Machine Learning for Predicting Radiotherapy Treatment Plan Complexity in Lung Cancer

Abstract

This thesis investigates the application of machine learning models, specifically Random Forest and XGBoost, to predict the complexity of volumetric modulated arc therapy (VMAT) treatment plans for lung cancer. Accurate prediction of treatment complexity could optimise radiotherapy planning, as VMAT complexity varies considerably with patient-specific characteristics and diverse planning approaches. This variability impacts treatment consistency and quality assurance (QA) processes, highlighting the need for predictive tools to manage complexity within clinical workflows. Using a dataset of historically treated patients that includes demographics, tumour characteristics, and radiotherapy planning parameters, this study developed and evaluated binary and multiclass classification models for VMAT complexity. Data curation involved systematic imputation, feature optimisation via principal component analysis (PCA), and training and validation across independent datasets. The probability distributions assessed the confidence of the model, providing insight into the predictive reliability. The results demonstrate that both models were effective in predicting VMAT complexity during cross-validation, with Random Forest showing a slight advantage in accuracy (0.77) and precision (0.76 for binary, 0.70 for multiclass) over XGBoost. However, both models faced challenges in classifying intermediate complexity levels, and calibration issues were observed, particularly in multiclass tasks. These findings highlight the effect of diverse planning practices on complexity prediction and suggest the need for refined feature engineering to improve classification robustness. In conclusion, this research highlights the potential of machine learning models to support radiotherapy planning by identifying cases with greater complexity, aiding in planning consistency and resource allocation. Future studies should expand dataset diversity and explore advanced modelling techniques to improve predictive accuracy across all complexity levels, supporting the integration of machine learning tools into clinical radiotherapy planning and optimising patient care.

Date of Award	8 Apr 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Alan Mcwilliam (Supervisor)