Summary
MRI has shown to be a very useful tool during the diagnosis of MB, with ADC being an important indicator when attempting to differentiate between similar posterior fossa tumours. ADC has also shown to posses some sensitivity to histological subgroups of MB, however the literature on this particular issue is limited. However, no studies have attempted to classify molecular subgroups based on singular MR measures such as ADC. Although, using more image features and applying advanced analysis techniques have shown high sensitivity to subtyping Glioblastoma (references), which has prompted investigations in medulloblastoma.
Radiomics and Radiogenomics
Distinguishing medulloblastoma from other posterior fossa tumours
The earliest study combining multiple MR measures to distinguish MB from other posterior fossa tumour types was published by Schnieder et al., (2007) \cite{Schneider2007}. 17 tumours were investigated, comprising of 7 medulloblastoma, four infiltrating glioma, two ependymoma and four pilocytic astrocytoma. The average tumour ADC was combined with six normalised metabolites obtained from single-voxel spectroscopy, and linear discriminant analysis was used in order to distinguish tumour types. All tumour types were correctly classified with a -2 log-liklihood of <1x10-9. However, the use of mean ADC alone could not separate the tumour types with adequate accuracy.
A series of short publications by Fetit et al., (2013, 2014, 2015) \cite{Fetit2013,Fetit2014,Fetit2015} investigated the use of texture analysis in distinguish paediatric brain tumours. The texture analysis process used 302 texture features and used machine learning approaches with support vector networks to distinguish tumour types. The initial study \cite{Fetit2013} assess T1w and T2w images, and found that SVM and Bayes algorithms are promising for separating tumour types with classification rates of >85%. This group extended their work in 2014 \cite{Fetit2014} where 3D texture analysis was performed rather than 2D. This study found that 3D texture analysis provided increased classification compared to that of 2D, with one machine learning technique providing 100% sensitivity and specificity distinguishing all tumour types. However, the author believes that this particular analysis method could have been biased during analysis. This work was further expanded to include heterogeneous data sets obtained from MRI scans from difference vendors and field strengths. 3D texture analysis was performed, and metrics were fed into a support vector machine where leave-one-out cross validation was applied. Overall classification accuracy was found to be 70%, where sensitivity and specificity of MB and PA were >71%, however sensitivity to EP was very low (11%). These results suggest that the use of different scanners have a significant impact on the classification via machine learning, as the prior study demonstrated higher classification rates.
Gutierrez et al., (2014) \cite{Rodriguez2014} investigated 40 tumours including 17 medulloblastoma, 16 pilocytic astrocytoma and 7 ependymoma. Of the 17 MB, 14 were of classic MB and 3 were large cell/anaplastic. ADC maps were analysed, where a 34 features were extracted and used in a support vector machine approach in order to distinguish the brain tumour type. It was found that the 25th percentile, 75th percentile and histogram skewness obtained the highest classification rates for all tumours, obtained >95.8% for medulloblastoma, 96.9% for pilocytic astrocytoma and 94.3% of ependymoma.
Distinguishing the subtypes of medulloblastoma
Wefers 2014
The study presented by Gutierrez et al., (2014) \cite{Rodriguez2014} extended their work to distinguish MB subtypes. When classifying BM subtypes, the ADC textural features provided high classification rates than histogram features, where the sum average and sum variance provided classification rates of 89%. The small number of medulloblastoma limits the impact of the subtype classification rates, which is acknowledged by the author.
Aside from DWI, location of tumour origin has been investigated in order to predict the MB subgroup. Perreault et al., (2014) \cite{Perreault2014} investigated 47 medulloblastoma, which were categorised into histological subgroups. Of the histological subgroups there were 31 classic, 10 large cell / anaplastic, 4 desmoplastic and 2 others. Of the molecular classification, there were 4 WNT, 13 SHH, 13 Group 2 and 17 Group 4. In this study, a multivariate logistic regression approach was used in order to identify significant predictors of MB subgroups. A validation cohort was then assessed which contained a further 52 MB, comprising of 37 classic MB, 4 large cell/anaplastic, 11 desmoplastic and 0 others within the histological classification, and 10 WNT, 11 SHH, 12 Group 3 and 19 Group 4 within the molecular classification. Results demonstrated that location, pattern of enhancement and tumour margin were significant predictors of MB subgroups, where 69% were accurately classified.
Patay 2015
Zhao et al., (2017) \cite{Zhao2017} investigated a cohort of 60 paediatric MB consisting of 8 WNTs, 17 SHH, 15 Group 3 and 20 Group 4 tumours. It was found that SHH, WNT and Group 4 significantly correlated with localisation pattern, and Group 4 MBs also demonstrated a significant association with minimal/no enhancement. Group 3 MB were not assessed for correlations with localisation patterns and other conventional MRI features. Despite demonstrating a promising model for predicting the molecular subgroup in adults, this same strategy was not applied for predicting paediatric cohort. Aside from a few differences between the SHH MBs, the localisation patterns were found to be similar between the paediatric and adult cohorts in the other subgroups.
In 2017, Fetit et al., (2017) \cite{Fetit2018} investigated 134 paediatric brain tumours, where 45 were MB, 71 were pilocytic astrocytoma and 18 were ependymoma. This study employed the use of 3D texture analysis in order to differentiate the tumour types using support vector machines. Despite the robust classification and validation procedure, the MB sensitivity was reported at only 57%, however the specificity was 91%. The pilocytic astrocytoma and ependymoma obtained sensitivity and specificity >83%. The overall classification accuracy was reported at 77%.
Dasgupta et al., (2018) \cite{Dasgupta2018} assessed 111 medulloblastoma, which comprised of 17 WNT, 44 SHH, 27 Group 3 and 23 Group 4. Each tumour was qualitatively assessed using 19 different features describing characteristics such as tumour size, location and MR intensities. 76 cases were used as a training data set for multinomial logistic regression in order to create a model which could predict the molecular subgroup of a validation cohort. Validation demonstrated that accurate classification was achieved in 71% of cases; SHH and Group 4 achieved high classification accuracy (95% and 78%), but Group 3 and WNT achieved low classification accuracy (56% and 41%).
Summary