Abstract:

The segmentation of liver and liver tumor from the image is one of the important means of liver disease diagnosis. The existing methods based on Convolutional Neural Network（CNN） achieve the segmentation of liver and liver tumor by assigning category labels to each pixel in the input image. In the process of classifying each pixel, CNN fails to use other pixel category information in the neighborhood, which is prone to suffer from missing detection of small objects and fuzzy segmentation boundaries. To address these issues, a conditional energy-based GAN is proposed for liver and liver tumor segmentation. This method, based on Energy-Based Generative Adversarial Network（EBGAN） and Conditional Generative Adversarial Network（CGAN）, uses a CNN-based segmentation network as a generator and an auto-encoder as a discriminator. The discriminator is used as a loss function to measure and improve the similarity between segmentation result and ground truth. In the process of adversarial training, the discriminator takes the segmentation output by the generator as the input and the raw image as the conditional constraint, improves the segmentation accuracy by learning the higher-order consistency between pixel categories, and uses the energy function as the discriminator to avoid vanishing gradient or exploding gradient, so it is tractable to train. The proposed method is evaluated on the dataset MICCAI 2017 Liver Tumor Segmentation（LiTS） Challenge and 3DIRCADb dataset. The experimental results show that, the proposed method can not only extract liver and liver tumor automatically, but also make use of the higher-order consistency between pixel categories to improve the segmentation accuracy of tumor and liver boundary. Furthermore, the missing detection of small tumors is reduced.

Key words: liver segmentation, liver tumor segmentation, Unet; , Energy-Based Generative Adversarial Network（EBGAN）

摘要：

从图像中分割出肝脏和肝肿瘤是肝部疾病诊断重要手段之一，现有基于卷积神经网络（Convolutional Neural Network，CNN）方法通过为输入图像中每个像素分配类别标签来实现肝脏和肝肿瘤分割。CNN在对每个像素分类过程中没有使用邻域内其他像素类别信息，容易出现小目标漏检和目标边界分割模糊问题。针对这些问题，提出了条件能量对抗网络用于肝脏和肝肿瘤分割。该方法基于能量生成对抗网络（Energy-Based Generative Adversarial Network，EBGAN）和条件生成对抗网络（Conditional Generative Adversarial Network，CGAN），使用一个基于CNN的分割网络作为生成器与一个自编码器作为判别器，通过将判别器作为一种损失函数来度量并提升分割结果与真实标注之间的相似度。在对抗训练过程中，判别器将生成器输出的分割结果作为输入并将原始图像作为条件约束，通过学习像素类别之间的高阶一致性提高分割精度，使用能量函数作为判别器避免了对抗网络训练中容易出现的梯度消失或梯度爆炸，更易于训练。在MICCAI 2017肝肿瘤分割（LiTS）挑战赛的数据集和3DIRCADb数据集上对提出的方法进行验证，实验结果表明，该方法不仅实现了肝脏与肝肿瘤的自动分割，还利用像素类别之间的高阶一致性提升了肿瘤和肝脏边界的分割精度，减少了小体积肿瘤的漏检。

关键词: 肝脏分割, 肝肿瘤分割, Unet, 能量生成对抗网络（EBGAN）