The role of diffusion tensor imaging metrics in machine learning-based characterisation of paediatric brain tumors and their practicality for multicentre clinical assessment

Abstract

Aims

Magnetic resonance imaging (MRI) is a valuable tool for non-invasive diagnosis of paediatric brain tumours. The rarity of the disease dictates multi-centre studies and imaging biomarkers that are robust to protocol variability. We investigated diffusion tensor MRI (DT-MRI), combined with machine learning, as an aid to diagnosis and evaluated the robustness of the imaging metrics.

Method

A multi-centre cohort of 52 clinical DT-MRI scans (20 medulloblastomas (MB), 21 pilocytic astrocytomas (PA), 11 ependymomas (EP)) were analysed retrospectively. Histograms for regions of solid tumour for fractional anisotropy (FA), mean diffusivity (MD), pure anisotropic diffusion (q) and pure isotropic diffusion (p) were compared to assess diagnostic capability. Linear discriminate analysis (LDA) was used for classification and validated using leave-one-out-cross-validation (LOOCV).

Results

Histogram medians for FA, MD, q and p were all different between tumor groups (P<.0001, Kruskal Wallis test). Median MD, p and q values were highest in PA, then EP and lowest in MB (P<.0001, Pairwise Wilcox test). FA median was higher for EP than PA (P=.004) with no significant difference between EP and MB (P=.591). ROC analysis showed that median MD, q and p perform best as a diagnostic marker (AUC= 0.92 to 0.99). LOOCV showed an overall accuracy of the LDA classification, ranging between 67% - 87%. FA values were highly dependent on protocol parameters, whereas pure anisotropic diffusion, q, was not.

Conclusion

DT-MRI metrics from multi-centre acquisitions can classify paediatric brain tumours. FA is the least robust metric to protocol variability and q provides the most robust quantification of anisotropic behaviour.