(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 3
Purpose: Knowledge-based methods have been shown to standardize, automate and improve the quality of treatment planning for external-beam radiotherapy, but are only now being explored for gynecologic brachytherapy. The purpose of this work was to develop a knowledge-based voxel-wise dose prediction system using a convolution neural network (CNN) for high-dose-rate brachytherapy cervical cancer treatments.
Methods: A 3D U-NET CNN was utilized to output voxel-wise dose predictions based on organ-at-risk (OAR), high-risk clinical target volume (HRCTV), and possible source location geometry. The available dataset was a five-year sample of previously-treated tandem-and-ovoid treatments comprising 364 cases (310 training:54 validation). HRCTV, OARs (bladder/rectum/sigmoid), and source train locations were extracted along with corresponding dose for each voxel. Structures and dose were interpolated to 1mmx1mm dose planes with 2 channels: one for dose emitting structures (possible source positions) and the other for OAR voxels. We examined the discrete DVH metrics utilized for tandem-and-ovoid plan quality assessment: HRCTV D90%(dose to hottest 90% volume), HRCTV V100%(%volume at prescription dose), and bladder/rectum/sigmoid D2cc, with ?Dx=Dx,actual-Dx,predicted mean and standard deviation quantified model performance. To assess 3D dose prediction accuracy, we evaluated the dose difference [Dactual(x,y,z)-Dpredicted(x,y,z)] in both training and validation sets across the clinically-relevant dose range for cervical cancer brachytherapy (50-120% of prescription), mean and standard deviation.
Results: Relative DVH metric prediction in the training (validation) set were HRCTV ?D90=0.11±0.48Gy(?D90=-0.05±0.75Gy), HRCTV ?V100=2%±5%(?V100=1%±8%), bladder ?D2cc=0.05±0.42Gy (?D2cc=-0.23±0.80Gy), rectum ?D2cc=-0.01±0.34Gy (?D2cc=-0.06±0.52Gy), and sigmoid ?D2cc=0.09±0.29Gy (?D2cc=-0.03±0.51Gy). Voxel-wise accuracy for 50-120% dose range inside contours volumes for training (validation) was bladder:1.3%±6.1%(-0.3%±10.7%), rectum:1.1%±6.0%(-0.1%±9.2%), sigmoid:2.2%±5.6%(0.8%±8.1%), HRCTV:1.1%±10.0%(-1.6%±17.7%).
Conclusion: 3D knowledge-based dose predictions for brachytherapy provide accurate voxel-level and DVH metric estimates that could be used for treatment plan quality control and eventually fully-automated plan generation. Reductions in modeling error are likely through quality filtering and/or re-planning of the raw clinical sampling of plans.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by Padres Pedal the Cause.
Keywords
Brachytherapy, Treatment Planning
Taxonomy
TH- Brachytherapy: Treatment planning using machine learning/automation
Contact Email
