【作者】 李军；

【导师】 刘萍； 邵浩靖；

【作者基本信息】 太原理工大学 ， 软件工程（专业学位）， 2024， 硕士

【摘要】 细胞分割是图像处理领域的一项重要任务,在生命科学、生物信息学和生物医学等多个领域具有广泛的应用价值。近年来,空间转录组技术的蓬勃发展为细胞分割的发展带来了全新的视角,通过细胞分割进一步阐明研究组织的细胞类型组成以及空间分布,明确细胞群体间相互作用,了解组织成分之间的分子相互作用。细胞分割作为空间转录组流程中的上游分析,成为整个流程分析中最基础、最关键的一步,而且空间转录组的基因表达量信息可以为细胞的精确分割助力。因此开发高效准确的分割软件工具也一直是空间转录组细胞分割领域的主要焦点之一。然而,每种软件工具都有其独特的特点和适应性,没有一种通用的细胞分割软件能达到完美的效果。基于上述情况,本文开展了以下研究:(1)细胞分割算法的基本信息与友好度分析。选取总共10个细胞分割算法(Cellpose、Omnipose、StarDist、DeepCell、Plantseg、Icy、CellProfiler、Ilastik、SCS和Watershed)进行功能以及软件特点阐述。并深入对比了所选的软件工具之间的基本信息和友好度。(2)常规生物细胞图像的算法对比分析。从10个细胞分割算法中选用6个细胞分割算法(Cellpose、Omnipose、StarDist、DeepCell、Plantseg和Ilastik)对DSB2018、Cellpose＿cyto和Ph C-C2DL-PSC三个公开常用的具有不同特点的常规细胞图像数据集进行测试。针对这三个含有真实标签的数据集,选取了常用的分割指标(准确率、召回率、精度和F1)进行软件评估。通过图像尺寸统一化、图像灰度化以及图像归一化的预处理操作,发现这些软件工具在DSB2018数据集上指标有较好的提升。同时也发现不同软件工具在不同数据集上对预处理操作的适应性差异很大。此外,结果还发现Cellpose和StarDist两个软件工具的整体分割性能较好,对不同的数据集适应性较强,能具备良好的分割精确性和稳定性。(3)空间转录组的算法对比分析。从10个细胞分割算法中选用不同的6个细胞分割算法(Cellpose、Omnipose、StarDist、DeepCell、SCS和Watershed)进行评估,其中SCS为基于空间转录组的细胞分割算法。通过选取两个有代表性的包含细胞染色图像数据以及相应的基因表达量测序数据的空间转录组数据集(Stereo-seq和Seq-scope)进行性能比较工作。对两种数据集进行了基因表达量热图生成、热图与细胞染色图像对齐、图像灰度化处理等预处理操作。由于所选的2个空间转录组数据集缺乏细胞分割的真实值标签,所以引入新兴的评估指标,测定2个软件工具之间的细胞分割结果的相关性以及基因表达相似性来判断分割效果的好坏。结果发现SCS、Cellpose、StarDist与DeepCell四个软件工具对两个数据集的分割结果相似,分割性能相近,且分割细微,可处理不同类型的数据集,均可以很好适应地空间转录组数据集,并获得良好的分割效果。(4)使用Icy软件设计了一个细胞分割算法流程,使用CellProfiler软件设计了一个细胞分割管道,选取Stereo-seq空间转录组数据集进行基因表达量热图生成、热图与细胞染色图像对齐、图像灰度化处理等预处理操作后评估相关性指标和P值。并与选用6个可以用于空间转录组的细胞分割算法(Cellpose、Omnipose、StarDist、DeepCell、SCS和Watershed)的比较结果表明,Icy细胞分割流程表现出最佳性能,在8个软件对比中转录相关性最高为第一,CellProfiler为第三。表明了Icy和CellProfiler在细胞分割准确性方面的优势,在处理空间转录组数据时表现突出。综合结果发现,SCS、Cellpose、StarDist、DeepCell、Icy和CellProfiler流程软件工具的整体分割较好,且分割性能较为接近,研究者们可以针对不同特点的空间转录组数据集选用适合的进行实验并加以创新,从而更好地完成空间转录组的细胞分割,助力空间转录组学研究。更多还原

