Robust Liver Segmentation Using Boundary Preserving Dual Attention Network

Abstract

AbstractAccurate liver segmentation is essential for the diagnosis and treatment planning of liver diseases. Recently, deep-learning-based segmentation methods have achieved state-of-the-art performance in the medical image domain. However, due to low soft tissue contrast between the liver and its surrounding organs, most existing segmentation methods are difficult to capture the boundary information of the liver well. Considering boundary information is necessary for liver edge localization in liver segmentation, we propose a novel deep-learning-based segmentation network, called Boundary Preserving Dual Attention network (BPDA). In the proposed BPDA, the boundary is being concerned with the dual attention module which consists of the spatial attention module and the semantic attention module at each layer of U-Net with the feature maps of current and its neighboring layer as inputs. The spatial attention module enhances the perception of target boundaries by focusing on salient intensity changes. Meanwhile, the semantic attention module highlights the contribution of different filters in semantic feature extraction by weighting important channels. Moreover, to overcome the phenomena that the segmentation of left lobe performance bad especially in new datasets from different sources in practical use, we analyze the attribution of the liver, we devise a data augmentation strategy to expand the data of the left lobe of the liver under the guidance of professional radiologists. The comparative experiments have done on the self-build clinic liver dataset, which includes 156 clinic case images with Iterative Decomposition of Water and Fat with Echo Asymmetry and Least-squares Estimation Magnetic Resonance Imaging(IDEAL-IQ MRI). Experiments show that our proposed method outperforms other state-of-the-art segmentation methods.KeywordsLiver segmentationAttention mechanismBoundary preservingGuided augmentation