Improving Treatment of Glioblastoma: Distinguishing Progression From Pseudoprogression

Improving Treatment of Glioblastoma by Distinguishing Progression From Pseudoprogression by Applying Machine Learning Techniques to Routine Clinical Data

Glioblastoma is the most aggressive kind of brain cancer and leads on average to 20 years of life lost, more than any other cancer. MRI images of the brain are taken before the operation, and every few months after treatment, to see if the cancer regrows. It can be hard for doctors to tell if what they see in these images represent growing cancer or a sideeffect of treatment. The similarity of the appearance of the treatment side-effects to cancer is confusing and is known as "pseudoprogression" (as opposed to true cancer progression).

If doctors mistake the appearance of treatment side-effects for growing cancer, they may think that the treatment is failing and change the patient's treatment too early or put them into a clinical trial. This means that patients may not be given the full treatment and the results from some clinical trials cannot be trusted.

The aim of this study is to provide doctors with a computer program that will use MRI images of the brain that are routinely obtained throughout treatment, in order to help them more accurately identify when the cancer regrows.

The impact of pseudoprogression is significant on patient care and medical research. The existing evidence shows that it is feasible to use Support Vector Machine and Deep Learning classification models for predicting survival using routine MRI images as well as differentiating progression from pseudoprogression. The investigators wish to capture signal changing over time in routine MRI images using parametric response maps (via a state-of-the-art postoperative-to preoperative image registration method that they have developed) and use such classifiers to differentiate progression from pseudoprogression. The research the investigators are proposing is needed in order to provide a solution to the problem of pseudoprogression and be implemented across the NHS easily and efficiently. Importantly, this does not depend on advanced imaging techniques.

Data collected at KCH from the last 24 months shows that, even at a leading glioma imaging centre, only 66% of patients had advanced imaging (e.g. DSC-MRI) performed at the time of increase in contrast-enhancement i.e. possible progression. The primary aim of this research is to use routine clinical MRI data in order to train the classifier. This will increase the utility of the classifier, as such routine MRI data can be acquired by all imaging centres, and the new classifier can therefore provide a much more cost-efficient solution than an alternative classifier which may depend on advanced imaging techniques.

Initial training, testing and cross validation of a classification model will be carried out using MRI data of glioblastoma obtained from publicly-accessible imaging archives and King's College Hospital (KCH), London. For clinical validation, the trained model will undergo testing using MRI data from patients recruited prospectively.

Inclusion Criteria:

• Diagnosed with glioblastoma (World Health Organisation grade IV)
• Patient undergoing the standard Stupp treatment regimen
• Have had a pre-surgery scan and at least one follow-up scan post-chemoradiation

Exclusion Criteria:

• Insufficient clinical and radiological follow-up
• The patient's treatment deviates greatly from the standard Stupp regimen, such as they are recruited into interventional trials and sufficient information is not known about the patient's trial treatment
• Patients receiving treatment with Angiogenesis inhibitors such as bevacizumab prior to completion of the Stupp regimen