面向肝肿瘤高精度自动语义分割的深度建模及模型跨域迁移算法研究

Automatic semantic segmentation, including detection and segmentation, of hepatic lesions from three dimensional medical images is a prerequisite for computer aided diagnosis, planning of treatment strategies, and monitoring of disease progression, providing valuable information for medical applications. However, it is a challenging task for a number of reasons. The ambiguous boundaries of tumors, the clustered backgrounds and the heterogeneous appearance of lesions including the large variability in location, size, shape and frequency, make it difficult to devise effective and automatic segmentation algorithms. So currently liver tumors are mainly segmented by manual delineation and semi-automatic methods which require additional initialization and supervision during segmentation. Deep Convolutional Neural Network (CNN) is a typical end-to-end learning method with strong ability to learn discriminative feature representation from images. However, the CNN method for segmentation is still limited for fine grained segmentation, which is usually relieved by post-processing with traditional segmentation methods. Moreover, training a deep CNN requires a large number of fine quality segmentation annotations, which is hard to collect for medical images. The performance of CNN models often degrades when they are applied on new data from different domains, such as medical images with different imaging protocols or from different modalities and scanning periods. Manually annotating new data from each test domain is not a feasible solution. To address these problems and limitations, we propose to investigate the accurate semantic tumor segmentation from clustered backgrounds, the semi-/weakly supervised tumor segmentation and the cross domain transfer learning of segmentation models under the framework of deep CNN. Based on this research project, we will design new network architectures for the targets, that is, the automatic semantic segmentation of the liver tumor with high accuracy, the effective method for transferrable knowledge learning, a validation system for computer aided liver cancer diagnosis. As a result, we can apply our model on diverse domains with little annotation efforts, paving the way for intelligent diagnosis of liver cancers.

三维医学图像中肝肿瘤的自动定位和分割是计算机辅助诊断、手术规划和监测的重要前提，具有重要应用价值，但也是具有挑战性的问题。由于图像中肿瘤的数量和分布位置不定、边界不清晰、背景复杂等难点，目前主要以手工和半自动分割为主。深度神经网络是一种具有强大图像特征学习能力的端到端机器学习方法，但当前深度分割模型的局限具有：1）分割精度不高，需要传统方法做后处理；2）模型训练需要大量人工标注的数据。医学图像多来自不同数据域(不同的扫描中心、扫描期以及模态)且数据标注代价昂贵，因此标注每个数据域并不现实。针对以上难点和瓶颈，本项目研究在复杂背景、弱监督/半监督以及跨数据域迁移等现实场景要求下基于深度神经网络模型的高精度自动语义分割。通过本项目研究实现肝脏肿瘤的高精度自动识别和分割，建立深度神经网络模型的跨数据域迁移机制，提高模型在多样数据域的适用性，为肝癌的智能临床诊断提供可靠的技术手段和验证评价体系。

实现三维高精度语义分割是计算机辅助诊断、手术规划和监测的重要前提，具有重要应用价值。但不规则多目标（例如肿瘤等）的分割是具有挑战性的问题：临床图像数据往往比较复杂、肿瘤形态和表观特征复杂、不同数据集的分布差异较大、缺少足够的数据标注。临床应用对分割模型的精度和计算复杂度均有较高要求。针对本项的关键科学问题，本项目：1）建立了新颖的多尺度多模态多视角特征融合方法，提出了新颖的轻量化多视角学习框架，不仅有效提升了分割精度，还实现了模型的大幅度轻量化；2）针对不规则多目标分割问题引入了新颖的弱先验，提出了融合实例尺寸和形状、中心线等弱先验知识的分割模型，引入了新颖的实例感知分割损失，引入了软标签分割，有效提升了对小目标（例如小肿瘤）和形状不规则目标的分割效果；3）针对三维数据分割问题提出了新颖的伪三维模型，利用轻量的二维运算和一维运算实现高性能三维分割的同时极大地降低了模型参数量和计算复杂度；4）引入了层间注意力模块，提出了新颖的与三维模型性能相当的二维分割方法，大幅度提升了二维模型的性能；5）提出了基于多任务多空间（形状空间、表观空间、语义特征空间）的无监督域迁移模型，提出了高性能的基于稀疏点监督的域迁移分割模型；6）提出了新颖的级联结构上下文随机森林学习模型，提出了新颖的上下文特征；7）对少监督分割方法进行了系统综述和分析，参与撰写了国内首本数理医学专著；8）针对多模态特征融合和筛选问题，提出了新颖的结构稀疏多核学习诊断模型，引入了新颖的结构稀疏范数，并提出了有效的求解算法；9）针对辅助肝脏手术，进行了临床应用，并获得了教育部高校科技进步二等奖。项目在研期间，在领域重要SCI/EI期刊和会议上发表论文15篇，其中SCI论文9篇、EI期刊和会议论文5篇；授权发明专利2项，申请发明专利3项；软件著作权1项；培养研究生5人获得硕士学位；新任一本SCI期刊编委。