Nuclei Detection Of Breast Cancer Pathological Images Based On Deep Learning

Cancer is one of the major diseases threatening human health in the world,and female breast cancer has become the cancer with the highest proportion of new cases in the world,and the morbidity and mortality are also the highest among women.Among them,the number of mitotic cells is an important indicator for pathological diagnosis and grading of breast cancer.However,mitosis detection in breast cancer slides is currently largely performed manually.On the one hand,this method is very cumbersome and time-consuming.On the other hand,this method is greatly affected by subjective factors,and the reproducibility of the results is poor.Different pathologists have different quantitative results due to different experiences,which may lead to different breast cancer grading results.In response to the above deficiencies,the introduction of computer-aided automatic detection,especially deep learning methods,has attracted more and more attention from researchers in recent years,which can help reduce the workload of doctors and improve diagnostic efficiency.At present,with the deepening of research,a large number of deep learning methods have been proposed to realize the automatic detection of mitotic nuclei in breast cancer slices,but due to the complexity of mitotic nuclei,the interference of other cells with similar shapes,and the influence of extremely low cell density in a single patch,the current various methods have not achieved good application effects of deep learning in other fields.In this thesis,through the object detection method based on deep learning,a deep learning model for breast cancer mitotic nuclei detection is constructed.And computer-aided automatic detection technology is used instead of manual identification by pathologists to obtain image-level classification results,which can greatly save human resources and precious medical resources.It can also help doctors complete disease screening more quickly and accurately,and greatly improve the efficiency of disease diagnosis,which has good practical significance and medical value.The research content and main innovations of the paper are as follows:(1)We propose RTR-Net,a breast cancer mitotic nuclei detection algorithm based on a two-stage cascaded network.It mainly includes two parts: the first-stage detection model and the second-stage classification network.The previous stage uses a backbone network that fuses multiple modules to extract features,and uses a classification subnet that fuses multiple branches to make predictions.The latter stage uses a variety of improved classification networks for further classification,which improves the overall detection performance.(2)We propose FSPR-Net,a breast cancer mitotic nuclei detection algorithm based on parallel voting.It mainly includes parallel feature extraction network,voting mechanism and post-processing classification network.First,we use two parallel backbone networks to extract features,and then the detection results are scored through a voting mechanism,and different slices are sent to different post-processing classification networks for screening according to the scores.(3)We cooperated with the hospital to launch a new breast cancer mitotic dataset GZMH.It mainly contains 55 whole slide images from 22 patients,which is divided into training dataset and test dataset,and is marked and reviewed by senior pathologist to ensure the accuracy of the dataset.