Introduction

Introduction Brain tumours are the largest cause of mortality in the paediatric oncological population, with a variable survival rate dependant on tumour type(1). Many challenges are faced when assessing survival risk factors, and studies focusing on imaging features such as perfusion, a biomarker of angiogenesis – commonly elevated in high grade tumours, or diffusion, a marker of cellularity – commonly restricted in high grade tumours, imaging have defined features that show significantly elevated hazard ratios from classical statistical analysis using Cox regression(2,3). In this study we combine both perfusion and diffusion imaging with Bayesian survival analysis coupled with unsupervised machine learning to provide a model that defines two separate clusters of high and low risk tumours in our cohort, with significantly different survival statistics between the two groups.

Methods

68 children with a suspected brain tumour were enrolled into this study across 4 clinical sites (Ethics number: 04/MRE0/41, Birmingham Children’s Hospital, Alder Hey, Royal Victoria Infirmary, Nottingham). Tumours varied in grade and type (detailed in Table 1), confirmed by biopsy. Diffusion and perfusion imaging covering the tumour volume were performed in all participants, with imaging parameters shown in table 2.
T₂ weighted imaging was used to drawn regions of interest outlining the tumour, and mean, standard deviation, kurtosis, and skewness of the whole brain and ROI were calculated in Matlab (2018b The Mathworks, MA) for ADC and perfusion maps. Tumour volume was calculated from the T2 weighted imaging. Iterative Bayesian survival analysis (R, iterativeBMAsurv) were used to determine key imaging features to predict high- and low-risk patients, with subsequent unsupervised clustering (k means) and features then used with supervised learning (Random Forest, AdaBoost, Support Vector Machine, performed in Orange) to predict low/high risk class. Low/high risk clusters were also used for Kaplain-Meier analysis and Cox regression to determine group hazard ratio.

Results

Diffusion and perfusion imaging provide significant features in assessing survival
Cox regression revealed perfusion features to have significant Hazard ratios, for example Uncorrected CBV ROI mean (HR = 3.1, Confidence Intervals (CI) = 1.5-6.6, p = 0.003). Bayesian analysis revealed the five most likely features to predict survival (posterior probability, posterior mean coefficient) to be Uncorrected CBV ROI mean (96%, 0.85), K2 ROI mean (39%, -0.17), Uncorrected CBV whole-brain mean (40%, 0.3), tumour volume (27%, 0.05), and ADC ROI kurtosis (20%, 0.02).

Unsupervised clustering detects distinct groups with significantly different survival characteristics
Using the Bayesian imaging features, k means clustering revealed two distinct clusters, shown in figure 1A, which when used with Kaplan-Meier analysis revealed a significant difference between a high and low-risk population (see figure 1B, p = 0.0015). There were a number of significant differences in imaging features between the high and low-risk populations, for example, ADC kurtosis (10.1 ± 5.3 vs 4.3 ± 1.8, p < 0.001) and K2 ROI mean (0.0186 ± 0.008 vs 0.028 ± 0.018 s-1, p = 0.007) as well as there being a mixture of high-and low-grade tumours in both clusters. Full results showing significant high and low-risk features and cluster information are shown in table 3.

Supervised machine learning can be used to distinguish between high/low-risk clusters
Supervised machine learning using imaging features showed that the Bayesian features combined with a neural network, after 10-fold stratified cross-validation, provided the most accurate classification of high/low risk patients (precision = 98%, F-statistic = 0.98). Further combinations of all diffusion and perfusion ROI features with principal component analysis with a neural network yielded similar results (precision = 94%, F-statistic = 0.94) and the addition of tumour volume to the aforementioned did not provide increased precision or F-statistic results. The full classifier results for features combinations are presented in table 4.

Discussion

This study has demonstrated the potential for using functional imaging to assess for novel survival populations in a mixed pediatric brain tumour cohort. Here we have demonstrated a novel sub-class of brain tumours which demonstrate significantly lower survival probability, and may present an opportunity for targeted therapeutic intervention and research study recruitment. Indeed, the high-risk cohort had a number of high-grade survivors present which either lack long term follow-up or, more interestingly, were low-grade and had imaging features more akin to high-grade tumours.
This work represents a novel step forward in the understanding of survival in paediatric brain tumours, with potential to be expanded through the addition of ASL derived perfusion imaging or spectroscopic data.

Acknowledgements

We would like to acknowledge funding from the Cancer Research UK and EPSRC Cancer Imaging Programme at the Children’s Cancer and Leukaemia Group (CCLG) in association with the MRC and Department of Health (England) (C7809/A10342), the Cancer Research UK and NIHR Experimental Cancer Medicine Centre Paediatric Network (C8232/A25261), the Medical Research Council – Health Data Research UK Substantive Site and Help Harry Help Others charity. Professor Peet is funded through an NIHR Research Professorship, NIHR-RP-R2-12-019. Stephen Powell gratefully acknowledges financial support from EPSRC through a studentship from the Physical Sciences for Health Centre for Doctoral Training (EP/L016346/1). Theodoros Arvanitis is partially funded by the MRC (HDR UK). We would also like to acknowledge the MR radiographers at Birmingham Children’s Hospital, Alder Hey Children’s Hospital, the Royal Victoria Infirmary in Newcastle and Nottingham Children’s Hospital for scanning the patients in this study. We would also like to thank Selene Rowe at Nottingham University Hospitals NHS Trust for help with gaining MRI protocol information. Dr James Grist is funded by the Little Princess Trust (CCLGA 2017 15).