Machine Learning for Radiotherapy QA

Important results from the project

We have a system that simultaneously measures the radiation dose in a phantom and the flow of high-energy particles (HP) that creates the radiation dose. The aim of the project was to train an ML model so that you can predict the radiation dose from the HP flow. We then wanted to analyze how we could implement a data pipeline that supports the ML model. Finally, we wanted to increase our ability to operate and develop future ML and AI-based project through the implementation of the project together with hired specialists .

Expected long term effects

After this project, we see that it is possible to filter out the non-therapeutic part of HP if we train the ML filter against theoretical data generated from a model. But it turns out that the tools we have to convert the filtered HP signal to phantom dose do not provide sufficient compliance with verification data. ScandiDos has increased its knowledge in ML and AI technology so we are better equipped to run similar projects in the future.

Approach and implementation

The project was divided into four parts, WP1, WP2, WP3 and WP4. WP1 aimed to improve and train an ML-based model to predict phantom doses based on the measured HP flow, as well as implement a prototype in existing software. As the results from WP1 were not good enough, we chose not to do WP2 and WP3; analyze and implement a data pipeline. WP4 would increase ScandiDo´s knowledge in ML and AI technology. This was integrated into WP1 through collaboration with hired specialists.