Research Of Ki67 Detection Based On Convolutional Neural Network

Breast cancer is one of the major cancers that causes death in modem women.Immunohistochemical evaluation of the cell dyeing score of nuclear antigen Ki67 has become the most widely used method for comparing the level of proliferation between tumor samples.Therefore,the use of Ki67 score has important medical effects on the diagnosis and prognosis of breast cancer.At present,the hospital pathology department detects and calculates the Ki67 score of breast cancer using artificial detection in a microscope.This method is not only inefficient but also inaccurate.In order to improve the work efficiency of pathologists and improve the accuracy of detection,we propose a deep learning method based on convolutional neural networks.We use it for to perform Ki67 detection on breast cancer Ki67 cell images.Experiment is divided into two parts:cell segmentation and tumor region segmentation.The mainly includes the following work:(1)We combine the doctor’s opinion to regulate the production of breast cancer Ki67 cell labeling data.(2)We propose to use Gaussian filtering to convert cell labeling data into heatmaps,and use point labeling instead of traditional labeling methods.We use conversion to reduce the problem of low data generation efficiency caused by traditional labeling methods and the difficulty of cell labeling,so that the network can focus on extracting target features.(3)Based on the existing network,we make improvements and propose a network training strategy that combines active learning to reduce the amount of data required for network training.(4)We propose to distinguish Ki67 cell staining results by H&E,DAB channel separation,and then calculate the Ki67 score to achieve Ki67 detection.(5)We propose a method of cell segmentation combined with tumor region segmentation to optimize cell segmentation results.We use Segnet to segment tumor regions and improve the accuracy of Ki67 score calculation results.Experimental results show that compared with traditional deep learning image segmentation networks FCN and U-Net,using improved networks combined with active learning strategies,the segmentation result F1 score can reach 88.58%.We combine cell segmentation and tumor region segmentation results to have higher accuracy on Ki67 detection.This method can save a lot of manpower and time costs for Ki67 detection by pathologists,and has better development prospects.