信息工程与人工智能学院知识库

  人工智能的快速发展使得计算机辅助诊断治疗在医学领域实现广泛使用，各种医学图像分割技术也在医学研究中显现出来并扮演着重要的角色。近年来，相关学者们对卷积神经网络特征提取功能展开了坚持不懈的探索和挖掘，医学图像分割任务也通过利用多种深度学习技术实现了自身的更新和优化，在具体实践中取得了比传统图像分割法更好的结果。但是，由于医学图像中人体组织的位置、大小、形状和纹理复杂，对比度不清晰，正负样本分布不均衡等许多问题层出不穷，单纯的全卷积神经网络（Fully Convolutional Network, FCN）网络结构模型在一些医学图像分割任务中并没有达到实际临床应用的效果。

  为了让计算机获得更准确的肝脏及肝脏肿瘤分割结果，提出了两种深度学习分割方法。具体方法如下：（1）提出了一种基于双分支网络（DB-Net）结构的肝脏分割方法和下采样过程中的特征融合方式。首先为了缓解肝脏分割的过分割问题，利用了CT图像中肝脏的边界信息，提出了一个边界和区域的双分支网络，其中边界提取分支的主要作用是约束优化区域分割分支的效果。之后将其在LiTS 2017肝脏分割数据集上进行实验，该方法的分割结果显示肝脏分割的边界有所改善，指标效果也得到了提升；此外，还将目前分类性能较好的ExFuse网络中的特征融合方式迁移到了腹部CT图像的肝脏分割任务当中，在网络侧监督的同时融入更多语义信息，来提高网络特征提取能力，易于网络识别更有用的信息；并且引入并改进交叉熵损失函数解决医学图像分割中背景区域远大于目标区域导致的模型训练困难的问题。（2）提出一种基于多任务学习网络的肝脏肿瘤分割方法并使用了一种新的损失函数。该方法通过分别提取肿瘤区域和非肿瘤区域的多任务学习方法来提高网络学习性能，并使用了一种有效的损失函数，该损失函数集合了交叉熵损失、JS散度、互斥损失，解决了样本分布不均，并且增强了这两个任务中的互补特征信息，也减少了肿瘤分割的收缩问题。所提出的方法不仅降低了网络的参数数量和训练速度，并且在LiTS 2017数据集上取得了较好的分割结果。提出的这两种模型对肝脏和肝脏肿瘤分割的形态变化具有鲁棒性，也一定程度上提高了肿瘤分割的精度和性能。

  With the quick advancement of artificial intelligence, computer -aided determination and treatment are broadly utilized within the medical field. Different medical image segmentation technologies show up and play an imperative part in medical research. In recent years, relevant scholars have carried out unremitting exploration and mining on the feature extraction function of convolutional neural network. The medical image segmentation task has also achieved its own update and optimization by using a variety of deep learning technologies, and achieved better results than the traditional image segmentation method in practice. In any case, due to the complexity of the location, size, shape and texture of human tissue in medical images, blurring contrast, uneven distribution of positive and negative samples and numerous other issues emerge in perpetually, the unadulterated Full Convolutional Network (FCN) structure model show has not accomplished the real clinical practice impact in a few medical image segmentation tasks.

  In order to get more accurate segmentation results of liver and liver tumor, two deep learning segmentation methods are proposed. The specific methods are as follows: (1) A new method of liver segmentation based on DB-Net structure and feature fusion in the process of down sampling are proposed. Firstly, in order to alleviate the over segmentation problem of liver segmentation, a double branch network of boundary and region is proposed by using the boundary information of liver in CT image. The main function of boundary extraction branch is to constrain and optimize the effect of region segmentation. After that, the experiment is carried out on the liver segmentation dataset of LiTS 2017. The segmentation results show that the boundary of liver segmentation is improved, and the index effect is also improved; In addition, the feature fusion method of ExFuse network, which has better classification performance at present, is transferred to the liver segmentation task of abdominal CT image. More semantic information is integrated into the network side supervision to improve the network feature extraction ability and facilitate the network to identify more useful information; The cross entropy loss function is introduced and improved to solve the problem that the background region is much larger than the target region in medical image segmentation, which makes the model training difficult.  (2) In this paper, a method of liver tumor segmentation based on multi task learning network is proposed, and a new loss function is used. This method improves the network learning performance by extracting tumor regions and non-tumor regions, and an effective loss function is used, which combines cross entropy loss, JS divergence and mutual exclusion loss to solve the uneven distribution of samples, enhance the complementary feature information in the two tasks, and reduce the shrinkage problem of tumor segmentation. The proposed method not only greatly reduces the number of network parameters and training speed, but also achieves good segmentation results on the LiTS 2017 dataset. These two models are robust to the morphological changes of liver and liver tumor segmentation, and improve the accuracy and performance of tumor segmentation to a certain extent.