IMG-06. PREDICTING SURVIVAL FROM PERFUSION AND DIFFUSION MRI BY MACHINE LEARNING

Abstract

INTRODUCTION Magnetic Resonance Imaging (MRI) is routinely used in the assessment of children’s brain tumours. Reduced diffusion and increased perfusion on MRI are commonly associated with higher grade but there is a lack of quantitative data linking these parameters to survival. Machine learning is increasingly being used to develop diagnostic tools but its use in survival analysis is rare. In this study we combine quantitative parameters from diffusion and perfusion MRI with machine learning to develop a model of survival for paediatric brain tumours. METHOD: 69 children from 4 centres (Birmingham, Liverpool, Nottingham, Newcastle) underwent MRI with diffusion and perfusion (dynamic susceptibility contrast) at diagnosis. Images were processed to form ADC, cerebral blood volume (CBV) and vessel leakage correction (K2) parameter maps. Parameter mean, standard deviation and heterogeneity measures (skewness and kurtosis) were calculated from tumour and whole brain and used in iterative Bayesian survival analysis. The features selected were used for k-means clustering and differences in survival between clusters assessed by Kaplan-Meier and Cox-regression. RESULTS Bayesian analysis revealed the 5 top features determining survival to be tumour volume, ADC kurtosis, CBV mean, K2 mean and whole brain CBV mean. K-means clustering using these features showed two distinct clusters (high- and low-risk) which bore significantly different survival characteristics (Hazard Ratio = 5.6). DISCUSSION AND CONCLUSION Diffusion and perfusion MRI can be used to aid the prediction of survival in children’s brain tumours. Tumour perfusion played a particularly important role in predicting survival despite being less routinely measured than diffusion.