摘要
目的 免疫组化病理图像存在数据稀缺与标注困难的问题,带标签的图像生成算法为其提供了解决方法。但现有图像生成方法虽关注语义信息,却忽略颜色特征,造成标注准确性不足。为此,笔者研究提出了一种融合扩散模型与颜色增强技术的带标签图像数据增强方法(DM-color),以提升生成病理图像的质量,并提高标注的准确度。方法 实验数据来源实验室自建的小鼠皮肤免疫组化染色的病理图像数据集,通过ImageScope软件对全数字扫描切片图像进行读取,在20倍镜下进行高分辨率截图,获取合适大小的病理图像制作成实验用数据集,包含细胞型122幅与分泌型160幅。采用ImageJ软件对这282幅图片进行阳性细胞或区域的标注,导出标注区域的坐标数据,并用自行编写的算法将其绘制成二值图像(像素0为背景,255为病变区域)制作为标签。基于扩散模型DatasetDM框架,引入颜色增强卷积(CEC)与颜色相似性注意力(CSA)双模块构建DM-color图像生成算法。采用弗雷谢特初始距离(FID)评估图像真实性,医学专家等级评分衡量标注准确度,并结合下游图像阳性细胞/区域分割任务,以平均交并比(mIoU)为主要评价指标,综合评价真实性与准确度。结果 与DatasetDM相比,改进的DM-color模型生成的细胞型图像,FID指标从42.31优化至37.51,分泌型图像改进前后FID指标分别为42.12和42.78。标注准确性评价显示,细胞型与分泌型图像的“优秀”占比分别提升13%和11.67%,语义分割任务中细胞型图像mIoU提升了6.96%,分泌型图像提升了6.10%。结论 改进的DMcolor方法显著提升了生成免疫组化图像的准确度和真实性,有望在人工智能病理图像识别应用中提供新的方法。
Objective The aim of this study was to propose a labeled image data enhancement method(DM-color)that combines diffusion model and color enhancement technique for improving the quality of generated pathological images and accuracy of annotation,under the consideration that labeled image generation algorithm provides solution to problems of data scarcity and difficult annotation in immunohistochemical pathological images,and the existing image generation methods focus on semantic information while ignores color features,resulting in insufficient annotation accuracy.Methods The data was obtained pathological image data set of immunohistochemical staining of mouse skin self-built in the laboratory.The full-digital scanned slice images were read by ImageScope software,and high-resolution screenshots were taken under the 20-fold microscope,the pathological images of the appropriate size were obtained and made into an experimental data set,including 122 cellular and 160 secretory.ImageJ software was used to label the positive cells or regions of 282 images,and the coordinate data of the labeled region was exported.The binary image(pixel O as background,255 as lesion area)was drawn by the self-written algorithm and made as the label.Based on the DatasetDM framework of the diffusion model,the dual modules of color-enhanced convolution(CEC)and color similarity attention(CSA)were introduced to construct DM-color image generation algorithm.The Frechet inception distance(FID)was used to evaluate the image authenticity,medical expert grade score was used to measure the annotation accuracy,and combined with the downstream image positive cel/region segmentation task,the mean intersection and union ratio(mloU)was used as the main evaluation index to comprehensively evaluate the authenticity and accuracy.Results Compared with DatasetDM,the FID index of the cell-type image generated by modified DM-color model was optimized from 42.31 to 37.51,and FID indexes before and after modification of the secretory image were 42.12 and 42.78,respectively.The evaluation results of annotation accuracy showed that the excellent proportion of cellular and secretory images increased by 13%and 11.67%,respectively.In semantic segmentation task,the mloU of cellular images increased by 6.96%,and the secretory images increased by 6.10%.Conclusion It is demonstrated that the modified DM-color significantly improve the accuracy and authenticity of generated immunohistochemical images,which is expected to provide a novel approach for artificial intelligence-based pathological image recognition application.
作者
胡典
阴慧娟
HU Dian;YIN Hui-juan(Integrrative Regeneration Laboratory,Institute of Biomedical Engineering,Chinese Academy of Medical Sciences&Peking Union Medical College,Tianjin 300192,China)
出处
《生物医学工程与临床》
2025年第4期441-448,共8页
Biomedical Engineering and Clinical Medicine
基金
国家自然科学基金资助项目(62175261)。
关键词
病理图像识别
数据增强
扩散模型
语义分割
图像标注
pathological image recognition
data enhancement
diffusion model
semantic segmentation
image annotation
