Automatic segmentation of the liver in abdominal CT images is critical for guiding liver cancer biopsies and treatment planning. Yet, automatic segmentation of CT liver images remains challenging due to the poor contrast between the liver and surrounding organs in abdominal CT images. In this paper, we propose a novel network for liver segmentation, and the network is essentially a U-shaped network with an encoder–decoder structure. Firstly, the complementary feature enhancement unit is designed in the network to mitigate the semantic gap between encoder and decoder. The complementary feature enhancement unit is based on subtraction, which enhances the complementary features between encoder and decoder. Secondly, this paper proposes a new cross attention model that no longer generates value by convolution, which reduces redundant information and enhances the contextual information of single sparse attention by encoding contextual information by \(3\times 3\) convolution. The dice score, accuracy, and precision of our network on the LiTS dataset were 95.85\(\%\), 97.19\(\%\), and 97.11\(\%\), and the dice score, accuracy, and precision on the dataset consisted of 3Dircadb and CHAOS were 93.65\(\%\), 94.38\(\%\), and 97.53\(\%\).

腹部 CT 图像中肝脏的自动分割对于指导肝癌活检和治疗计划至关重要。然而，由于腹部 CT 图像中肝脏与周围器官之间的对比度较差，CT 肝脏图像的自动分割仍然具有挑战性。在本文中，我们提出了一种新的肝脏分割网络，该网络本质上是一个具有编码器-解码器结构的 U 形网络。首先，在网络中设计了互补特征增强单元，以减轻编码器和解码器之间的语义差距。互补特征增强单元基于减法，增强编码器和解码器之间的互补特征。其次，本文提出了一种不再产生价值的新交叉注意力模型通过卷积，通过\(3\times 3\)卷积对上下文信息进行编码，减少冗余信息，增强单个稀疏注意力的上下文信息。我们的网络在 LiTS 数据集上的骰子得分、准确度和精确度分别为 95.85 \(\%\)、97.19 \(\%\)和 97.11 \(\%\)，骰子得分、准确度和精确度在由 3Dircdb 和 CHAOS 组成的数据集上，分别为 93.65 \(\%\)、 94.38 \(\%\)和 97.53 \(\%\)。