11. Performance evaluation of an atlas-based machine learning approach for automated prostate radiotherapy planning

150 150 Techna Symposium

Leigh Conroy 1, Chris McIntosh 1,2, Alejandro Berlin 1,2, Michael C. Tjong 1, Peter Chung 1, Tim Craig 1, and Thomas G. Purdie 1,2

1 Radiation Medicine Program, Princess Margaret Cancer Centre, Toronto, ON, Canada
Techna, University Health Network, Toronto, ON, Canada

To evaluate the behavior and dosimetric performance of an in-house developed atlas-based automated planning approach applied to prostate treatment planning.

The automated planning pipeline was trained on 94 sets consisting of CT scans, structure sets, and dose distributions. Image features were extracted and used to train two atlas-selection models: one to determine atlases for spatial dose prediction using an equal weighted combination of the Gamma metric and a dose volume histogram (DVH) difference, and one to predict atlases with the closest dose distribution (‘dose prior’) using mean DVH difference. For each independent testing set (n=20) the 4 closest spatial atlases were used to create a probabilistic dose distribution and the 4 closest dose prior atlases were subsequently used to determine the most probable dose distribution. Finally, dose-mimicking was used to create clinically-deliverable plans in a commercial treatment planning system. Automated plans were evaluated using our institutional clinical criteria. To investigate the underlying performance of the atlas-based technique, we determined the frequency of atlas selection and investigated atlas distance metrics for predicting automated plan quality.

Automated plans were successfully generated for all 20 testing sets, 3 were discarded due to hip implants. 16/17 automated plans met all clinical DVH goals. One plan failed to meet dosimetric criteria (D30%) for rectum wall. The distance metrics were not predictive of this failure and may not be suitable for inter-patient plan quality prediction. In this cohort, most patient plans were well-predicted by a small number of representative atlases: of 94 training atlases, 21 were used for spatial dose prediction and 13 were used for dose prior construction.

Atlas-based automated planning is capable of producing clinically-deliverable prostate VMAT plans that exceed clinical goals for the majority of patients. Future work will further investigate metrics to predict the quality of automated plans.