Click here to


Are you sure ?

Yes, do it No, cancel

Fully Automated Head-And-Neck Contouring and VMAT Planning with Integrated and Comprehensive QA

R McCarroll1*, B Beadle2 , P Balter3 , H Burger4 , C Cardenas3 , S Dalvie5 , D Followill3 , K Kisling3 , M Mejia6 , K Naidoo7 , C Nelson3 , C Peterson3 , K Vorster4 , J Wetter4 , L Zhang3 , J Yang3 , L Court3 , (1) University of Maryland Medical Center, Baltimore, MD, (2) Stanford University, Stanford, CA, (3) MD Anderson Cancer Center, Houston, TX, (4) University of Cape Town, Groote Schuur Hospital, Cape Town,(5) University of the Free State, Bloemfontein, (6) University of Santo Tomas Hospital, Metro Manila, Manila, (7) Tygerberg Hospital, Bellville, Cape Town


(Saturday, 3/30/2019) 10:30 AM - 12:30 PM

Room: Osceola Ballroom C

Purpose: To develop and validate the automation of the entire head-and-neck treatment planning process.

Methods: In-house and commercial software were integrated to automatically create head-and-neck VMAT plans in <30 minutes requiring only the treatment prescription, CT scan, and GTV contour. OARs and CTVs are automatically contoured using an in-house independently verified multi-atlas technique which has been implemented clinically. For QA, machine-learning algorithms and deep-learning are used to assess OARs and CTVs for accuracy and need for review. Plans are optimized using a modified commercially-available RapidPlanᵀᴹ model. Eclipse API and dicom transfer techniques automate the process from CT import to dose calculation.

Results: On retrospective analysis, 98%(1903/1949) and 95%(381/400) of automatic OARs and CTVs were physician approved with no or minor edit. During clinical implementation, 50% of 1,415 OAR contours were not edited for clinical use and 86% had a maximum distance-to-agreement of <5mm. In identifying contours requiring edit, random-forest models detected >95% of autocontoured OARs with clinical edits exceeding a maximum distance-to-agreement of 10mm. For CTV contour QA, the mean distance-to-agreement of autocontoured CTVs compared to those from an independent deep-learning algorithm was significantly correlated to agreement with clinical CTVs (CCC=0.83,p<0.01). When autocontoured OARs were used for treatment planning only 5/898 clinical structures exceeded DVH criteria. Finally, comparing 30 autoplans to those from a clinical trial, the autoplans performed significantly better considering brainstem and spinal cord D(max), D(1cc) of the high-dose PTV, and D(mean) of contralateral submandibular and parotid glands (paired Wilcoxon rank-sum test, p<0.01). For no structure did autoplans perform significantly worse. Physicians from four institutions rated 90% of plans as clinically acceptable without edit, the remaining plans required only minor edit.

Conclusion: All planning processes have been successfully automated, with clinically acceptable plans created in less than 30 minutes. Independent validation of each step ensures safe plan delivery.

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.

Contact Email