Room: Davidson Ballroom A
Purpose: To fully automate head-and-neck contouring, VMAT planning, and independent plan checks in under 45 minutes.
Methods: VMAT plans are automatically created and checked using a combination of in-house algorithms and Eclipse functions. Only a CT scan, patient prescription, and GTV contour(s) are required. The prescription and CT are first checked for consistency, then normal structures and intermediate- and low-risk CTVs are automatically contoured using an in-house multi-atlas algorithm. Normal tissue contours are then checked using machine learning (random forest). Finally, knowledge-based VMAT optimization and secondary 3D-dose calculation are completed.
Results: For contouring validation, physicians rated 95%(381/400) of automatically delineated intermediate- and low-risk target volumes and 98%(1903/1949) of normal tissue contours as acceptable with no or minor edit. Upon clinical implementation (200+ patients over 18 months), normal tissue contours were edited resulting in a mean-surface-distance of only 1.8Â±0.8mm; 49% of contours were used for planning without edit. Potentially significant autocontouring errors (Hausdorff Distance>1cm) were detected using random forest models with sensitivity and specificity >95%. Planning on structures requiring small edits (Hausdorff Distance<1cm) resulted in an average change to clinically relevant DVH metrics of only 2.4%. The quality of automated VMAT planning was evaluated using 20 patients originally treated on a clinical trial protocol. Automated plans performed significantly better (Wilcoxon-paired rank-sum-test, p<0.01) considering Dmax to brainstem and spinal cord, D1cc of the high-dose PTV, and Dmean to the contralateral parotid and submandibular glands. Physicians from four institutions rated 90%(35/39) of plans as clinically acceptable without edit, the remaining 4 plans required minor edit.
Conclusion: Fully automatic head-and-neck contouring, VMAT planning, and independent plan checks in under 45 minutes has been achieved. Requiring only the CT, GTV contour(s), and patient prescription, the process requires no user intervention. Automatic contouring is robust for treatment planning, and the resulting plans outperform clinically delivered plans.
Funding Support, Disclosures, and Conflict of Interest: This work was funded by the NCI (UH2 CA202665). Additional support provided by Varian Medical Systems and Mobius Medical Systems.