Deep Generative Adversarial Network(GAN) On Histological Image Analysis And Its Application

The segmentation of histologic images is an important prerequisite for medical image analysis,and has important value in disease research,clinical diagnosis,treatment and prognosis.The traditional segmentation method is generally a manual or semi-automatic method,which is far from meeting the clinical needs in terms of segmentation efficiency or accuracy.Compared with natural images,pathological image recognition is more difficult.The current deep learning segmentation model still has the problem of low segmentation accuracy and robustness in the segmentation task of pathological images.The main reason for these problems is that the general deep learning segmentation model mainly relies on the pre-defined loss function training mechanism for training,so it cannot measure the error between the model output and the true label and guide the model training.In this paper,based on the basic principles of generative adversarial networks(GAN)and conditional generative adversarial networks(c GAN),segmentation models at the tissue and cell levels are constructed in a variety of pathological image analysis tasks.Based on the results of segmentation,many disease-related visual or sub-visual Visual features,and use these features for clinical auxiliary diagnosis and prognosis.In the pathological image analysis task,this paper constructed the following segmentation model and framework based on c GAN:(1)Constructed the semantic segmentation of nuclei in H & E stained pathological images and the instance segmentation model SIc GAN,which can carry out precise semantics on all nuclei Segmentation;(2)Constructed an automatic segmentation model EPSc GAN of epithelial and interstitial regions in H & E stained pathological images of breast.Compared with the current mainstream image segmentation model,EPSc GAN can achieve optimal segmentation performance;Detection and segmentation of myeloblast(Myeloblast)in the pathological smear of leukemia(AML),constructed the model and framework of AMLc GAN segmentation and detection,and then extracted the image features of Myeloblast based on the segmentation result of Myeloblast to be effective for bone marrow metastasis Sex was predicted.(4)For the diagnosis of chronic myeloid leukemia(CML),a megakaryocyte(Megakaryocyte)detection and segmentation model in CML stained pathological sections was constructed,and then the statistical characteristics of Megakaryocytes were extracted based on the segmentation results.diagnosis.(5)In the H & E stained pathological image of oral cancer,a multi-cell nuclei(MN)segmentation and detection model MNSDc GAN was constructed,and tested in a total of 758 full-scan image(WSI)cases provided by multiple units,and finally found The extracted MN features are closely related to the 5-year survival rate of oral cancer patients.