Click here to


Are you sure ?

Yes, do it No, cancel

Data-Driven Dose Volume Histogram Prediction

M Polizzi1*, R Watkins2, W Watkins3, (1) Virginia Commonwealth University, Richmond, VA, (2) Virginia Polytechnic Institute and State University, Department of Computational Modeling and Data Analytics, (3) University of Virginia, Charlottesville, VA


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

Room: AAPM ePoster Library

Purpose: To introduce methods to query and interpret prior patient data; and demonstrate the utility of prior data in predicting prospective clinical treatment dosimetry.

Methods: An Oncospace database was constructed utilizing clinical data for advanced lung cancer (N=122 patients). The database can be queried and returns DVH data for the most similar organs in the database. We introduce a 2-point overlap volume histogram (OVH) query at distances d=0,2cm. The OVH(0cm) is OAR/PTV overlap; OVH(2cm) is relative volume 2cm away from PTV. We examine OVH-variation as a function of number of similar organs returned from the query; and we examine the ability to predict DVH utilizing this method. The OVH(cm) and DVH(%) queried from the database are reported as the inter-quartile-range (IQR) and non-outlier range (NOR)=3*IQR and average error (err).

Results: OVH variations increase as a function of increasing patients and increasing distance. In esophagus and heart, IQR<=20% for distances ranging from (-2cm,10cm) when querying up to 20 similar organs; NOR>40% at d=2cm. With a 20-patient query, the IQR, NOR were <4%,16% in total lung and <1%,5% for external at all distances <2cm from the PTV. The ability to predict clinical delivered DVH based on a query of the 20-most-similar organs was excellent in lungs and external contour but varied in heart and esophagus for many patients. The err was <10% for 11/23 esophagus, 13/23 hearts; but for external the maximum difference was 7% and for 21/23 err<4.6%. Lungs were also predicted accurately with <3% err 14/22 cases.

Conclusions: Utilizing prior data to prospectively predict DVH dose is of increased interest, but model- and data-driven methods have severe limitations. We show prediction is reasonable in organs containing tumor (with known overlap), but for non-overlapped structures planning preference and plan design may be the dominant factors in determining dose.


Not Applicable / None Entered.


TH- Dataset Analysis/Biomathematics: Informatics

Contact Email