Click here to


Are you sure ?

Yes, do it No, cancel

Effective Automated Treatment Planning and Quality Assurance for Cervical and Breast Cancer for Limited-Resource Clinics

K Kisling1*, L Zhang1, B Anderson1, D Anderson2, P Balter1, B Beadle3, H Burger2, C Cardenas1, M Du Toit4, N Fakie2, R Howell1, A Jhingran1, J Johnson, R McCarroll5, K Schmeler1, S Shaitelman1, H Simonds4, T Thebe2, C Trauernicht4 , J Yang1 , L Court1, (1) MD Anderson Cancer Center, Houston, TX, (2) University of Cape Town, Cape Town, South Africa, (3) Stanford University, Stanford, CA, (4) Stellenbosch University, Cape Town, South Africa, (5) University of Maryland Medical Center, Baltimore, MD


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

Room: Osceola Ballroom C

Purpose: To develop automation algorithms that increase the availability of high-quality treatment plans in clinics with limited staff and without access to advanced techniques, such as breath-hold gating and VMAT.

Methods: We used atlas-based segmentation, deep learning, optimization, and classification techniques to create automation algorithms that (1) create treatment plans from a patient CT and (2) perform plan QA. We automatically planned treatments using our algorithms (retrospective validation cohort). Plans were scored for clinical acceptability by radiation oncologists in the US and South Africa, and dose metrics were evaluated. To evaluate risk in our automated workflow, we performed Failure Modes and Effects Analysis (FMEA).

Results: For cervical cancer, 150 four-field box treatments were automatically planned, and 89% were scored acceptable by physicians. The dose distributions were more homogeneous (p<0.001) using automatically-optimized beam weights compared with equal beam weights (commonly used in Africa). Our automatic QA technique flagged 90% of unacceptable plans (false-positive rate: 16%). For breast cancer, we automatically planned three-field treatments on free-breathing CTs for 19 left-sided, post-mastectomy patients. Seventeen (89%) met constraints for lung and heart dose and target coverage. Physicians accepted the tangent beam angles for all plans. The inferior border required median changes of 4.0mm (max: 34mm), indicating that this should be identified manually. Otherwise, physicians accepted a majority of plans (12/19) with no or minor changes. Automatic QA of the dose flagged 6/7 unacceptable plans. Physicians indicated that 3/7 unacceptable plans needed a more complex treatment technique due to anatomy. FMEA showed that automated QA reduced risk (number of high-risk failure modes was reduced by half) and that manual plan review is still vital for safety.

Conclusion: Automated treatment planning and QA was effective for the majority of patients tested. Our algorithms will be implemented clinically at our partner hospitals in South Africa next year.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by a grant from the National Cancer Institute (UH2-CA202665). We have received equipment and technical support from Varian Medical Systems and Mobius Medical Systems.

Contact Email