D Rhee¹*, A Jhingran¹, B Rigaud¹, K Huang¹, K Kisling², B Beadle³, C Cardenas¹, S Kry¹, S Vedam¹, L Zhang¹, K Brock¹, W Shaw⁴, D O'Reilly⁴, H Simonds⁵, L Court¹, (1) University of Texas MD Anderson Cancer Center, Houston, TX, (2) UC San Diego, La Jolla, CA, (3) Stanford University, Stanford, CA, (4) University of the Free State, Bloemfontein, ,ZA, (5) Stellenbosch University, Stellenbosch, ,ZA,

• 
  D Rhee

(Sunday, 7/12/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose:
To develop an automatic contouring and planning solution for cervical cancer radiotherapy that does not require user intervention or edits.

Methods:
We developed a two-step convolutional neural network (CNN)-based model that automatically contours 11 normal structures and CTVs (primary and nodal) on CT images of patients with cervical cancer. High-quality clinical contours were curated from 2,254 CT scans using a semi-automated approach for the normal structures. Four physicians contoured the CTVs (250 CTs) based on the GEC-ESTRO EMBRACE-II guidelines with peer-review to ensure consistency. The Inception-ResNet-V2 architecture was used to classify the existence of the normal structures and the CTVs in each CT slice. Then, a combination of 2D (FCN-8s) and 3D (V-Net) architectures was used to segment the structures from the CT slices that had been classified to contain these structures. Using overlap metrics, we evaluated the accuracy of the autocontouring model on an independent set of 140 CT scans.
PTVs were generated from the automatic CTVs and projected into the beam’s-eye-view. Beam weight optimization was automatically performed to achieve a homogenous PTV dose distribution. An experienced radiation oncologist scored (no-edits/minor-edits/major-edits) the automatically generated plans on 10 patients from a second institution.

Results:
The mean Dice±1s between manual and automatic contours were 0.84±0.08 for primary CTV, 0.82±0.03 for nodal CTV, 0.88±0.09 for bladder, 0.80±0.09 for rectum, 0.94±0.03 for femurs, 0.90±0.02 for spinal cord, 0.93±0.011 for kidneys, 0.91±0.02 for pelvic bone, 0.91±0.01 for sacrum, 0.93±0.03 for L4 vertebrae, and 0.92±0.02 for L5 vertebrae. 10/10 automatically generated plans were scored as no-edits. End-to-end plan generation time was 10.1±0.7 minutes with a single GPU.

Conclusion:
We developed a streamlined data-driven methodology to curate high-quality clinical contours. This approach provided the necessary data to create a comprehensive and reliable auto-treatment planning solution for cervical cancer.

Funding Support, Disclosures, and Conflict of Interest: This work was partially funded by NCI (UH3CA202665) and Varian Medical Systems.

Keywords

Treatment Planning, Segmentation, Data Acquisition

Taxonomy

TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email