| Breast cancer is one of the most threatening cancers to women’s health.Pathological examination is the gold standard for breast cancer diagnosis.Due to the high resolution,diversity and complexity of pathological slices,manual diagnosis not only consumes time and manpower,but various subjective and objective factors will also affect the accuracy of diagnosis.With digital pathology and With the rapid development of computer technology,the advancement of computer-aided diagnosis technology has become inevitable.This paper is based on the deep learning method,focusing on the characteristics of breast cancer pathological images,researching at the tissue level and cell level,and completely realizes the reading of pathological slices,the positioning of the cancerous area,the nucleus segmentation in the cancerous area,and the mitotic cells.The main content of the whole process of determining the pathological grade through the mitotic cell count is as follows:(1)A method based on dilated UNet is proposed to segment the normal,benign,carcinoma in situ and invasive carcinoma in WSI slides,and to locate the cancerous regions.Aiming at the high resolution,scale diversity and feature complexity of breast cancer images,this method uses UNet as the basis for semantic segmentation.The feature extraction part is improved to Mobile Net,which realizes the weight reduction of parameters.Before up-sampling,a structure of concatenated dilated convolution is designed to expand the receptive field.A good segmentation effect was achieved in the ICIAR2018 data set,and the location of the cancerous area was realized.(2)A 2DVNet segmentation network is proposed to realize cell detection and segmentation.Pathological images contain complex information such as cells,matrix and background,and traditional segmentation methods have poor results in segmenting cell contours.Use the 2DVNet network improved based on the VNet network to train with manually annotated data sets and auxiliary data sets to segment the cell nucleus to locate the cell location.This method reduces the 3D voxels of3 DVNet to a 2D convolution kernel to improve the information extraction ability of 2D images,inherits the structure including residual structure,full convolution instead of pooling and other structures,and connects additional layers in series.The branch input of the fusion of high-level and low-level information.After segmentation,a dice coefficient of 86.1% was obtained,and the contour and centroid of the cell were obtained by post-processing.Nuclear segmentation realizes cell positioning,and the cell image block at the centroid position also provides a data set for the subsequent mitotic detection of mitotic non-mitotic cell two-classifier.(3)A cascading method based on classification and segmentation is proposed for the detection of mitotic cells.The four stages of mitotic cells have different shapes,small proportions,and similar characteristics to other cells,so it is difficult to distinguish.Therefore,we use the improved UNet++network and the classifier based on Res Net-50 to segment the candidate set on the data first,and then use the cell image block obtained by segmenting the cell nucleus as the data set to train the classifier,and secondly screen out the mitotic cells.Among them,the candidate set segmentation network adds the output of the deep supervision average pruning to make the cell location information more obvious,and the improved loss function makes the model more sensitive;the classification standard uses the ratio of regional pixels instead of pixel-by-pixel classification to filter.The combined method obtained 0.836 and 0.571 F-Scores on the ICPR2012 and MITOSIS-ATYPIA-14 data sets,respectively,which has improved the detection effect compared with the existing methods.Mitosis detection reads the number of mitotic cells per unit area,realizes the intelligent and automatic classification of the pathological grade of breast cancer patients,and then obtains the corresponding prognostic information. |
PDF Full Text Request