Session Item

The majority of radiotherapy treatment planning systems provide automated treatment planning solutions with various underlying algorithms. The validation, clinical introduction, and maintenance of these algorithms can, however, be relatively time consuming. We aimed to efficiently implement an automated treatment planning algorithm to create high quality and consistent treatment plans for rectal cancer VMAT non-inferior to the “manually” optimized VMAT plans.

A contextual atlas random forest machine learning optimization (MLO) algorithm was used to fully automatically predict dose distributions and create corresponding treatment plans for new rectal cancer patients. A MLO model using clinical 6 and 10MV VMAT plans of 51 previously treated patients with a prescription of 5000 cGy in 25 fractions was trained at the Princess Margaret Cancer Centre (PMCC, Toronto, Canada). At Radiotherapiegroep (Arnhem/Deventer, the Netherlands), the MLO model was tuned to conform the local clinical goals (table 1) and contouring guidelines using 10 previously treated rectal cancer patients. MLO tuning was done by applying generic postprocessing objectives to the MLO predicted dose distribution and the consecutive dose mimicking step.

Prospectively, MLO plans were created with the tuned MLO model and compared to the manually optimized clinical plans. The machine configuration (e.g. single or dual arcs) of the MLO plans were copied from the clinical plans. If needed, limited postprocessing was allowed as part of the automated planning procedure. For each patient, a radiation oncologists blindly reviewed both plans.

After tuning the original PMCC MLO model, 8/10 MLO plans were clinically acceptable following the local clinical goals. The tuned MLO model showed a higher degree of modulation with, on average (range), 466 (360 – 645) and 651 (563 – 834) monitor units for the manual and MLO plans, respectively (p<0.001). 

Next, MLO plans for 19 consecutive patients were prospectively created. The radiation oncologists classified 15/19 MLO plans non-inferior to the clinical plans. Postprocessing was only needed in 4/19 plans (on average within <10 min), consisting of a subtle reduction of the maximum bladder dose. Two MLO plans did not meet predefined goals due to relatively high bowel bag overlap with the PTV, but were considered clinically acceptable by the radiation oncologist. The clinical goals as well as the bladder and bowel bag mean dose between the clinical plans and MLO plans were not statistically significant different (figure 1). All plans fulfilled the clinical goals of the femoral heads and the bladder V5150<0.1cc. 

For rectal cancer VMAT, a machine learning automated planning model can be shared across institutes while conforming to the local clinical goals without laborious retraining, but by applying a generic postprocessing strategy. The resulting MLO plans for rectal cancer VMAT were non-inferior to the manual plans for the majority of patients (79%).