| Artificial intelligence has gradually become an important auxiliary means to diagnose medical images.Hematoxylin-eosin(H&E)staining images and immunohistochemical(IHC)staining images are two types of pathological images.Although generous achievements have been made in the field of H&E staining image analysis,the analysis of IHC staining images is rarely explored so far.This paper focuses on the IHC image analysis through deep learning.Specifically,the extraction and analysis of cytological features,plasma cells detection,and the lymphocytes segmentation are discussed,repectively.The contributions can be summarized as follows:1.In the task of cytological feature extraction and analysis,aiming at the unique image representation of IHC images,this paper proposes a high-quality feature fast extraction network which takes a lightweight network as the backbone,supplemented by low-cost multi-scale fusion and costless attention module.This network is able to provides effective feature for subsequent detection and segmentation prediction while meeting the speed requirement.2.In the task of plasma cell detection,a novel framework combining the global scanner and local discriminator is proposed based on the highly sparse distribution of plasma cells.It can alleviate the contradiction between speed and accuracy in the analysis of ultrahigh-resolution pathological images.Moreover,a novel gridded random cropping method is proposed,which can improve the sample balance,diversity,and generation efficiency of the training set.3.In the task of lymphocyte segmentation,algorithms based on transfer learning,semi-supervised learning,and self-supervised learning are proposed respectively to deal with the problem of data scarcity.Specifically,based on the characteristics of the IHC image,the improved interpolation consistency training and embedded selfsupervised training are proposed.The ablation experiments and comparison experiments on multiple public data sets verify the effectiveness of the proposed algorithm. |
