MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automation of Palliative Radiotherapy Treatment Planning Using Independent Models to Prevent Errors From Propagating Through the Planning Process

T Netherton1*, D Rhee2, C Cardenas3, C Chung4, A Klopp5, L Colbert6, C Nguyen7, V Kolluru8, R Douglas9, C Peterson10, R Howell11, P Balter12, L Court13, (1) MD Anderson Cancer Center, Houston, TX, (2) MD Anderson Cancer Center, Houston, TX, (3) University of Texas MD Anderson Cancer Center, Houston, TX, (4) MD Anderson Cancer Center, Houston, TX, (5) MD Anderson Cancer Center, Houston, TX, AF, (6) Md Anderson Cancer Center, ,,(7) MD Anderson Cancer Center, Houston, TX, (8) Ut Md Anderson Cancer Center, ,,(9) The University Of Texas Md Anderson Cancer Center, ,,(10) The University of Texas MD Anderson Cancer Center, Houston, TX, (11) The University of Texas MD Anderson Cancer Center, Houston, TX, (12) UT MD Anderson Cancer Center, Houston, TX, (13) UT MD Anderson Cancer Center, Houston, TX

Presentations

(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 3

Purpose: To create and test a fully-automated treatment planning approach for labeling, contouring, and planning within the vertebral spine.

Methods: A fully-automated deep-learning solution was developed to localize and contour vertebral bodies in CT scans. For localization, we created X-Net, a dual-input architecture which inputs orthogonal intensity projection images and outputs labeled vertebral centroids for all vertebral bodies within the scan. To leverage shared cranial-caudal information between each orthogonal projection image, X-Net concatenates convolutional layers from sagittal and coronal feature maps between encoding and decoding network arms. X-Net was trained/validated/tested on 359/86/94 scans. To verify localization accuracy, an independent model was trained/validated/tested on 83/41/94 scans from a separate training/validation patient cohort. For contouring, an ensemble of V-Net models was trained/tested on 172/30 scans. Scripts were written to automatically generate radiotherapy plans based on standard institutional practices. To test the entire solution, treatment plans with targets located throughout the spine were created on 60 patient scans. Two radiation oncologists scored 35 of the plans on a three-point scale (acceptable/minor-edits/major-edits).

Results: X-Net labeled the spine with a mean identification-rate and localization-error of 91.9% and 2.6mm for a variety of patients, including those with widespread metastases and surgical implants. Mislabeled vertebral bodies were detected with a sensitivity/specificity of 93.3/82.3%. Vertebral bodies were auto-contoured with median Dice 0.92. Target vertebral bodies were mislabeled for 2/60 cases(one with heavy disease; another with no L5), but all of these were automatically identified prior to plan generation. Radiation oncologists scored plans 30/4/1 as acceptable/minor-edits/major-edits. Median end-to-end planning time was 3.25(range: 2.1—4.8) minutes.

Conclusion: We developed a multi-stage deep-learning approach that reliably labels, verifies, contours, and plans treatments for bony metastases of the spine within 5 minutes. Furthermore, we demonstrated the benefit of using two independent deep-learning models to prevent mislabeled vertebral bodies from propagating to treatment planning.

Funding Support, Disclosures, and Conflict of Interest: Our research group receives funding from the NCI and Varian Medical Systems

Keywords

Treatment Planning, Segmentation, Computer Vision

Taxonomy

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

Contact Email