Hierarchical Feature-learning Graph-based Segmentation of Fat-Water MR Images

In this paper, we proposed a deformation-Iregistration-free method for multilabel segmentation of fat-water MR images without need to prior localization or geometry estimation. This method employed a multiresolution (hierarchical) feature- and prior-based Random Walker graph and a hierarchical conditional random field (HCRF). To incorporate both aspatial (intra-patch) and spatial (inter-patch neighborhood) information into the image segmentation, the proposed random walker graph was made of a multiresolution spatial and a multiresolution aspatial (prior-based) sub-graph. Edge weights and prior probabilities of this graph as well as the energy terms of the HCRF were determined by a hierarchical random decision forest classifier. This classifier was trained using multiscale local and contextual features extracted from fat-water (2-channel) magnetic resonance (MR) images. The proposed method was trained and evaluated for simultaneous volumetric segmentation of vertebral bodies and intervertebral discs on fat-water MR images. These evaluations revealed its comparable accuracy to the state-of-the-art while demanding less computations and training data. The proposed method was, however, generic and extendible for segmenting any kind of tissues on other multichannel images.