多域染色风格转换及其在肾小球检测中的应用  

Multi-domain staining style transfer and its application in glomerulus detection

作　　者：苗一民 吴永飞 MIAO Yimin;WU Yongfei(College of Data Science,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学大数据学院,晋中030600

出　　处：《上海理工大学学报》2024年第6期677-685,共9页Journal of University of Shanghai For Science and Technology

基　　金：国家自然科学基金资助项目(61901292)。

摘　　要：临床实践中由于组织学染色耗时费力且具有不可逆性,导致肾脏病理学图像的数量相对较少,从而限制了医疗诊断和深度学习方法的应用。为此,提出了一种基于生成对抗网络的深度学习模型,通过对模型进行单次训练,实现了不同染色之间的风格转换。随后,将多域染色风格转换模型引入肾小球检测流程。实验通过染色转换模型进行染色转换,利用不同风格染色特征之间的相互补充,提高肾小球检测模型的准确性和泛化性。实验结果表明,多域染色风格转换模型生成的图像具备可靠性,并且可以有效提高肾小球检测的性能。Histological staining is time-consuming and irreversible in clinical practice,which leads to a relatively limited quantity of renal pathology images,thus restricts the application of medical diagnosis and deep learning method.A deep learning model based on generative adversarial networks was proposed,style transfer between different staining techniques was achieved through a single training process.Subsequently,a multi-domain staining style transfer model was introduced into the glomerulus detection workflow.The staining transformations were performed using the staining transfer model in experiments,and the accuracy and generalization of the glomerulus detection model were improved by the the mutual complementarity of staining features from different styles.The experimental results show that the the images generated by the multi-domain staining style transfer model are reliable and effective in improving glomerulus detection performance.

关 键 词：肾脏病理学 深度学习 生产对抗网络 染色转换 

分 类 号：F830[经济管理—金融学]