CASIA OpenIR  > 毕业生  > 硕士学位论文
显微细胞图像的鲁棒实例分割算法
周雅婷
2023-05-22
页数76
学位类型硕士
中文摘要

显微图像中单细胞实例分割是一项关键的细胞生物学技术,它能够将显微镜图像中的细胞划分到对应的实例中,获取每个细胞的位置和形态信息,进而深入、详细地了解细胞内部结构和功能。同时,作为细胞分析的基础,细胞实例分割技术广泛应用于疾病诊断医治、药物筛选开发等领域,展现出重要的应用价值。

 

细胞实例分割的准确性直接影响生物学研究中定量分析等任务的可靠性。因此,自动、高效、精确的细胞实例分割技术已成为生命科学和医学研究领域的一种有效工具。目前,研究人员致力于开发和完善各种基于传统图像处理方法和深度学习方法的细胞实例分割技术,但以下挑战性困难的存在使其难以精确分割细胞实例。

 

细胞粘连:细胞粘连是指多个细胞在显微镜下相互粘附,形成一个难以分割的连续结构。细胞边界的模糊性使得难以将其分为独立单元,产生分割错误,从而影响分割准确性。

 

弱信号:因荧光强度不同或受到目标在显微镜中的位置影响,荧光显微镜拍摄的细胞往往存在一些低强度区域。这些区域通常难以检测出来,从而降低了图像分析的准确性。

 

缺乏大规模数据集标注:生物医学图像标注的专业性较强,通常依赖专家进行,人工标注成本高。因此实际应用中往往缺乏足够数量的标注数据,使得机器学习模型的训练和性能评估受到限制。

 

细胞的形态和颜色变化复杂:在显微镜图像中,细胞的形态和颜色会因为不同的生理状态、细胞类型以及拍摄条件等而发生变化,现有的模型大多针对特定数据集设计,难以适配细胞形态和风格的变化。

 

为克服上述挑战性问题,本文提出了新的实例分割架构,包括创新的数据增强模块、语义感知模块、Transformer骨干网络以及域适应策略等组成部分。实验证明,该架构有效地解决了上述场景中的细胞实例分割问题,并在LIVECell数据集与Hek293T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:

 

为解决细胞粘连与弱信号问题,本文提出了语义感知模块以提高粘连细胞实例辨别度与弱信号区域分割完整性。该分支能识别前景、细胞内部和细胞边界的语义信息。通过将骨干网络输出与语义特征相结合,为检测和分割任务提供额外信息。前景和背景类的语义信息提高了弱信号区域的分割完整性,而边界语义信息则提高了重叠和接触细胞之间的实例区分度,使得实例分割更加精确。

 

为解决标注数据量不足问题,本文提出了一种名为“空间填充”的新型数据增强模块。该方法首先基于现有标注创建细胞数据库,然后随机挑选适量细胞实例并进行调整。最后,将这些实例叠加到训练图像的空白区域上,从而不断生成新的训练数据。该方法充分发挥现有标注实例的作用,可以大幅增强训练数据的多样性,减轻对人工标注的依赖,同时提升模型的泛化表现。

 

为解决细胞形态和颜色变化的问题,本文使用Swin Transformer主干网络,通过泛全局的注意力机制,帮助模型适应不同形状、大小的细胞。与此同时,提出域适应策略来减小不同细胞类型图像特征之间的差异。具体来说,域适应策略分为两阶段:第一阶段通过对抗网络拉近源域与目标域间的特征分布。第二阶段生成高质量的伪标签并采用伪标签对目标域图像进行像素级监督训练。通过注意力机制、对抗网络和伪标签增强训练,提高模型鲁棒性并优化模型在目标数据上的分割性能。

 

英文摘要

Single-cell instance segmentation in microscopy images is a crucial cell biology technology that enables the division of cells in microscopy images into corresponding instances, obtaining the location and shape information of each cell, and subsequently prompting for an in-depth understanding of the internal structure of cells and function. Simultaneously, as the foundation of cell analysis, cell instance segmentation technology has extensive applications in fields such as disease diagnosis and treatment, drug screening and development, demonstrating significant practical value.

