Preliminary evaluation of using machine learning to prioritise cancer patients for proton radiotherapy by predicting dose to organs at risk

Abstract

The investment in proton radiation therapy raises the question of how cancer patients should be prioritised for this treatment method. A large advantage to proton therapy is that one can minimise the radiation received by organs at risk. Machine learning might be used for predicting the impact of organs for different radiotherapy methods. Thus, machine learning could be a helpful tool in the prioritisation process.

In this thesis, spatial features were extracted from medical images of head and neck cancer patients. These features were used for analysis of photon dose distributions of target volumes and organs at risk. Additionally, correlations between features were investigated. It was confirmed that the distance between radiation target volumes and organs at risk correlated with the dose received by the organs. Further, the analysis contributed to an understanding of which spatial features were expected to affect the dose given to organs at risk.

Within machine learning, both classification and regression algorithms were tested for dose prediction by using the extracted spatial features. In addition, different combinations of algorithms and features were evaluated. The algorithms with the highest accuracy scores were Logistic Regression and Random Forest.

However, the machine learning models either overfitted or underfitted. Thus, other features and machine learning methods should be tested. A possibility for future work is to use deep learning for constructing spatial features and prioritising patients for proton therapy.