Research On Cell Detection Method Based On Convolutional Neural Network

In recent years,with the development of computer technology and medical imaging,medical image processing has become an important adjunct to clinical trials and diagnosis.In the field of medical image processing,deep learning methods occupy an important place.Among them,Convolutional Neural Network(CNN)is the most commonly used deep learning technique,which can automatically extract disease features from medical images to achieve disease classification and segmentation of lesion areas.Medical images generally have a small data sample and high data dimension,which poses a great challenge to the automated diagnosis of diseases.In specific medical tasks,breast cancer grading is an important step in the patient’s prognosis and depends mainly on three features of the patient’s histological section: nucleus polymorphism,tubule formation status,and mitotic cell count.Among them,the number of mitotic cels is an important indicator to assess the aggressiveness of breast cancer.However,it is clinically tedious and time-consuming for doctors to manually count the number of mitotic cels,and a method to automatically detect mitotic cels needs to be developed.Induced Pluripotent Stem Cells(iPSCs)are somatic cels transformed by inducible factors,and iPSCs are similar to embryonic stem cels in terms of gene expression,protein transcription,and differentiation capacity.The health quality of undifferentiated iPSCs is necessary for further experimentation and treatment,but it is time-consuming and subjective for physicians to judge the health quality of cels based on protocol,so there is a need to develop a method that can automatically detect iPSCs and identify their health quality.In this background,this paper makes full use of convolutional neural networks to make studies in mitotic cell auto-detection methods and induced Pluripotent Stem Cell auto-detection methods.The main research elements and contributions are as follows.First,this paper uses two rapid and effective methods to automatically detect mitotic cells from histological images of patients.One,an improved object detection algorithm is utilized for automatic detection of mitotic cels.Specifically,CNN was used to automatically extract mitotic cell features and represent mitotic cels using the fused features,which in turn used the Regional Proposal Network(RPN)to obtain a set of category-agnostic mitotic cell candidates.Finally,mitotic cels were screened from mitotic cell candidates using an improved multibranch classification subnet.Second,based on the above method,a spatial attention module was used instead of a feature fusion module and combined with a multi-branch classification subnet to detect mitotic cels.The present method was validated on the dataset of the ICPR2012 mitotic cell detection competition and achieved the best results compared to other methods.Second,a method is used to detect iPSCs and identifies their health status.The method is divided into two steps,first,the location of iPSCs is detected using a constructed attentionguided multibranch convolutional neural network,and then,the health status of each i PSC is identified by using a constructed attention-guided multi-scale deep supervised classification network.Specifically,the features of iPSCs are extracted via using multiscale convolution operation,and then the features are recoded using attentional modules to construct more discriminating features,and the health status of iPSCs is identified in conjunction with deep surveillance mechanisms.The proposed method is validated on a hospital-provided dataset of iPSCs and the experimental results demonstrate that the proposed method is effective and feasible for the task of automatic detection of iPSCs.In this paper,we have made a preliminary exploration of both mitotic cell detection and induced Pluripotent Stem Cell detection according to a specific medical task,and the method in this paper has achieved better results than the compared methods and provided a basis for subsequent studies.