MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Plan Metric and 3D Dose Prediction for Improved Online Adaptive Decisions in MR-Guided Radiation Therapy

A Thomas1*, Y Fu1 , P Parikh2 , D Yang1 , (1) Department of Radiation Oncology, Washington University School of Medicine, Saint Louis, MO, (2) Department of Radiation Oncology, Henry Ford Cancer Institute, Detroit, MI

Presentations

(Thursday, 7/18/2019) 10:00 AM - 12:00 PM

Room: 225BCD

Purpose: Plan re-optimization in online adaptive RT often cannot receive comparable time and scrutiny as fully developed, offline plans. While it is usually straightforward to determine that adapted plans are safe for delivery, it is more difficult to assess plan quality in relation to more traditional plans established offline. We developed several prediction models with various levels of complexity in order to not only understand the range of plan quality but also better inform online decision-making in adaptive MR-IGRT.

Methods: Approximately 100 patients from our institution undergoing adaptive MR-IGRT for various abdomen cancers (pancreas, liver, adrenal) were included. Plans from both Co-60 and linac-based RT were modeled, though separately. Simple linear models with patient anatomy/geometry inputs and plan metric (V95, D95) outputs were initially developed and tested. Then, voxel-level 3D dose prediction models were also initiated using both shallow and convolutional neural networks.

Results: Various tumor sites could generally be combined without causing degradation to the model’s predictive ability. However, the sole removal of pancreas plans to their own separate model enabled the greatest accuracy of prediction for all other tumor sites. A careful balance between model accuracy and robustness was required – e.g. a 3-variable linear model could predict plan metrics such as V95 and D95 within ~6% uncertainty for a 50 patient cohort. Similarly, shallow neural network design was required to be simple in order to avoid overfitting. Interestingly, shallow neural networks could predict 3D dose values with less than 10% error in most cases, yet they performed no better than simple linear models for predicting plan metrics such as V95 and D95.

Conclusion: Various prediction models for treatment plans in online adaptive MR-IGRT were developed and tested. Such models will enable improved adaptive planning strategies and workflows through more informed plan optimization and evaluation in real time.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email