Click here to


Are you sure ?

Yes, do it No, cancel

Analysis of a Machine Learning Based Planning Tool for Parotid Dose Prediction and Sparing

N Shaheen*, R Bayliss, P Hill, University of Wisconsin, Madison, WI


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

Room: AAPM ePoster Library

Purpose: One use of machine learning to assist treatment planning incorporates the expertise and historical tendencies of a clinic to predict and create high-quality, deliverable plans. This work aims to validate an internally developed, machine learning based dose prediction tool for use in parotid sparing and optimization improvement in head and neck treatment planning.
Methods: Treatment planning data was collected from 289 head and neck patients treated with TomoTherapy. Dosimetrically-based groupings were determined for each parotid by applying principal component analysis (PCA) and k-means clustering. Anatomical features were used to predict one of three dose groups (low, medium, high) using a prescription and clinic independent method. Of the 575 parotids in the dataset, 152 were used as a validation cohort to determine the accuracy of the trained model. Dose-volume histogram (DVH) goals for the dose groups were extracted from the average of each cluster, and correctly classified plans were reoptimized to verify that further dose sparing could not be achieved without compromising target coverage. The resulting DVHs were transformed with the same PCA and k-means to compare against the initial model.
Results: The highest predictive accuracy for the laterality agnostic parotid group was 92%, while the validation set performed at 70%. Reoptimizing within the predicted group goals results in maintained target coverage. Of the 18 parotid volumes that were replanned, 17 fell within the same dose group, while dose to one parotid was reclassified from high dose to medium dose following reoptimization.
Conclusion: The plans that were correctly classified and reoptimized using the DVH goals predicted by the trained model saw little improvement in parotid sparing and maintained target coverage. Plans that were reoptimized using too restrictive of DVH goals as determined by the model showed a loss of target coverage becoming clinically unacceptable.

Funding Support, Disclosures, and Conflict of Interest: This project was supported by the Specialized Program of Research Excellence (SPORE) program, through the NIH National Institute for Dental and Craniofacial Research (NIDCR) and National Cancer Institute, grant P50DE026787. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.


Treatment Planning, Optimization, Quality Assurance


TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email