Research On Cell Segmentation Method Of Lung Cancer Pathological Image Based On Deep Learning

Pathological diagnosis is the most intuitive and accurate gold standard for cancer diagnosis.Cell segmentation of pathological images is of great significance for cancer diagnosis,grading,and prognosis.However,when computer analyzes pathological images,it also encounters difficulties and challenges.For pathological images,manual annotation not only requires professional knowledge,but also obtains very little annotation data.In addition,there are often problems in pathological images,such as different cell size and morphology,low contrast of images,cell adhesion or overlapping,etc.,which cause many cell segmentation algorithms to fail to accurately segment cells.In view of the above problems,the main work of this thesis is as follows:Aiming at the problems of large workload and low efficiency in the process of cell labeling in pathological images,a semi-automatic labeling method for lung cancer pathological images based on deep learning is proposed to quickly label pathological images.First,the labeling of pathological images is divided into cell center point labeling and cell boundary labeling.The cell center point labeling model and cell boundary labeling model are trained through the public pathological image dataset,and then the cell center point labeling model is used to obtain the cell center point coordinates.The cell boundary label is obtained by combining the cell center point coordinates and the cell boundary labeling model.The semi-automatic labeling method achieved good labeling results on public datasets and lung cancer pathology image datasets.Aiming at the difficulty of overlapping cells segmentation by classical methods in the field of cell segmentation,a cell segmentation method based on multi-task learning is proposed.Firstly,the multi-task network structure is adopted.On the basis of semantic segmentation of pathological image cells,the distance prediction branch is added to extract the boundary information of cells,so as to improve the accuracy of cell segmentation.Then,on the basis of this network,the distance prediction results and semantic segmentation results are post-processed,and the boundary distance map obtained by the distance prediction branch is extracted as foreground markers,and the semantic segmentation results are used as background markers.Combined with the marker-based watershed algorithm,the pathological Accurate segmentation of overlapping cells in images.The experimental results on the lung cancer pathology image dataset verified the effectiveness of the proposed method.This thesis designs and implements a lung cancer pathological image cell segmentation system.The system provides the pathological image labeling function,and can use the cell center point labeling algorithm and the cell boundary labeling algorithm to assist in labeling to improve labeling efficiency.At the same time,the system is based on the realization of the pathological image viewing function,it provides cell segmentation function.