毕业生知识库

随着信息技术的飞速发展，机器学习方法在医学图像分析领域展现了巨大的潜力。大多数传统机器学习方法的优异表现依赖于训练数据和测试数据服从相同分布的基本假设。然而，在实际场景中，由于图像数据采集方式、采集环境的不同，训练数据（源域）和测试数据（目标域）往往难以满足上述假设。此外，图像数据标注费时、费力、成本高昂，导致实际场景中较难获取大量带标签的样本。域自适应学习以源域和目标域的内在相似性为桥梁，旨在不同分布的两个域之间实现知识迁移和复用。域自适应学习提升了机器学习模型的泛化能力，也能缓解当今“大数据”时代的标注稀缺问题，具备显著的理论和应用研究价值。近年来，学术界提出了许多面向图像的域自适应学习方法。这些方法使目标域的分类性能大幅提高的同时，仍然暴露出三个亟待解决的问题：欠适配、欠拟合和负迁移。针对这三个问题，本文研究相应的解决方法，并在此基础上选择肺炎影像分类、乳腺癌病理图像分类和皮肤癌图像分类三个典型的医学图像分类任务，进行应用验证。本文的主要研究工作与创新成果如下：

1. 针对现有分布差异度量在匹配源域和目标域时存在的欠适配问题，提出了一种源域与目标域类簇结构的类别质心匹配 CMMS(class Centroid Matching and local Manifold Self-learning）方法。首先， CMMS 研究了类别质心匹配策略，匹配更为复杂的分布模式，缓解欠适配问题；其次，引入源域判别结构保留策略和目标域局部流形自学习策略来分别提高源域和目标域的类簇质量，提升类别质心匹配的效率；接着，设计了一种迭代优化求解算法，并证明了该算法的收敛性；最后，提出了一种简洁的半监督扩展方案，使得优化求解算法可直接适用于 CMMS 的半监督扩展。在六个公开图像数据集上进行了实验，结果验证了 CMMS 在缓解欠适配问题方面的有效性，以及在无监督和半监督域自适应学习中的优异性。

2. 针对直接使用常规分类器预测目标域样本伪标签时引发的欠拟合问题，提出了一种自适应判别图学习的跨域标签传播 CDGS（Cross-Domain label propa- gation with discriminative Graph Self-learning）方法。首先，构建了一个统一的学习框架，联合优化域不变特征学习、邻接矩阵构建与跨域标签传播，实现邻接矩阵质量提升，从而有效地缓解欠拟合问题；其次，引入了一种自适应判别图学习策略，不仅可以捕捉跨域样本的内在相似性，而且还能充分利用源域样本真实标签和目标域样本伪标签中含有的判别信息；最后，对 CDGS 设计了一种迭代优化求解算法，可直接适用于半监督域自适应学习。在六个公开图像数据集上的实验表明： CDGS 能有效地缓解欠拟合问题，并且在无监督和半监督域自适应学习中均显著优于对比方法。

3. 针对源域和目标域的标签分布失配时出现的负迁移问题，提出了两种标签分布匹配的域自适应学习方法。其一，针对两个域的标签种类一致但同一类别的样本比例差异较大的情况，提出了一种两阶段域自适应学习 TSDA（Two Stage Domain Adaptation）方法。通过边缘分布适配和自适应分类器学习两个阶段，分别适配边缘分布和后验分布，从而实现联合分布适配，缓解负迁移问题；其二，针对源域类别完全覆盖目标域类别的情况，提出了一种共享类别渐进目标域样本学习 PLSC（Progressive target sample Learning of Shared Classes）方法。首先，设计了自适应阈值的共享类识别策略，缓解源域私有类导致的负迁移问题，对于目标域样本，引入自步学习机制，提升子空间学习的质量，促进更准确的共享类识别。在五个公开图像数据集上的实验表明：所提方法能有效地缓解标签分布失配时出现的负迁移问题。

4. 选择肺炎影像分类、乳腺癌病理图像分类和皮肤癌图像分类三个典型的医学图像分类任务，进行了上述四种方法的应用验证。对于每一个任务，构建了跨中心医学图像数据集，用于客观评测域自适应学习方法的有效性；精心设计了相应的域自适应学习任务，选择合适且公平的评价指标，探究本文方法在跨中心医学图像数据集间的迁移效果；将本文所提方法与当前相关的先进方法进行了对比。实验结果验证了所提方法在三个典型任务上的有效性和优异性，显示了所提方法在实际医学图像分类中的显著应用价值。

With the rapid development of information technology, machine learning (ML) methods have presented great potential in medical image analysis. The superior performance of most traditional ML methods relies on the basic assumption that training data and test data follow the same distribution. However, in practical scenarios, training data (source domain) and test data (target domain) are often difficult to satisfy the above assumption as a result of the differences of ways and environments for image data collecting. Besides, image data annotating is time-consuming, laborious and costly, which makes it difficult to obtain a large number of labeled samples in practical scenarios. Domain adaptation (DA) takes the internal similarity of source domain and target domain as a bridge, aiming to realize the knowledge transfer and reuse between two domains. DA can improve the generalization ability of ML methods, and can also alleviate the annotation scarcity problem in the era of big data, which has signifcant value in both theory and practice.

