Magnetic Resonance Imaging

Prior to the concept of Radiomics, magnetic resonance imaging (MRI) studies investigated the nature of posterior fossa tumours using single quantifiable features. The primary aim of these studies was to distinguish tumour types by hypothesising that singular extracted values would differ significantly enough to allow accurate classification. This section focuses specifically on those studies which have used singular measures in order to investigate medulloblastoma.

Distinguishing medulloblastoma from other posterior fossa tumours

By far the most investigated metric is apparent diffusion coefficient (ADC), acquired using diffusion weighted imaging (DWI). DWI is a non-contrast MR technique which provides spatial information on the Brownian motion of water molecules in tissue, including intra-, trans- and extra-cellular diffusion as well as perfusion (reference). Acquiring fully quantitative value of diffusion requires long scan times, so ADC is used as a semi-quantitative marker for diffusion which can be acquired in a much shorter time.
In order to achieve an accurate diagnosis of the type of posterior fossa tumour, a biopsy is required is order to assess histological features, one of which is cellularity. Cellularity describes the density of cells within the tumour and is known to vary based on tumour type. A meta-analysis performed by Surov et al., (2017) \cite{Surov2017} demonstrated that all investigated tumour types show a significant inverse relationship between cellularity and ADC, suggesting ADC could be used as a surrogate for cellularity and reduce the need for a biopsy.
The assessment of ADC in paediatric posterior fossa tumours has been thoroughly investigated (references), with the earliest study being conducted in 2001 by Gauvin et al., \cite{Gauvain2001}. The ADC within the MB is commonly compared with other posterior fossa tumours, notably pilocytic astrocytoma (PA) and ependymoma (EP) \cite{Gauvain2001,Yamasaki2005,Rumboldt2006,Schneider2007,Yamashita2009,Jaremko2010,Gimi2012,Bull2012,Mohamed_2013,Rodriguez2014,Pierce2014,Pierce2014a,Assis2015,Dom_nguez_Pinilla_2016,Zitouni2017}.
The consensus of this literature is that MB possesses a significantly reduced mean ADC when compared to PA and EP, with PA possessing a significantly higher ADC than EP. However, in some cases no significant difference is observed between MB and EP \cite{Jaremko2010,Bull2012}, or EP and PA \cite{Mohamed2013}. In addition, some papers have instead considered the minimum ADC as a key feature of tumour types \cite{Yamashita2009,Jaremko2010,Yeom2013,Pierce2014,Pierce2014a}, which have reported similar trends between MB, EP and PA. These results demonstrate that ADC has the potential to be a useful marker for differentiating tumour types. Numerous studies have attempted to classify MB from other tumour types by calculating an upper ADC threshold, whereby a tumour containing a mean ADC lower than this threshold is classified as MB.
When a threshold distinguishing MB and EP is concerned, Yamasaki et al., (2005) \cite{Yamasaki2005} reported that a threshold if 1.00x10-3 mm2/s which can distinguish MB (N=9) from EP (N=6) with a sensitivity and specificity of 100%. Similarly, Pierce et al., (2014) \cite{Pierce2014} reported the same threshold, however a lightly lower sensitivity and specificity was observed (97%/90%). This could possible be due to a larger sample size which increased the likelihood of observing more extreme cases. Gimi et al., (2012) \cite{Gimi2012} reported that a slightly lower threshold of 0.91x10-3 mm2/s can distinguish MB from EP with a sensitivity of 79% and sensitivity of 93%.
A threshold separating MB and PA was also reported in numerous studies. Jaremko et al., (2010) \cite{Jaremko2010} reported that a threshold of 0.80x10-3 mm2/s can differentiate MB from PA with a sensitivity of 97% and specificity of 90%. Finally, Rumboldt et al., (2006) \cite{Rumboldt2006} reported that a threshold of 0.90x10-3 mm2/s was able to distinguish MB from non-MB (EP and PA)  with a sensitivity and specificity of 100%. However, the relatively small sample size should be considered when interpreted the high sensitivity and specificity.
The reported thresholds for distinguishing MB from other posterior fossa tumours range from 0.80-1.00x10-3 mm2/s, suggesting a good agreement between studies on the suggested upper limits.
A meta-analysis of the reported values shows that MBs have an average ADC of 0.70 plus/minus 0.16 x10-3 mm2/s, whereas PA and EP have ADCs of 1.60 plus/minus 0.35 x10-3 mm2/s and 1.07 plus/minus 0.22 x10-3 mm2/s. A total of 261 MBs, 78 EPs and 196 PA were included in this meta-analysis. Using the mean and standard deviations to produce a normal distribution, mean ADC values were simulated for each tumour type, where the number of samples for each tumour type match the proportions of tumour types from the meta-analysis. Using ROC analysis (Figure #), optimal thresholds can be found which minimised the sum of square difference between 100% sensitivity and specificity between difference tumour types. When considering a threshold between MB and EP, it can be found that 0.87 x10-3 mm2/s provides an optimum sensitivity and specificity of 84% and 82%. Likewise, an optimum threshold for distinguishing between MB and PA was found to be 1.01 x10-3 mm2/s, providing a sensitivity and specificity of 97% and 95%. Finally, MB can be distinguished from non-MB (EP and PA) using a threshold on 0.94 x10-3 mm2/s, providing a sensitivity of 93% and specificity of 90%.