Background: In robotic radiation therapy, ionizing radiation is delivered by a linear accelerator that is
mounted on a robotic arm. This allows to deliver dose from practically arbitrary many directions.
Additionally, multi-Leaf-Collimators (MLCs) offer a great flexibility in shaping beams. However, find-
ing optimal treatment beams is a computationally challenging task. However, similarities be-
tween patients exist and convolutional neural networks (CNNs) offer a method to identify beneficial
beams. Once trained CNNs show comparatively fast inference. Hence, they can be used to speed up treatment planning times and improve treatment plan quality.
Taks: The current approach for generating MLC-beams should be evaluated and training data
should be generated. The existing framework should be extended to generate image features for
MLC-beams. Using these features, a CNNs should be trained to predict each beam’s quality. Infer-
ence on the trained network for treatment plan generation should be implemented in the existing
Java framework and the performance evaluated.
Specific steps include:
- Familiarize with the treatment planning framework.
- Evaluate conventional MLC-beam generation.
- Generate training data.
- Implement a pipeline to train a CNN.
- Integrate inference into the treatment planning framework.
- Evaluate the CNN-based approach with respect to computation time and treatment plan
Requirements: Good programming skills (ideally in Python+Java), ability to work independently, ideally experience with machine learning
Not required is experience in medicine or radiation therapy.
References: Gerlach, S., Fürweger, C., Hofmann, T. and Schlaefer, A. (2020), Feasibility and analysis
of CNN‐based candidate beam generation for robotic radiosurgery. Med. Phys., 47: 3806-3815.
Please contact me if you want to know more about the topic.