RNN models such as gated recurrent unit (GRU) are capable of carrying long-term information by mimicking human memory to make prediction. The VAE was first applied to the dataset to perform feature extraction, in replacement of manual feature selection or without introducing prior domain knowledge, and subsequently the compressed latent representation was fed into 4 GRU units for predicting xerostomia occurrence, where the outcomes were scored by 1-year xerostomia (grade 1 or higher). This approach was then evaluated using 5-fold cross-validation.

Results: The (nonlinear) dimensionality reduction of data using unsupervised VAE, via variational encoder and decoder of two neural networks with the training epochs set at 500 using 32 neurons, demonstrated efficient latent-variable inference and robustness to data noise. The RNNs were then able to predict xerostomia with an average AUC=0.79 (95% CI: 0.77, 0.81). This is an improvement over our previous best standard radiomics model AUC=0.71 (0.60-0.82), indicating benefits from incorporating time correlations via GRU.

Conclusion: The combination of VAE and RNN was demonstrated effective in the sequential learning task of xerostomia. A VAE was successfully able to retain necessary information into a lower-dimensional representation and a RNN was able to apply the reduced data to predict xerostomia from longitudinal CBCT information.


Radiation Therapy, Pattern Recognition, Treatment Planning


Not Applicable / None Entered.

Contact Email