1st supervisor: Thomas Booth, King’s College London
2nd supervisor: Marc Modat, King’s College London
Aim of the PhD Project:
What: MRIs play an important role in diagnosis and treatment response assessment of brain tumours.
Why: Unknown whether the MRI performed at each stage of the patient pathway after initial treatment changes outcomes (morbidity, mortality or health economics).
Aim: Develop in silico approach modelling each MRI time-point.
How: Deep learning based spatio-temporal model of tumour progression
Project Description / Background:
Glioblastoma is the most common primary malignant brain tumour in adults with ~2,500 cases diagnosed annually in England [1]. It has one of the worst prognoses of any cancer [2].
The current standard-of-care for newly diagnosed patients consists of surgery followed by radiotherapy and chemotherapy [3]. MRI performed at regular intervals after treatment is believed to play an important role in treatment response assessment [4-8]. However, it is unknown whether the MRI performed at each stage of the patient pathway after glioblastoma treatment changes outcomes (morbidity/mortality/health economics) whilst frequent imaging investigations can contribute to some of the financial difficulties that patients with glioblastoma face and can cause anxiety.
Optimizing the pathway for post-treatment imaging in patients by discovering which imaging performed provides evidence-based benefit, will not only reduce anxiety surrounding imaging investigations, but may reduce some of the financial burden associated with this disease for patients and the healthcare system.
Randomised controlled trials (RCTSs) provide the highest clinical evidence [9], however performing one for each imaging time-point would be challenging for multiple practical reasons including excessive cost and poor patient recruitment – many patients and clinicians believing that missing routine imaging follow-up would be disadvantageous.
Alternatively, using machine learning with Bayesian methodology, it is envisaged that the predicted contribution (with confidence intervals) of each co-variate, including each discrete imaging time-point, can be computed in terms of morbidity, mortality and health economic outcomes. MRI co-variates of increasing granularity include (1) scan performed or not; (2) individual sequences, segmented volume, brain location and reported treatment response outcome; (3) radiomic features. Non-imaging co-variates include clinical (Karnofsky Performance Status), treatment type given, pathology (molecular markers) and demographics (sex, age). The strength of an in silico approach (as opposed to an RCT) is that interrogation of the MRI at different levels of granularity and the non-imaging co-variates is more feasible.
Alternatively, using machine learning with Bayesian methodology, it is envisaged that the predicted contribution (with confidence intervals) of each co-variate, including each discrete imaging time-point, can be computed in terms of morbidity, mortality and health economic outcomes. MRI co-variates of increasing granularity include (1) scan performed or not; (2) individual sequences, segmented volume, brain location and reported treatment response outcome; (3) radiomic features. Non-imaging co-variates include clinical (Karnofsky Performance Status), treatment type given, pathology (molecular markers) and demographics (sex, age). The strength of an in silico approach (as opposed to an RCT) is that interrogation of the MRI at different levels of granularity and the non-imaging co-variates is more feasible.
Gaussian Processes, which have the advantage to implicitly capture uncertainly associated with their predictions, may prove to be optimal to achieve this. Modelling would allow predictions as to whether imaging at any particular time-point is of value or not. It would also help to predict which key co-variates, including imaging time-points, should be targeted in future non-CDT RCTs thereby optimizing research resources. Creation of a spatio-temporal model of tumour progression using retrospective clinical data will be central to this process (Box 1).
This study will also demonstrate a key priority of the James Lind Alliance Priority Setting Partnership (of patients, carers and clinicians) to determine the value of neuro-oncological interval scanning. This will help inform future clinical practice and ensure that all imaging performed is appropriately, timely and has an evidence base.
It is envisaged that the results, once disseminated, will inform standard practice in all the neuro-oncology centres throughout the UK. Furthermore, given that this is a value-based healthcare project, CE marking of models produced can be avoided with quicker implementation into hospitals. The expected timeline to achieve the initial results is within three years from the proposed start date.
Box 1. Spatio-temporal model methodology
| First, a convolutional neural network will be trained to identify tumour area within a patient’s brain. Second, a spatio-temporal model will be train to track the evolution of the tumour across different timepoints. Our model will learn from retrospective data rather than rely on physical model, as previously presented in the literature [10]. Third, we will condition our model in order to simulate various longitudinal patterns, such as tumour growth or reduction arising from treatment. As an output, we will be able to simulate any tumour evolution given an input image, simulated or real. Lastly, these tools will enable the study of the effects of a change of management in each subject, providing insights on how different treatments can impact the subject’s outcome. |
Candidate Profile
Highly motivated and academically excellent (minimum upper 2nd or 1st class undergraduate degree). Computational experience required. Knowledge of python is advantageous, in particular libraries such as pandas and scikit learn.
References:
[1]. Brodbelt, A., et al., Glioblastoma in England: 2007-2011. Eur J Cancer, 2015. 51(4): p. 533-42.
[2]. Stupp, R., et al., Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. N Engl J Med, 2005. 352(10): p. 987-96.
[3]. Stupp, R., et al., Effects of radiotherapy with concomitant and adjuvant temozolomide versus radiotherapy alone on survival in glioblastoma in a randomised phase III study: 5-year analysis of the EORTC-NCIC trial. Lancet Oncol, 2009. 10(5): p. 459-66.
[4]. Sanghera, P., et al., The concepts, diagnosis and management of early imaging changes after therapy for glioblastomas. Clin Oncol (R Coll Radiol), 2012. 24(3): p. 216-27.
[5]. Stupp, R., et al., High-grade glioma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol, 2014. 25 Suppl 3: p. iii93-101.
[6]. Weller, M., et al., European Association for Neuro-Oncology (EANO) guideline on the diagnosis and treatment of adult astrocytic and oligodendroglial gliomas. Lancet Oncol, 2017. 18(6): p. e315- e329.
[7]. NCCN Guidelines. NCCN guidelines for treatment of cancer by site. Central Nervous System Cancers. National Comprehensive Cancer Network. http://www.nccn.org/professionals/physician_gls/f_guidelines.asp#cns. Updated March 2018. Accessed July 2018.
[8]. NICE pathways. https://pathways.nice.org.uk/pathways/brain-tumours-and-metastases/brain-tumours-and-metastases-oNICE verview#content=view-node%3Anodes-follow-up&path=view%3A/pathways/brain-tumours-and-metastases/brain-cancer-glioma.xml [Accessed 1 December 2018].
[9]. Howick J, et al. The Oxford 2011 Levels of Evidence. Oxford Centre for Evidence-Based Medicine, Oxford; 2016. Available at: http://www.cebm.net/index.aspx?o=5653. [Accessed 1 August 2018].
[10] Menze BH, et al. A generative approach for image-based modeling of tumor growth. Inf Process Med Imaging. 2011;22:735-47.
