毕业生知识库

语义分割是生物医学图像处理领域中的一项关键的技术。该技术的主要目 的是区分图像中的目标前景与背景，从而精准地提取出特定语义区域。通过为医 生提供对组织区域，如细胞、血管等的精确分割结果，语义分割极大地推动了针 对病变组织结构的研究。作为生物医学图像分析的基石，生物医学图像的语义分 割融合了生物医学影像、数据建模、图像处理以及人工智能等多个学科的技术， 这些技术的协同发展推动着生物医学图像分析领域的不断进步。 近年来，生物医学图像的语义分割为疾病检测等下游任务提供了坚实的基 础，是生物医学图像分析中极为重要的一环。目前，研究人员致力于开发和完善 各种生物医学图像的语义分割技术，但面临以下技术挑战：1.数据分布的差异： 不同的生物或医疗中心采用的扫描设备在数据互通上存在壁垒，不同生物医学 数据集的数据分布也存在偏差；2.缺乏大规模数据集标注：生物医学图像标注的 专业性较强，通常依赖于专家进行标注，人工标注成本高；3.源域数据的缺失： 传统的无监督领域自适应可应用的数据包括有标注的源域数据和无标注的目标 域数据，但在实际应用场景中，由于隐私或者存储空间等限制，可能无法获得其 他医疗或采集中心的数据；4.鲁棒性迁移学习的缺失：对抗性样本是针对模型精 心设计的，因此对模型的性能会造成较大的影响。鲁棒性迁移学习在自然图像中 已有一定的研究，但是在生物医学图像领域的研究是非常有限的。 为解决上述挑战性问题，本文提出了新的面向图像分割的领域自适应架构， 包括对抗性学习模块、图像翻译模块、创新的伪边界模块、图像增强模块以及伪 标签策略等组成部分。实验证明，该架构有效地解决了上述场景中生物医学图像 分割的领域自适应问题，并在公开的生物医学图像分割数据集上取得了优于其 他方法的性能表现。 本文主要工作和创新点包括： (1) 为解决领域分布差异和缺乏标记数据的问题，本文在传统的无监督领域 自适应的设定下，针对癌细胞跨域分割的问题，提出了基于图像翻译和对抗性学 习的无监督领域自适应算法来解决细胞形态、大小多样等问题，在可以使用有标 注的源域数据的场景中提高了模型在无标注的目标域上的性能。 (2) 为解决无源域数据和鲁棒性迁移学习的问题，本文在目标域自适应阶段 无法使用有标注的源域数据的场景中，针对视网膜视杯视盘的图像分割提出了 基于伪标签和伪边界的无源域的无监督领域自适应算法，解决了由于隐私或者 存储空间导致的源域数据缺失的问题。本文提出的领域自适应算法，在源域模型 训练阶段，采用了对抗性样本增强的策略，旨在提升源域模型在不同数据分布上 的泛化能力。然后本文利用传统的数据增强技术，生成标准源域模型所对应的伪 标签和伪边界。然后，本文将经过上述步骤训练的鲁棒源域模型迁移到目标域数 据上，以得到鲁棒的目标域模型。这一迁移过程充分利用了源域模型的泛化性能，并将其有效地应用于目标域，从而提升了目标域上的性能。最后，为了进一 步发挥目标域数据的作用，本文引入了自监督学习的机制。通过自监督学习，模 型能够从目标域数据中提取有用的信息，进一步提升模型的泛化表现。

Semantic segmentation is a key technology in biomedical image processing. The main function of this technology is to distinguish between the foreground, i.e., image objects, and the background in the image, thereby accurately extracting specific areas. Byproviding doctors with accurate segmentation results of tissue regions such as cells, blood vessels, semantic segmentation has greatly promoted research on the structure of diseased tissues. As the cornerstone of biomedical image analysis, semantic segmen tation of biomedical images integrates technologies from multiple disciplines such as biomedical imaging, data modeling, image processing, and artificial intelligence. The collaborative development of these technologies is driving continuous progress in the f ield of biomedical image analysis. Recently, the semantic segmentation technology of biomedical images has pro vided a solid foundation for downstream tasks targeting specific disease detection and is an extremely important step in biomedical image analysis. At present, researchers are committed to developing and improving various biomedical image semantic seg mentation techniques based on traditional image processing methods and deep learning methods, but face the following technical challenges: 1. Differences in data distribution: Due to barriers in data exchange between different biological or medical centers using scanning equipment, different biomedi cal datasets also exhibit significant shifts in data distribution. 2. Lack of large-scale dataset annotation: Biomedical image annotation is highly specialized and usually re lies on experts for annotations, resulting in high costs of manual annotations; 3. Lack of source domain data: Traditional unsupervised domain adaptation methods can use both labelled source data and unlabdlled target data, however, in real applications, due to limitations such as privacy or storage space, it may not be possible to obtain data from other medical centers or data collection centers; 4. Lack of robust transfer learn ing methods: Adversarial samples are designed elaborately for the model, adversarial attacks can have a significant impact on the performance of the model. Robust transfer learning has been studied in image classification to some extent, but research in the field of biomedical image segmentation is highly limited. Toovercometheaforementionedchallenges,thispaperproposesnewdomainadap tive architectures based on image segmentation, including an adversarial learning mod ule, image translation module, innovative pseudo boundary module, image enhance ment module, and pseudo label module. Experimental results have shown that this ar chitecture effectively solves the domain adaptation problem of biomedical image seg mentation in the aforementioned scenarios, and achieves better performance than other methods on KIRC datasets and RIM-ONE datasets. The main contributions and innovations of this thesis include: (1) To address the differences in domain distributions and the lack of labeled data, this paper proposes an unsupervised domain adaptation algorithm based on imagetrans lation and adversarial learning to address the problem of cross-domain segmentation of cancer cells in a scenario where annotated source domain data can be used during the target domain adaptation stage, under the traditional unsupervised domain adaptation setting. This improves the segmentation performance of the model on unlabeled target domains by addressing the diversity of cell morphology and size. (2) Toaddress the unavailability of source domain data and the lack of robust trans fer learning methods, this paper proposes an unsupervised domain adaptation algorithm based on pseudo labels and pseudo boundaries for image segmentation of optic cup and optic disc in a more challenging setting, where annotated source domain data cannot be used during the target domain adaptation stage. This algorithm solves the problem of missing source domain data due to privacy or storage space. During the training phase of the source domain model, this thesis adopts an adversarial sample augmenta tion strategy aimed at improving the generalization ability of the source domain model on different data distributions. Then this thesis uses traditional data augmentation tech niques to generate pseudo labels and pseudo boundaries corresponding to the standard source domain model. Secondly, this thesis transfers the robust source domain model trained through the above steps to the target domain data to obtain a robust target domain model. This transfer process fully utilizes the generalization performance of the source domain model and effectively applies it to the target domain, thereby improving performance on the target domain. Finally, to further leverage the role of target domain data, this thesis employs a mechanism of self-supervised learning. Through self-supervised learning, the model can extract useful information from the target domain data, further improving the generalization performance of the model.