Unfortunately we will not be able to share the data due to ethical issues. Codes might be available after the work is published. The folowing sections describe the technical problem and solution approach used for this project.
This work involves developing a machine learning tool for evaluating the quality of radiotherapy treatment plans, which is a highly subjective and experience-dependent task. The underlying dataset is high dimensional (162 by 782) consists of geometrical features, the clinical evaluation criteria and several engineered features based on domain knowledge. There are correlations and interactions between the features e.g. structures in close proximity are likely to receive similar doses and the deliberate sparing of a structure is likely to cause an increased dose to the other structure. Model stacking of various structural-based sub-models and extreme randomised trees are used to capture the interations and to ensure all relevant clinical criteria contribute to the model. Final classification model achieves an 80% cross validation accuracy score.
Radiotherapy treatment is one of the most effective non-surgical cancer treatment modality. It uses external ionising radiotherapy to control or eliminate cancerous cells.
The planning process for radiotherapy treatments consists of repeated plan optimization and plan evaluation. The planner needs to set some optimization parameters in a commercial treatment planning software in order to generate a plan. The plan is evaluated based on a list of treatment evaluation criteria, the so-callede dose-volume histograms and the spatial dose distributions (examined through 2D slices of the treatment site).
Due to patient specific geometrical structural variations, evaluating the quality of a plan is difficult. Clinical evvaluation protocols are based on population-based statistics and are not specifically designed to a patient. If a plan is not deemed satisfactory, the planner adjusts the optimization parameters in order to improve the plan. The optimization process can take serveral minutes and it is uncertain if the adjustment will lead to an improvement. As a consequence, a treatment plan tends to be accepted as long as it achieves the minimal requirements.
In this work, we develop a data-driven plan evaluation tool using machine learning techniques. We build a classification model that predicts plan acceptability based on the geometrical features and the radiation dose to different structures. The model results are feed into a model interpreter which will show the strengths and weaknesses of a given plan. The planner will be able to make unbiased evaluation on the plan quality hence saves time in the planning process and ensures the plan quality is truly optimal.
The figure above is an annotated screen shot of the model interpreter for a plan. In this example we can see the classification score is less than 0.5, suggesting the plan is unsatisfactory. The bar plot shows the top criteria affecting the predicted plan quality. Criteria in red/green are the criteria contributing negatively/positively to plan quality. The planner can read from this plot to identify the strength and weaknesses of a given plan to make an informed evaluation on plan quality and improvement potential.