Machine Learning for Radiotherapy Treatment of Prostate Cancer

Abstract

External beam radiotherapy (EBRT) and brachytherapy (BT) are both forms of radiation treatment used for prostate cancer to destroy cancer cells. EBRT applies the radiation externally while BT involves placing radioactive seeds inside the prostate. At Guy’s Cancer Centre, both treatment modalities are performed depending on various factors. Each of the treatment modalities involves different imaging modalities used for treatment planning, delivery and follow-up. However, both have some overlapped clinical tasks such as defining the clinical target volume (CTV) and organs at risk (OARs) from imaging data. The work described in this thesis aims to perform research to promote clinical translation of machine learning (ML) techniques to streamline workflows in EBRT and BT. The first piece of work in this thesis focuses on an ML-based segmentation model for prostate MRI. One of the main challenges affecting clinical adoption of ML in MRI segmentation is the domain shift problem. The findings of this piece of work reveal for the first time the significant impact on model performance of using different acquisition/annotation protocols, even if using the same scanner vendor/field strength. It is shown that training an ML model with data that covers the important sources of domain shift can produce a robust model with good generalisability performance. The next piece of work investigates the possibility of race bias in ML-based prostate MRI segmentation. Through experiments on a controlled dataset of White and Black patients, it is shown that the model performance gap between Black and White subjects is dependent on the level of (im)balance between Black and White subjects in the training data. Again, it is shown that training using demographically balanced data can produce a fair and robust model. The conclusion from both of these pieces of work is that model performance can be robust if the training data is sufficiently diverse, both in terms of image characteristics and patient demographics. Building upon these analyses, the thesis next investigates the clinical utility of a diagnostic prostate MRI model trained on diverse data and externally validates it on in-house clinical data. The evaluation of this model encompasses not only standard quantitative metrics but also measurement of inter-observer variability in manual segmentation and assessments of performance on downstream clinical tasks. Next, the thesis investigates the clinical utility of multi-organ ML-based segmentation models. Here, two models are investigated: one for planning MRI called the “FIMRAa-P” model and another radiotherapy CT model called the “PelvisMA-CT” model. Both models are extensively evaluated quantitatively and qualitatively by five observers. The agreement between the quantitative metrics and the qualitative clinical metrics is also investigated for each clinical structure, revealing generally poor agreement between the two. It is also shown that this agreement is dependent on the structure being segmented and the profession of the clinicians who perform the evaluations. One of the main clinical translation outcomes of this thesis is the deployment of PelvisMA-CT by the Clinical Scientific Computing (CSC) group at GSTFT, and its integration into a contouring application called GSTTAutoSeg. This model is currently being used clinically at Guy’s Cancer Centre and the thesis presents the results of a monitoring and enhancement study based on its ongoing clinical use. Overall, the thesis presents a number of key contributions, all aimed at promoting clinical translation of ML in EBRT and BT. It is hoped that the work performed will accelerate the benefits of ML in radiotherapy treatment planning and delivery and ensure that all patients benefit from the introduction of the thoroughly evaluated new technology.