A User-Specific Automatic Contouring, Planning, and QA Solution for Cervical Cancer Radiotherapy
D Rhee1*, A Jhingran1, B Rigaud1, K Huang1, C Anakwenze1, K Kisling2, B Beadle3, C Cardenas4, S Kry1, S Prajapati1, L Zhang1, K Brock1, W Shaw5, H Simonds6, L Court1, (1) MD Anderson Cancer Center, Houston, TX, (2) UC San Diego, La Jolla, CA, (3) Stanford University, Stanford, CA, (4) The University of Alabama at Birmingham, Birmingham, AL, (5) University of the Free State, Bloemfontein, ZA, (6) Stellenbosch University, Stellenbosch, Cape Town, ZA
Presentations
TU-H-BRA-2 (Tuesday, 7/12/2022) 2:45 PM - 3:45 PM [Eastern Time (GMT-4)]
Ballroom A
Purpose: To develop and test an automatic contouring and planning solution for cervical cancer radiotherapy with three different planning techniques so that users can create plans based on their resources and practice standards. An automatic contour QA process was also developed for risk reduction.
Methods: We developed two (primary/secondary) convolutional-neural-network-based autocontouring systems that automatically generate 3 CTVs and 11 normal structures contours for cervical cancer radiotherapy. The surface-DSC (Dice-Similarity-Coefficient) between the primary and the secondary autocontours was calculated to flag incorrectly generated contours. Contours that passed QA were used to automatically create one of the following plans:(1) 4-field-box plan based on projected bony structures, (2) 3D-CRT based on projected PTV, and (3) VMAT based on soft tissue structures. Furthermore, an automated field-in-field algorithm was developed and used to reduce hotspots in 4-field-box and 3D-CRT plans. For VMAT, a commercial knowledge-based planning tool was used for automatic plan generation. We generated the plans on 35 retrospective patients from 3 hospitals for each treatment technique. Each plan was automatically fine-tuned by the field-in-field algorithm and/or plan normalization based on each physician’s practice standard. Five experienced radiation oncologists from 3 countries evaluated these plans.
Results: Overall, 93% of the primary autocontours were clinically acceptable, and the mean volumetric-DSC was 0.80 for the CTVs and 0.91 for the normal structures. The average accuracy/sensitivity/specificity of the contour QA approach was 91%/80%/97%, respectively. Finally, 97%/100%/94% of the automatically generated 4-field-box/3D-CRT/VMAT plans were clinically acceptable, respectively. The median end-to-end plan generation time was 8.73/17.9/52.9 minutes for 4-field-box/3D-CRT/VMAT plans, respectively.
Conclusion: We developed an automatic contouring, planning, and QA solution that works robustly in various clinical settings. Moreover, each plan could be optimized based on the physician’s practice standard. This solution can respond properly to resource-stratified guidelines and is helpful for all hospitals, especially those with limited resources.
Funding Support, Disclosures, and Conflict of Interest: This work was partially funded by Varian and the Nuclear Technologies in Medicine and the Biosciences Initiative, a national technology platform that is developed and managed by the South African Nuclear Energy Corporation and funded by the Department of Science and Technology through the Technology Innovation Agency.
Keywords
Segmentation, Quality Assurance, Treatment Planning
Taxonomy
Contact Email