Machine Learning for Automatic Prediction of Tumour Growth from CT Images





1.   Background for Image Guided Radiotherapy


Radiation treatment plans for the treatment of cancers are designed primarily from CT datasets. The tumour within the patient is imaged using a cross-sectional scanner (CT) creating a three-dimensional electron density map (Figure 1). The radiation oncologist, using fused magnetic resonance imaging (MRI) and positron emission tomography (PET) scans as well as clinical information creates a gross tumour volume (GTV outlined in red) (Figure 2).


 CT_GTV                                         CTPET_GTV


                                                                                             Figure 1                                                                                                      Figure 2


Using geometric expansion tools, a margin around the GTV is created and modified according to the likelihood of tumour involvement by microscopic disease. This region is called the Clinical Target Volume (CTV outlined in blue in Figure 3). The physician must judge the probability that the tumour will spread into surrounding regions, which depends on the density of surrounding objects. For example, it is more difficult for a tumour to traverse bone than fat. The creation of the CTV is time consuming and would require several hours of contouring to create repeated treatment plans during a course of treatment as well as the original initial treatment plan.



Figure 3: Planning CT scan with GTV (red), CTV (blue) and isodose lines from conventional radiotherapy beam arrangement.



The density of tissue is proportional to proton density, which in turn is proportional to electron density. Higher molecular structures such as calcium containing bone are denser than fat which contains a higher proportion of water on a CT scan (e.g. bone is white and air is black with fat being grey).


The aim of modern radiotherapy is to treat the tumour yet causing the least amount of damage to the surrounding normal structures. Adapting the treatment and reducing the clinical target volume may reduce the dose that surrounding tissues receive, allowing sparing of normal tissue functions such as saliva production in the treatment of head and neck cancers. With the introduction of image-guided radiotherapy, images of the patient in their treatment position are recorded intermittently during their treatment. Responding tumours shrink during their treatment and so the original tumour size and position change with respect to their original status. Their relationship to surrounding structures also changes temporally.


This daily adaptation of the GTV and CTV is time consuming and would rely upon daily physician adjustment of the contoured tissue. Automation of this process would increase the feasibility of DART (Dynamic adaptive radiotherapy).  


2.   Summary and Research Objectives


The aim of modern radiotherapy is to treat the tumour whilst causing the least amount of damage to the surrounding normal structures. The region for tumour treatment is called the Clinical Target Volume (CTV), which is a drawn manually using CT scan images.  The creation of the CTV is very time consuming and requires several hours of manual contouring to create an initial treatment plan.


Adaptive radiotherapy, in which the treated region is modified as treatment progresses, remains the holy grail of personalised radiotherapy, but as yet is out of reach of daily practice because of the extensive cost of repeated manual definition of the CTV.  The recent development of machine learning and the availability of powerful computers hold great promise that this research question can be addressed systematically.    


The purpose of this project is to develop and assess novel machine learning algorithms to automate the creation of CTVs. This project will initiate a collaborative group comprising a clinical oncologist from Royal Devon and Exeter Hospital and computer scientists from the University of Exeter, with aims to develop intelligent machine learning algorithms for personalised radiotherapy.  In summary, the main objectives of the project are: 


·       Developing image segmentation algorithms to automatically detect tumour regions from CT images.

·       Designing dynamic and sequential machine learning algorithms to model tumour growth from CT images and therefore automatically produce CTVs.

·       Validating the proposed methods using the available CTVs created by radiation oncologists.




People:     David Hwang (Royal Devon Hospital)

                Yiming Ying (Exeter University)

                Richard Everson (Exeter University)




Return to Yiming Ying's home page