MENU

Click here to

√ó

Are you sure ?

Yes, do it No, cancel

Multi-Strategy Machine Learning Auto Planning for Liver SBRT: Improving Quality, Consistency and Efficiency for a Complex Treatment

H Prichard*, J Wo, T Hong, Y Wang, Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To investigate if a machine learning model employing multiple strategies can auto plan liver SBRT at various locations to provide similar or improved plan quality over the state-of-the-art manual planning using multicriteria optimization (MCO).

Methods: A machine learning optimization (MLO) model was trained in RayStation 9B using 75 MCO plans (34 IMRT + 41 VMAT) with a single lesion receiving 50 Gy in 5 fractions. MLO employs a random forest algorithm to predict voxel dose, with 3D dose distribution further refined by generalized dose-volume rules (called strategy). We developed five strategies, one balancing PTV coverage and organ dose and the other four specialized in sparing chest wall, heart, gastrointestinal organs and lung. The model was validated on 10 representative patients in the training set and tested on 11 consecutive new patients (7 IMRT + 4 VMAT). Two auto plans were created for each patient using the balanced and most relevant specialized strategies.

Results: For all five patients with lesion close to (but not abutting) an organ (chest wall, heart, small bowel, stomach and lung), the balanced auto plan provided similar quality to the clinical manual plan and was preferred over the specialized auto plan. For four patients with lesion abutting an organ (one for kidney and three for chest wall), the auto plan (balanced plan preferred for all but one lesion abutting chest wall) only needed a simple processing on one clinical goal. The last two patients had a large lesion, one abutting chest wall, the other abutting lung and close to heart. The specialized plan was preferred and needed processing of two clinical goals (PTV coverage + chest wall or heart dose).

Conclusion: With little or no human involvement, the multi-strategy MLO model trained by MCO plans can generate auto plans matching the quality of clinical MCO plans.

Download ePoster [PDF]

Keywords

Treatment Planning

Taxonomy

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

Contact Email