基于Lab色彩空间K均值聚类的全切片图像分割在肿瘤生物学效应空间分析中应用

Application of whole slide images segmentation in spatial analysis of tumor biological effects based on Lab color space and K-means clustering

目的探讨全切片图像分割在肿瘤生物学效应空间分布分析中的准确性和可行性。方法采用Lab色彩空间K均值聚类分别获取苏木精-伊红(HE)、TdT介导的dUTP缺口末端标记(TUNEL)和血小板内皮细胞黏附因子(PECAM-1/CD31)染色图像分割的颜色阈值关系式和亮度阈值,与基于颜色反卷积等分割方法比较,验证其分割性能[性能参数包括归一化互信息(NMI)、Kappa系数、平均交互比(mIoU)、平均精准率(mPr)、平均召回率(mRe)和平均准确度(mA)]。以光动力治疗小鼠乳腺癌为例,对来自科学数据银行(ScienceDB)的数据集,运用核密度估计热力图和分布密度的空间分析方法分别可视化和量化肿瘤生物学效应,并对坏死、凋亡和血管密度与光通量进行相关性分析。结果图像分割速率约为54.5 s/G;以基于颜色反卷积和Otsu的分割结果为金标准,HE、TUNEL和CD31染色图像分割的NMI为0.45~0.58,Kappa系数为0.60~0.80,mIoU为0.70~0.83,mPr为0.73~0.91,mRe为0.81~0.97,mA为0.94~0.96。光通量与坏死密度线性相关,其Pearson相关系数为0.88;进行线性回归分析,可得到关系式为Y=0.101 4X+22.470 0(其中:Y为坏死密度;X为光通量)。结论基于Lab色彩空间K均值聚类的方法实现了图像的精准分割,结合空间分析和统计学方法实现了肿瘤生物标记物的空间数据分析及可视化,在肿瘤图谱生物效应分析中具有巨大的潜力。

Objective To investigate the accuracy and feasibility of whole slide image segmentation in analysis of spatial distribution of tumor biological effects. Methods The color threshold relation and brightness threshold of hematoxylin-eosin(HE), TdT-mediated dUTP nick end labeling(TUNEL) and platelet endothelial cell adhesion molecule-1(PECAM-1/CD31)stained image segmentation were obtained by K-means clustering based on Lab color space. Compared with color deconvolution based segmentation method, its segmentation performance was verified [performance parameters included normalized mutual information(NMI), Kappa coefficient, mean intersection over union(mIoU), mean precision(mPr), mean recall(mRe) and mean accuracy(mA)]. The photodynamic therapy of breast cancer in mice was taken as an example, the spatial analysis method of nuclear density estimation heat map and distribution density were used to visualize and quantify tumor biological effects on data set from Science Data Bank(ScienceDB). The correlation between necrosis, apoptosis and blood vessel density and light fluence was analyzed. Results The image segmentation rate was about 54.5 s/G;the segmentation results based on color decon volution and Otsu were set as the gold standard, the NMI of HE, TUNEL and CD31 image segmentation was 0.45-0.58, Kappa coefficient was 0.60-0.80, mIoU was 0.70-0.83, mPr was 0.73-0.91, mRe was 0.81-0.97 and mA was 0.94-0.96.The light fluence was linearly related to necrosis density, and the Pearson correlation coefficient was 0.88. By linear regression analysis, the relationship was obtained as Y = 0.101 4 X + 22.470 0(Y = necrosis density;X = light fluence). Conclusion It is demonstrated that K-means clustering based Lab color space realizes accurate image segmentation. Combined with spatial analysis and statistical methods, spatial data analysis and visualization of tumor biomarkers are realized, which has great potential in the analysis of the biological effects of tumor atlas.