A Hierarchical Ensemble Classifier for Multilabel Segmentation of Fat-Water MR Images

Abstract

Automatic segmentation of organs on fat-water magnetic resonance (MR) images not only enables an analysis of their morphological characteristics but also their tissues pathogenesis demonstrated by their fat fraction ratios. So far, only a few methods have been designed based on these images and all proposed segmentation algorithms have only addressed one organ at a time. In this paper, we propose a hierarchical deformation-/registration-free algorithm for multilabel segmentation of fat-water MR images without need to prior localizations or geometry estimations. This method involved a hierarchical random forest classifier and a hierarchical conditional random field (CRF) encoding a multi-resolution image pyramid. This pyramid was formed by extracting multiscale local and contextual features from image patches at different resolutions. The classifier used penalized multivariate linear discriminants and SMOTEBagging to mitigate limited and imbalanced training data. The CRF refined the segmentations with regard to the spatial and hierarchical consistencies of the labels by using layer-specific significant features identified over the trained random forest classifier. Also, we incorporated resolution-specific hyperparameters to handle variable numbers or class mixtures of the image patches over hierarchical structures. This method was trained and evaluated for segmenting 10 thoracic and 5 lumbar VBs and IVDs on 30 training and 30 test volumetric fat-water (2 channel) MR images. Objective evaluations revealed its comparable accuracy to the state-of-the-art while demanding less computational burden.