【Abstract】 Cell segmentation is an important task in the field of image processing,which has a wide range of application value in many fields such as life science,bioinformatics and biomedicine.In recent years,the booming development of spatial transcriptome technology has brought a new perspective to the development of cell segmentation,through which cell segmentation further elucidates the cell type composition as well as the spatial distribution of the studied tissues,clarifies the interactions among the cell populations,and understands the molecular interactions among the components of the tissues.Cell segmentation,as the upstream analysis in the spatial transcriptome process,has become the most basic and critical step in the analysis of the whole process,and the gene expression information of the spatial transcriptome can help the accurate segmentation of cells.Therefore the development of efficient and accurate segmentation software tools has also been one of the main focuses in the field of spatial transcriptome cell segmentation.However,each software tool has its unique characteristics and adaptability,and no universal cell segmentation software can achieve perfect results.Based on the above,the following research is carried out in this paper:(1)Basic information and friendliness analysis of cell segmentation algorithms.A total of 10 cell segmentation algorithms(Cellpose,Omnipose,StarDist,DeepCell,Plantseg,Icy,CellProfiler,Ilastik,SCS,and Watershed)were selected for the elaboration of their functions as well as software features.An in-depth comparison of the basic information and friendliness between the selected software tools is also made.(2)Comparative analysis of algorithms for conventional biological cell images.Six cell segmentation algorithms(Cellpose,Omnipose,StarDist,DeepCell,Plantseg,and Ilastik)were selected from 10 cell segmentation algorithms for DSB2018,Cellpose＿cyto,and Ph C-C2DL-PSC,three publicly commonly used conventional cell image with different features datasets were tested.For these three datasets containing real labels,commonly used segmentation metrics(accuracy,recall,precision and F1)were selected for software evaluation.The pre-processing operations of image size unification,image greyscaling,and image normalisation were found to provide a better improvement in the metrics of these software tools on the DSB2018 dataset.It was also found that the adaptability of different software tools to the preprocessing operations varied greatly on different datasets.In addition,the results also found that the two software tools Cellpose and StarDist have better overall segmentation performance,are more adaptable to different datasets,and can have good segmentation accuracy and stability.(3)Comparative analysis of algorithms for spatial transcriptome.Different six cell segmentation algorithms(Cellpose,Omnipose,StarDist,DeepCell,SCS and Watershed)were selected for evaluation from 10 cell segmentation algorithms,of which SCS is a cell segmentation algorithm based on spatial transcriptome.Performance comparison work was performed by selecting two representative spatial transcriptome datasets(Stereo-seq and Seq-scope)containing cell staining image data as well as corresponding gene expression volume sequencing data.Pre-processing operations such as heatmap generation of gene expression quantities,heatmap alignment with cell staining images,and image greyscaling were performed on both datasets.Since the 2 selected spatial transcriptome datasets lacked the real value labels for cell segmentation,emerging assessment metrics were introduced to determine the correlation of cell segmentation results between the 2 software tools and the gene expression similarity to judge the segmentation effect.The results found that the four software tools SCS,Cellpose,StarDist and DeepCell have similar segmentation results for the two datasets,with similar segmentation performance,and the segmentation is subtle and can handle different types of datasets,and all of them can be well adapted to the geospatial transcriptome dataset with good segmentation results.(4)A cell segmentation algorithm process was designed using Icy,a cell segment ation pipeline was designed using CellProfiler,and the Stereo-seq spatial transcriptome dataset was selected to evaluate the correlation metrics and P-values after pre-processi ng operations such as gene expression heatmap generation,heatmap alignment with cel l staining images,and image greyscaling.And the comparison results with the selectio n of six cell segmentation algorithms that can be used for spatial transcriptome(Cellp ose,Omnipose,StarDist,DeepCell,SCS,and Watershed)showed that the Icy cell seg mentation process showed the best performance,with the highest transcriptional correla tion as the first out of the eight software comparisons,and CellProfiler as the third.It shows the superiority of Icy and CellProfiler in cell segmentation accuracy,which is outstanding when dealing with spatial transcriptome data.The comprehensive results found that SCS,Cellpose,StarDist,DeepCell,Icy and CellProfiler process software tools have better overall segmentation and closer segmentation performance,and researchers can choose the suitable ones for experiments and innovate for spatial transcriptome datasets with different characteristics,so as to better complete the spatial transcriptome cellular segmentation and help spatial transcriptomics research.更多还原