Session Item

Purpose

To describe steps required to introduce the machine learning treatment planning process into clinical practice.



Methods and Material

Machine learning treatment planning process was introduced in RayStation version 10A. Prostate cancer treatment plan generation model was created by the vendor based on Princess Margaret Hospital developed model incorporating the local hospital clinical goals and DVH parameters. Six local treatment plans were selected for the model fine-tuning. Ten patients were selected for model validation. After validation two plans (manual and autoplanned) were prospectively created for each prostate cancer patient. The plans were presented to two radiation oncologists who scored the plans and provided a feedback without prior knowledge of which plan is which.



Results

Initially, only about 25% of the plans selected were autoplanned due to worse PTV coverage and higher dose to the bladder. Moreover, review of 3D dose cube and clinical review has helped to identify the weaknesses and to improve the model further. After four model iterations 60-65% selected plans were autoplanned. Further 15% of plans have had a very minor dose distribution differences.



Conclusion

Current strategy is to do the autoplan and if the plan meats clinical goals then no manual plan is created. Machine learning treatment planning has improved the overall plan quality, saved time and reduced interplanner plan quality variability. It is important to incorporate into the model not only the DVH parameters, but also clinical feedback and 3D isodose distributions to improve the model during commissioning process.