Program Information
Fully-Automated Treatment Planning for Cervical Cancer Radiotherapy
K Kisling1*, L Zhang1 , A Jhingran1 , J Yang1 , H Simonds2 , R McCarroll1 , M du Toit2 , P Balter1 , R Howell1 , K Schmeler1 , O Bogler1 , B Beadle3 , L Court1, (1) MD Anderson Cancer Center, Houston, Texas, (2) Stellenbosch University, Stellenbosch, Western Cape, (3) Stanford University, Palo Alto, California
Presentations
TH-EF-FS1-7 (Thursday, August 3, 2017) 1:00 PM - 3:00 PM Room: Four Seasons 1
Purpose: To develop and clinically deploy automated planning for four-field box cervical cancer treatments.
Methods: We developed fully-automated planning for cervical cancer treatment in low-resource settings. This system creates a patient-specific treatment plan ready for physician review with only a CT and no human input. In-house algorithms were developed that automate every step, including isocenter localization, treatment field creation, and beam weight optimization, and were integrated into Eclipse using the Application Programming Interface. The algorithms that create treatment fields use multi-atlas auto-segmentation of bony anatomy and then set the beam apertures based on landmarks identified on the structure projections in each beam’s-eye-view. Automatic beam weight optimization uses least-squares fitting to minimize dose heterogeneity. All algorithms were developed with input from radiation oncologists from the US and South Africa. We tested the fully-automated planner retrospectively on 228 patients. A radiation oncologist evaluated the fields for clinical acceptability, and the dose distribution was compared for optimized and equal beam weights. Additionally, we deployed automated four-field box treatment planning into our clinic. To adhere to our normal workflow, physicians review and edit the automatically-generated fields prior to treatment. The extent of these edits was quantified.
Results: Fully-automated planning took 10.2 minutes (range=7.9-14.2minutes). Of the 228 auto-plans, 89.7% of resulting fields were rated acceptable for treatment without edits. Beam-weight optimization decreased the frequency of high maximum doses (≥107%) from 28% to 4% of auto-plans. To date, the automated planning tool has been used prospectively for 14 patients in our clinic. Compared to automatically-generated fields, the edited fields had a mean-distance-of-agreement of 3.7mm and 95%-Hausdorff-distance of 14.0mm.
Conclusion: Our findings suggest that automated treatment planning for cervical cancer radiotherapy is effective and may provide a reliable alternative for low-resource settings. Prospective studies are ongoing in the US and planned with partner clinics in South Africa.
Funding Support, Disclosures, and Conflict of Interest: This work is funded by the NIH (UH2CA202665). Additional support from Varian Medical Systems and Mobius Medical Systems.
Contact Email: