Abdominal organs segmentation based on multi-path fully convolutional network and random forests

Abstract

Fully convolutional network has predicted multiple class dense outputs in CT image labels and obtained significant improvements in segmentation tasks. In this paper, we present a joint multi-path fully convolutional network (MFCN) with random forests (RF) architecture for abdominal organs segmentation automatically. First, in coarse segmentation step, three FCNs are trained respectively with three orthogonal directions which consider contextual and spatial information of fusion layers adequately. In classification step, using features extracted from different layers of network and normalizing them to mean value as supervoxel representation to train RF. This allows the computation of supervoxel at each orientation achieve high efficiency. Finally, we aggregate the results of MFCN and RF on voxel-wise and perform conditional random fields (CRF) focuses on smoothing borders of fine segmentation regions. We exceeds the state-of-the-art methods and get achievable DSC values for our work is 90.1%, 88.4%, 88.0%, 88.6% represent liver, right and left kidney, spleen respectively.