| Pathological diagnosis plays a key role in clinical medicine.The analysis of pathological sections by pathologists can not only provide intuitive evidence for the diagnosis,the classification and the treatment of diseases,and also provide objective basis for judging the progress,the prognosis and the efficacy of diseases.Pathologists need to pay attention to the location of the lesion,the mitosis of the nucleus or gland formation and other regions of interest（ROI）.However,in the traditional diagnostic procedure,due to the huge amount of data of whole slide image（WSI）and the acute shortage of corresponding doctors,the process of looking for ROI will not only greatly increase the workload of doctors,but also make it easier to be influenced by subjective factors and even leads to clinical misdiagnosis[1].In recent years,with the development of deep learning,more and more researchers focus on the application of deep learning in the automatic analysis of pathological images.However,due to the characteristics of pathological images,such as few labeled samples,complex background noise,large staining differences,and large differences in cell morphology and different aggregation degree of different organs,it can not attain good results if the existing segmentation models are applied directly to pathological images.This thesis focuses on the two common problems in pathological image segmentation——focus segmentation and nuclear division.Based on the analysis and understanding of the classical segmentation model,it aims to improve the existing problems in the application of deep learning model.The main research is as follows:（1）The algorithm,the evaluation index and the data set of pathological image segmentation are investigated in this thesis.The traditional segmentation method and the classical segmentation method based on deep learning are explored.The model is selected according to the characteristics of pathological image.Also,the network structure analysis is carried out.The performance index of algorithm evaluation and the common data set of pathological image segmentation are introduced and selected.（2）Aiming at the segmentation of lesion region in pathological image of gastric cancer,an improved u-net algorithm based on level feature aggregation is proposed in this thesis.Through the decoder level feature aggregation,multi FCN head design and hard sample mining technology,u-net is improved to achieve the purpose of tumor segmentation in gastric cancer pathological section.（3）In this thesis,a self-supervised method based on convolutional neural network is proposed for the nuclei segmentation in pathological sections of different tissues.The channel attention mechanism and TTA（test time argument）technology are added to the self-supervised nuclear segmentation,and then through the morphological operation and watershed algorithm and other post-processing steps,the task of nuclei segmentation in pathological sections of different tissues is realized. |
