Pyramid Dilated Residual Pooling Convolutional Network for whole liver segmentation

Abstract

Segmentation of the liver in Computer Tomography (CT) images allows the extraction of three-dimensional (3D) structure of the liver structure. The adequate receptive field for the segmentation of such a big organ in CT images, from the remaining neighboring organs was very successfully improved by the use of the state-of-the-art Convolutional Neural Networks (CNN) algorithms, however, certain issue still arise and are highly dependent of pre- or post- processing methods to refine the final segmentations. Here, an Encoder-Decoder Dilated Poling Convolutional Network (EDDP) is proposed, composed of an Encoder, a Dilation and a Decoder modules. The introduction of a dilation module has produced allowed the concatenation of feature maps with a richer contextual information. The hierarchical learning process of such feature maps, allows the decoder module of the model to have an improved capacity to analyze more internal pixel areas of the liver, with additional contextual information, given by different dilation convolutional layers. Experiments on the MICCAI Lits challenge dataset are described achieving segmentations with a mean Dice coefficient of 95.7%, using a total number 30 CT test volumes.