Deep neural network for pancreas segmentation from CT images

Abstract

Automatic pancreas segmentation from Computed Tomography (CT) images is a prerequisite of clinical practices such as cancer detection, yet challenging due to the variability in shape. To address this challenge, we propose a Hierarchical Convolutional Neural Network (H-CNN) to fuse multi-scale features, which could remedy the lost image details in progressive convolutional and pooling layers. In our proposed H-CNN, a hierarchical fusion block is designed to fuse low-level and high-level features across different layers. The H-CNN is evaluated on NIH pancreas dataset and outperforms the current state-of-art methods by achieving 86.59% ± 4.33% in terms of DSC. The experimental results confirm the effectiveness of the proposed H-CNN.