Click here to


Are you sure ?

Yes, do it No, cancel

Longitudinal Analysis of Parotid Gland Anatomical Changes During Radiotherapy by Recurrent Convolutional Neural Networks

D Lee*, P Zhang, S Alam, J Jiang, S Nadeem, A Caringi, N Allgood, Y Hu, Memorial Sloan Kettering Cancer Center, New York, NY


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: radiation therapy (RT) of head and neck (HN) cancer, the shape and volume of the parotid glands (PG) may change significantly due to patient weight loss, resulting in clinically relevant deviations of delivered dose from the planning dose. Early and accurate longitudinal prediction of PG anatomical changes during RT can be valuable to inform decisions on plan adaptation. We developed a novel deep neural network for longitudinal predictions using only the displacement vector fields (DVFs) between the planning CT and weekly cone beam computed tomography (CBCT).

Methods: three HN patients treated with volumetric modulated arc therapy of 70Gy in 35 fractions were retrospectively studied. Each patient’s data contained the planning CT, planning contours and 4 weekly CBCTs. We calculated DVFs between week 1-3 CBCT and the planning CT by a demon deformable image registration algorithm. DVFs were subsequently used for input of a novel network combining convolutional neural networks and recurrent neural networks to predict the DVF of the 4th week. Finally, we reconstructed the warped PG contour using the predicted DVF. We split the patient cohort qinto the training/validation set (n=50) and testing set (n=13). For evaluation we calculated Dice coefficient (DICE) and volume change by comparing the predicted with manually-segmented PG contour data on the 4th week CBCT.

Results: average DICE between the predicted and the manual PG contours was 0.79±0.06 standard deviation (ipsilateral) and 0.81±0.05 (contralateral), respectively. The average difference in volume was 2.4±1.3 cc (Ipsilateral) and 2.5±1.2 cc (Contralateral), with a correlation of 0.88.

Conclusion: deep learning framework combining convolutional and recurrent neural networks is capable of making predictions of future PG volume and shape changes with clinically acceptable accuracy, and ready to be integrated into an adaptive radiotherapy workflow.

Funding Support, Disclosures, and Conflict of Interest: Master research agreement between Memorial Sloan Kettering Cancer Center and Varian


Cone-beam CT, Registration, Deformation


IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

Contact Email