Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Using Deep Learning Auto-Planning for Evaluating the Dosimetric Impact of Deep Learning Auto-Segmentation Without Human Intervention Over Multiple Dose Escalation Schemes for Lung SBRT Treatment Planning

M Holmstrom1*, E Samuelsson1, Y Wang2, (1) RaySearch Laboratories, Stockholm, Sweden, (2) Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA

Presentations

PO-GePV-M-238 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: This work presents an integrated solution for automatic segmentation and planning for lung SBRT. Running auto-planning with multiple dose escalation schemes allowed the dosimetric impact of auto-segmentation to be evaluated.

Methods: The DL planning model was trained in RayStation 10B, using 99 lung SBRT plans delivering 50 Gy in five fractions. The DL segmentation model originated from an official RaySearch model for heart, spinal cord, esophagus, and both lungs. This was extended with a sub-model for chest wall, trained in RayStation 10B using the same lung SBRT patients. The planning model was previously validated, allowing it to be used as a virtual planner. It generates three types of auto-plans, following different dose escalation schemes (110%, 115%, and 120%) to an ITV-4mm heater in the center of the lesion. For 15 test patients, auto-planning was run based on manual and DL ROIs, separately. The former was evaluated on manual ROIs only, while the latter was evaluated on both manual and DL ROIs. PTV, CTV, and heater were always manually delineated.

Results: When evaluating on manual ROIs, auto-planning based on manual and DL ROIs resulted in similar lung V20 dose, heart mean dose, and spinal cord maximum dose for all test patients and all dose escalation schemes. Studying target coverage and chest wall sparing, however, only 12/15 patients had adequate coverage in the DL ROI-based plans. The remaining three patients had lesions abutting the chest wall, for which the lesion was subtracted from the chest wall in the manual contours. This behavior was not captured by the DL segmentation model.

Conclusion: Generally, auto-planning based on DL ROIs instead of manual ROIs did not have significant dosimetric effects, except for some patients with lesions abutting the chest wall. Further, the choice of dose escalation scheme had no effect in these dosimetric analyses.

Funding Support, Disclosures, and Conflict of Interest: Mats Holmstrom and Elin Samuelsson are employees of RaySearch Laboratories, Stockholm, Sweden.

ePosters

    Keywords

    Treatment Planning

    Taxonomy

    IM/TH- Image Segmentation Techniques: Machine Learning

    Contact Email

    Share: