Room: AAPM ePoster Library
Purpose: To investigate if a machine learning optimization (MLO) model trained by multi-criteria optimization (MCO) plans can auto plan complex pancreatic VMAT with dose painting and clinically replace MCO manual planning.
Methods: An MLO model was trained in RayStation 9B by 109 MCO-VMAT plans, delivering 50.4 Gy to pancreatic tumor and regional nodes with dose painting of 58.8 Gy to involved vessels. MLO uses a random forest algorithm to predict voxel dose and a collection of dose-volume rules (called strategy) to refine 3D dose distribution. The model was validated on ten representative training patients and tested on 18 consecutive new patients. Eight key clinical goals were compared between the MLO auto plan and the MCO manual plan: the coverage of the low-dose and high-dose CTV and PTV, and the V50 of stomach, duodenum, small and large bowels outside PTV.
Results: The MLO model had a minor deficiency: it frequently under-dosed the outermost voxels in the low-dose PTV in the direction not constrained by any adjacent organ. This problem could be easily fixed in post-processing by boosting the dose to those voxels without causing extra dose to any organ. Three MLO auto plans had all eight goals fulfilled, seven only slightly missed the low-dose PTV coverage which can be easily fixed in post-processing, three had sufficient target coverage but missed one goal on organ dose, and the rest five missed only one goal in addition to the low-dose PTV coverage.
Conclusion: With a simple post processing, 10 of the 18 MLO auto plans can offer similar quality to the MCO manual plans for this very complex dose-painted VMAT treatment in which the size, shape and location of the high-dose PTV can vary dramatically. The other eight plans also provided the user very favorable starting points to quickly find the optimal plan.
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation