FPF-Net: feature propagation and fusion based on attention mechanism for pancreas segmentation

Automatic organ segmentation is a prerequisite step for computer-assisted diagnosis (CAD) in clinical application, which can assist in diabetes inspection, organic cancer diagnosis, surgical planning, etc. However, segmenting tiny organs like the pancreas is very challenging. Despite the success of convolutional neural networks (CNN) in automatic pancreas segmentation, the loss of the shape features impedes progress in clinical applications. Therefore, a novel pancreas segmentation network is proposed to extract features in a propagation and fusion manner, named FPF-Net. Firstly, the low-level features and high-level features are combined progressively to preserve and propagate the shape features of the pancreas. Secondly, instead of context-unaware addition or concatenation, we adopt attentional feature fusion (AFF) to alleviate the problems caused by the shape diversity and small size of the pancreas. Finally, a module consisting of Coordinate and multi-scale spatial attention (CMSA) is designed to exploit long-range dependencies and multi-scale spatial features. This module is used to extract salient information for pancreas segmentation. Experimental results validated on two pancreas datasets and a spleen dataset justify the superiority and generalization ability of our method and guarantee the reliability of our approach in clinical application.