Automatic segmentation of pathological region (tumour and oedema) in high grade glioma multi-sequence MR images through voted prediction from pixel level feature sets

Abstract

Automatic region segmentation of brain from the neuroimages is an active research area in the medical domain. Currently, different kinds of magnetic resonance imaging acquisition are performed such that each technique highlights a specific region in the brain making multi-sequence images a better candidate for investigation when compared to single sequence. The abnormal regions (tumour and oedema) in glioma images are segmented through hybrid technology involving preprocessing, feature extraction and classification. The extracted features are grouped and random forest procedure is applied on each set and the prediction is obtained that minimises the randomisation. The final prediction of a pixel is obtained by aggregation of individual predictions from feature set through maximum voting which increases the ensembling and improves the outcome appreciably. The average dice coefficient of tumour and oedema segmentation is 0.96 and 0.94 respectively with three-fold cross validation. The results show significant improvement when compared to earlier methodologies.