肝脏精准分割对于肝癌的定位与治疗至关重要, 针对肝脏形状尺寸不一以及边缘和病灶区域分割难度较大等问题, 提出了一个基于多层感知机和多尺度特征提取的肝脏分割网络(M2U-Net)。该网络分为卷积阶段和多层感知机阶段。在卷积阶段的编码器部分加入挤压激励模块, 突出特定的肝脏分割任务, 抑制无关背景区域; 在多层感知机阶段加入标记化多层感知机模块, 减小模型复杂度。过渡层增加了多尺度特征提取模块, 适应不同尺度肝脏的分割以及细节区域的分割; 在LiTS数据集和东方肝胆医院提供的数据集上的实验结果表明, 此分割网络在三个评价指标上结果均优于U-Net、U-Net++和CE-Net等分割网络。

Precise liver segmentation is crucial for the localization and treatment of liver cancer, and in view of the problems of different liver shapes and sizes, as well as the difficulty of segmentation of edges and lesion areas, a liver segmentation network based on multilayer sensor and multi-scale feature extraction (M2U-Net) is proposed. The network is divided into convolutional phases and multilayer perceptron phases. First, an extrusion excitation module is added to the encoder portion of the convolution phase to highlight specific liver segmentation tasks and inhibit irrelevant background areas. Secondly, a tokenized multilayer perceptron module is added to the multilayer perceptron stage to reduce the complexity of the model. The transition layer adds a multi-scale feature extraction module to adapt to the segmentation of livers at different scales and the segmentation of detail areas. Finally, experimental results on the LiTS dataset and the dataset provided by Oriental Hepatobiliary Hospital show that the segmentation network is better than the segmentation networks such as U-Net, U-Net++ and CE-Net on three evaluation indicators.