 

The accuracy of cell instance segmentation directly impacts the reliability of tasks such as quantitative analysis in biological research. As a result, automatic, efficient, and precise segmentation of cell instances has become an essential tool in the field of life science and medical research. Currently, researchers are dedicated to developing and enhancing various cell instance segmentation techniques based on traditional image processing methods and deep learning approaches. However, the presence of the following challenging difficulties makes it hard to accurately segment cell instances:

 

Cell adhesion: Cell adhesion refers to multiple cells sticking together under the microscope, forming a continuous structure that is difficult to separate. The ambiguity of cell boundaries makes it challenging to divide them into independent units, leading to segmentation errors that affect segmentation accuracy.

 

Weak signal: Due to different fluorescence intensity or affected by the position of the target in the microscope, the cells imaged by fluorescence microscope often have some areas of low intensity. These areas are often difficult to detect, which reduces the accuracy of image analysis.

 

Lack of large-scale dataset labeling: Biomedical image labeling is highly specialized, usually relying on experts, and the cost of manual labeling is high. Therefore, in practical applications, there is often a lack of sufficient labeled data, which limits the training and performance evaluation of machine learning models.

 

Complex shape and color changes of cells: In microscopy images, the shape and color of cells may change due to different physiological states, cell types, and imaging conditions. Most of the existing models are designed for specific datasets, making it difficult to adapt to changes in cell shape and appearance.

 

To address the aforementioned challenging issues, this thesis proposes a novel instance segmentation architecture, which includes an innovative data augmentation module, a semantic-aware module, a Transformer backbone network, and a domain adaptation strategy, among others. Experiments demonstrate that this architecture effectively tackles the cell instance segmentation problem in the given scenarios and outperforms other methods on the LIVECell dataset and the Hek293T dataset.

 

The primary contributions and innovations of this thesis include:

To address the issue of adhesion and weak signal, this thesis proposes a semantic-aware module to improve instance discrimination and the segmentation integrity of weak signal regions. This branch can identify semantic information of the foreground, cell interior, and cell boundary. It provides additional information for detection and segmentation tasks by combining the backbone network output with semantic features. The semantic information of foreground and background classes improves the segmentation integrity of weak signal regions, while the boundary semantic information enhances instance discrimination between overlapping and contacting cells, making instance segmentation more precise.

 

To address the issue of insufficient labeled data, this thesis proposes a novel data augmentation module called "Space Filling." The method first creates a cell database based on existing annotations, then randomly selects an appropriate number of cell instances and adjusts them. Finally, these instances are superimposed onto the blank areas of the training images, continuously generating new training data. This method fully utilizes the existing labeled examples, greatly enhancing the diversity of training data, reducing reliance on manual labeling, and simultaneously improving the model's generalization performance.

 

To address the issue of cell shape and color changes, this thesis employs the Swin Transformer backbone network to help the model adapt to cells of varying shapes and sizes through a pan-global attention mechanism. Simultaneously, a domain adaptation strategy is proposed to reduce differences between image features of different cell types. Specifically, the domain adaptation strategy is divided into two stages: the first stage narrows the feature distribution between the source domain and the target domain using an adversarial network. The second stage generates high-quality pseudo-labels and applies them for pixel-level supervised training on target domain images. The model's robustness is improved and its segmentation performance on target data is optimized through attention mechanisms, adversarial networks, and pseudo-label augmentation training.

 

关键词Microscopy Cell Images Instance Segmentation Domain Adaptation Cell Adhesion Data Scarcity
语种中文
七大方向——子方向分类生物特征识别
国重实验室规划方向分类生物进化与仿生计算
是否有论文关联数据集需要存交
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/51876
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
周雅婷. 显微细胞图像的鲁棒实例分割算法[D],2023.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
显微细胞图像的鲁棒实例分割算法.pdf(25433KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[周雅婷]的文章
百度学术
百度学术中相似的文章
[周雅婷]的文章
必应学术
必应学术中相似的文章
[周雅婷]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。