Object Detection On Cervical Cancer Cell Image Dataset Optimization Strategy Study

Medical images are used in a wide range of application scenarios,and are also a difficult research area for object detection in the medical field.Early stage cervical cancer cells are incompletely mutated,and their characteristics are not obvious,and many women contain precancerous lesions themselves,and cancerous cells are not detected and treated in time.Therefore,it is of great practical importance to improve the detection accuracy of medical images,especially cell images,and the use of target detection for medical images can both help doctors improve the efficiency of diagnosis and provide reference for doctors who lack experience.However,the current object detection network is not satisfactory for processing medical image datasets.In order to improve the detection accuracy of medical image datasets,this paper proposes an improved algorithm based on the YOLOv5 model by introducing the mirror filling strategy and EIo U loss function into the YOLOv5 model;by introducing the Focal Loss and Varifocal Loss loss functions into the YOLOv5 model to improve the detection capability of medical image boundary cells as well as category imbalance data,which improves the model detection effect overall;a YOLOv5-based cell image detection system is developed to assist doctors and other non-technical personnel in medical consultation.The main research works are as follows:(1)To address the problem that the boundary cells of medical images are sliced,the mirror filling strategy and EIo U loss function are introduced into the YOLOv5 model to improve the model detection effect.In this paper,the mirror filling makes more pixel points around the boundary cells to enrich the contextual information.In addition,the slender rectangular target frame caused by the cut score,the localization loss of YOLO model is difficult to fit the target box perfectly,which causes the error of detection results.In this paper,we propose to use an improved localization loss by calculating the difference value of width and height separately instead of the aspect ratio,which can make the prediction frame fit the target box better and reduce the localization loss overall.(2)To address the problem of category imbalance in medical image datasets,Focal Loss and Varifocal Loss loss functions are introduced into the YOLOv5 model to improve the detection effect of the model on category imbalance datasets.This paper solves the problem of extremely unbalanced number of positive and negative samples in target detection by introducing the Focal Loss function: since Focal Loss itself has difficulty in dealing with the problem of difficult and easy samples in multi-category object detection,Varifocal Loss is introduced to further optimize the deficiency of Focal Loss;finally,this paper combines the Focal Loss and Finally,this paper combines the two loss functions of Focal Loss and Varifocal Loss to propose FVF Loss,which can improve the detection effect of difficult and rare categories as much as possible while ensuring the detection ability of common categories.(3)In order to facilitate doctors and other non-technical personnel to truly appreciate that object detection can be a good aid to the consultation process,this paper develops a cell image detection system based on YOLOv5.The main responsibility for the development of the administrator role in the system,the administrator has patient management,physician management,model detection management,personal center and other functions,can be more convenient for the system to provide the doctor role in the perfect auxiliary functions.The system has been verified to be highly accurate,safe and reliable through extensive experimental testing.