Click here to


Are you sure ?

Yes, do it No, cancel

High Dose Rate (HDR) Vaginal Cylinder Treatment Planning by Using Machine Learning Algorithm

N Demez1*, (1) University of Kansas Hospital, Overland Park, KS


(Sunday, 7/14/2019)  

Room: ePoster Forums

Purpose: To incorporate machine learning concepts into endometrial cancer treatment with high dose rate (HDR) brachytherapy. The machine learning algorithm can learn from previous HDR treatment data and can estimate dwell times within acceptable accuracy

Methods: 50 endometrial cancer treatment plans which were planned on Oncentra™ TPS and treated with vaginal cuff cylinder were selected for this study. The dwell times, diameter of cylinder and fraction dose along with HDR Ir-192 source’s air kerma strength (AKS) were extracted from those patients’ DICOM files and exported into excel file to store as a training data. The training of patient data was implemented by using sklearn linear model multivariate LogisticRegression algorithm with python programming language. The predicted dwell times for different cylinder diameter and different cylinder treatment size were extracted from the algorithm and implemented into new or previously treated plans which were not part of the training datasets. The predicted surface data points on vaginal cuff cylinder which were used to review the HDR plan’s quality were compared with TPS plans in order to validate the predicted plans. The DVH comparison were also performed for Organs at risk (OAR) volumes.

Results: Comparison between TPS plans and predicted plans shows good agreement in surface data points. Although the predicted surface dose points on the cylinder top section shows higher dose than TPS dose points, this is not making big impact on OARs’ DVH due to sharp dose fall of HDR source.

Conclusion: This work indicates that there is a significant time saving with machine learning techniques in treatment planning and execution. It is also possible to treat safely without using TPS. More plan data are needed in order to train existing model better.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email