Research On Cervical Cell Image Detection And Segmentation Based On Deep Learning

Cervical cancer is a common gynecological disease endangering women’s health,and it is a malignant tumor of the reproductive tract.Fortunately,this type of malignant tumor has a clear cause,and early screening can reduce its incidence and achieve the effect of effective prevention or even elimination.In today’s medical field,cervical cell smear analysis,such as liquid-based cytology(TCT)smear analysis,is one of the main methods of manual screening for cervical cancer.Traditional manual reading mainly relies on doctors with rich clinical experience to make subjective judgments on the condition.Because the cervical cell images are complicated and difficult to understand,the continuous high-intensity work of the doctor can easily lead to missed detection,misdetection and inefficiency in the diagnosis of lesions.With the great progress of deep learning in medical image target detection,the intelligent detection and segmentation of cervical TCT images to assist disease diagnosis has important research significance for preventing patients from missing the best treatment opportunity and saving patients’ lives.This paper conducts an in-depth study on the two network frameworks with the highest accuracy in the field of target detection,Mask R-CNN and Faster R-CNN.Aiming at the problems of missed detection of small targets and insufficient segmentation,a multi-parameter loss based on NAS-FPN is proposed.Masker R-CNN,an improved algorithm for function and edge information,has effectively improved the accuracy of cervical cell detection and segmentation.The specific work of this article includes the following five points:(1)Data set production.The author cooperated with Changsha Wangwang Hospital,and with the help of the pathology examiner,used TCT smear technology to collect cervical cell images and make them into COCO format.And expand the data set size through data enhancement,so that the data can support the network for full training,avoid over-fitting,and improve the generalization ability.(2)This paper compares the training results of Faster R-CNN and Mask R-CNN on the self-made data set,and selects the basic framework of the improved model as Mask R-CNN.(3)NAS-FPN is proposed for the feature fusion scheme of Mask R-CNN.The best connection method of each layer of the pyramid is searched through the neural network search method,which reduces information loss and effectively improves the recall and accuracy of dense small targets..Experiments have proved that this improvement has increased the recall rate of small targets by up to 4.6%.(4)According to the characteristics of different scale targets’ sensitivity to classification and regression loss,a multi-parameter loss function is proposed to improve the detection accuracy of cells of different sizes in the same image.Experiments have proved that the accuracy and regression rate of target detection results have increased by 5.3% and 4.4% at the highest.(5)In view of the problem that Mask R-CNN divides the candidate area directly,which results in the incomplete fit between the edge of the mask and the real edge of the cell,this paper proposes to identify the edge of the target first,and then fill the closed area formed by the edge.The algorithm effectively improves the convergence of the target edge.Experiments show that the integration of edge information makes the segmentation accuracy 6.6% higher than the original algorithm.