In recent years, many DA learning methods for image have been proposed in academia. These methods greatly improve the classifcation performance of target domain, but are still exposed to three urgent problems: under-adaptation, underftting and negative-transfer. Aiming at the three problems, this dissertation studies the corresponding methods and on this basis, selects three typical medical image classifcation tasks, namely pneumonia image classifcation, breast cancer pathological image classifcation and skin cancer image classifcation, to conduct application verifcation. The main research works and innovation achievements of this dissertation are listed as below:

1. For addressing the under-adaptation problem existing when current distribution discrepancy metrics are utilized to align the source domain and target domain, a class centroid matching CMMS (class Centroid Matching and local Manifold Self-learning) method based on the cluster structures of source domain and target domain, is proposed. Firstly, CMMS investigates the class centroid matching strategy, such that the more complicated distribution pattern can be matched and the under-adaptation problem can be alleviated. Secondly, CMMS introduces the source discriminative structure preserving strategy and the target local manifold self-learning strategy to improve the quality of source cluster and target cluster, respectively. Next, an efcient iterative optimization algorithm is designed, and its convergence is proved. Finally, a simple extension scheme for semi-supervised DA is proposed, such that the optimization algorithm can be directly applied to the semi-supervised extension of CMMS. Experiments have been conducted on six public image datasets. The results have verified the effectiveness of CMMS to alleviate the under-adaptation problem, and the superiority in both unsupervised and semi-supervised DA learning.

2. For addressing the underfitting problem triggering when the standard classifer is used to predict the pseudo-labels of target samples, a cross-domain label propagation CDGS (Cross-Domain label propagation with discriminative Graph Self-learning) method based on adaptive discriminative graph learning is proposed. Firstly, a unifed learning framework is established, which integrates domain-invariant feature learning, affinity matrix construction and cross-domain label propagation, such that the quality of the affinity matrix can be improved and the underftting problem can be alleviated. Then, an adaptive discriminative graph learning strategy is introduced, which can not only capture the inherent similarity between cross-domain samples, but also make full use of the discriminative information contained in the ground-truth labels of source samples and the pseudo-labels of target samples. Finally, an effective iterative optimization algorithm is designed, and can be directly applied to semi-supervised DA. Experiments on six public image datasets have shown that CDGS can effectively alleviate the underftting problem, and is signifcantly superior to the comparison methods in both unsupervised and semi-supervised DA learning.

3. For addressing the negative-transfer problem appearing when the label distributions of source domain and target domain are mismatched, two DA methods based on label distribution matching are proposed. One, for the condition that the label types of the two domains are the same but the sample proportions of the same category differ greatly, a Two Stage DA (TSDA) method is proposed. The marginal distribution is aligned by the marginal distribution alignment stage, and the posterior distribution is aligned by the adaptive classifer learning stage. Thus, the joint distribution alignment can be achieved and the negative-trasfer problem can be alleviated. Second, for the condition that the labels of source domain completely cover the labels of target domain, a Progressive target sample Learning of Shared Classes (PLSC) method is proposed. Firstly, a shared classes identifcation strategy based on adaptive threshold is designed, which can alleviate the negative-transfer problem caused by the source-private classes. The self-paced learning mechanism is introduced for target samples to improve the quality of subspace learning, which helps to identify shared classes more accurately. Experiments on five public image datasets have shown that the proposed methods can effectively alleviate the negative-transfer problem appearing when the label distributions are mismatched.

4. This dissertation selects three typical medical image classifcation tasks, namely pneumonia image classifcation, breast cancer pathological image classifcation and skin cancer image classifcation, to conduct application verifcation for the above four methods. For each task, cross-center image dataset is constructed to objectively evaluate the effectiveness of DA methods; the corresponding DA tasks are carefully designed, and appropriate and fair evaluation metrics are selected, which aims to explore the effectiveness of the proposed methods in cross-center medical image datasets; the proposed methods are compared with existing related advanced methods. The experimental results have verified the effectiveness and superiority of the proposed methods in the three typical tasks, and also have shown the signifcant application value of the proposed methods in practical medical image classifcation.