Click here to

Session: Advances in SRS/SBRT [Return to Session]

Deep Reinforcement Learning Based Intelligent Virtual Planner of Gamma Knife Radiosurgery for Vestibular Schwannoma

Y Liu1, C Shen2, T Wang1, J Zhang1, X Yang1, T Liu1, S Kahn1, H Shu1, Z Tian3*, (1) Emory University, Atlanta, GA, (2) University of Texas Southwestern Medical Center, Dallas, TX, (3) University of Chicago, Chicago, IL


TU-B-202-1 (Tuesday, 7/12/2022) 8:30 AM - 9:30 AM [Eastern Time (GMT-4)]

Room 202

Purpose: Vestibular Schwannoma cases are often very challenging for treatment planning of Gamma Knife (GK) radiosurgery, due to the irregularly-shaped target and its proximity to both brainstem and cochlea. During inverse treatment planning, planners often need to repetitively adjust the priorities among multiple planning objectives via a trial-and-error process until a clinically optimal GK plan is achieved. This study aims to develop a deep reinforcement learning based method to automate the priority tuning process.

Methods: We built a virtual planner network using deep convolutional neural networks. It consists of multiple subnetworks, with each corresponding to a priority tuning action option. Our network is designed to take a vector of multiple plan metrics of the intermediate plan as input, and predicts the reward for each action option in terms of plan quality change. A scoring function was designed to evaluate the plan quality change to calculate the received reward of a tuning action. The action option that has the largest predicted reward is selected as the action at the current tuning step. An end-to-end deep reinforcement learning framework was employed to train the virtual planner.

Results: In this study, we used 5 training cases, 5 validation cases and 16 testing cases. The average plan scores obtained by an experienced GK planner through manual priority tuning were 5.28 ± 0.23, 4.97 ± 0.44 and 5.22 ± 0.26 for these three datasets, respectively. Our virtual planner network achieved competitive plan scores of 5.42 ± 0.11, 5.10 ± 0. 42, 5.28 ± 0.20, respectively.

Conclusion: Our virtual planner network can automatically generate GK plans of comparable or slightly higher quality comparing with plans generated by human planners for vestibular schwannoma cases. It can potentially be incorporated into the clinical workflow as a planning assistance to improve GK planning efficiency and reduce plan quality variation.

Funding Support, Disclosures, and Conflict of Interest: This study is supported by Winship Cancer Institute #IRG-17-181-06 from the American Cancer Society.


Gamma Knife, Inverse Planning, Radiosurgery


TH- External Beam- Photons: gammaknife

Contact Email