Research On Cervical Cancer Assisted Screening Based On Deep Nerual Networks

Cervical cancer is one of the most common female malignancies.Relevant medical research shows that early cancer screening can reduce the incidence and mortality of cervical cancer.Therefore,screening for precancerous lesions has become the key to cervical cancer prevention.Colposcopy is one of the critical medical methods for cervical lesion screening.Because of its convenience and good accuracy,it has become an indispensable diagnostic method worldwide.However,the traditional manual reading method is likely to cause problems such as missed diagnosis,misdiagnosis,and inefficiency in the diagnosis of lesions.Different clinical experiences of different physicians may easily lead to differences in diagnosis,thereby reducing the screening efficiency.In recent years,deep learning has made rapid progress in the field of medical image analysis and has achieved excellent results in many related image classification and image detection tasks.Based on the research of deep neural networks,this paper researches the classification and diagnosis of cervical precancerous lesions in colposcopy images.The main research contents are as follows:A method for detecting and diagnosing cervical lesions using the principle of image semantic segmentation is proposed.A U-net-based full convolutional deep neural network model is established for the CIN1 + diagnosis of cervical lesions.The model appropriately deepens the network depth and implements the segmentation and detection of the lesion area in the form of an encoder-decoder.Finally,the digital image post-processing of the segmentation result further realizes the classification of the lesion.The test set's diagnostic results show that the method can obtain diagnostic accuracy comparable to that of senior doctors and significantly improve diagnostic efficiency.This part proposes a method of cervical lesion image classification based on the pre-trained densely connected network.This method uses transfer learning technology to solve the problem of image classification of small data sets and thelabeling of artificial lesions.It constructs a CNN model with better performance for the more important CIN2 + diagnosis tasks.For neural network training on small data sets,the effects of different pre-trained models on fine-tuning are analyzed,and various training strategies,cross-validation,and setting and selection of hyperparameters such as data volume are discussed.Finally,the differences between the characteristics of the CNN model diagnosis and artificial diagnosis are analyzed,and the general training methods for the cervical lesion classification model are summarized.Experiments show that the proposed model can effectively achieve the classification goal and achieve good results.A classification method for cervical lesions based on the integration of multiple neural network decision features is proposed.Based on the previous work,a new data preprocessing strategy is used to optimize the category distribution of the training data by clustering in limited data to achieve the purpose of optimizing model training.Based on the cross-validation method,the model decision features of the multi-neural network are integrated from the data level and the model structure level.Two integration strategies are proposed based on the internal and external methods of the model output to achieve the performance optimization of classification tasks.The CIN2 + classification was further optimized,and its diagnostic results were more in line with the clinical screening requirements for cervical precancerous lesions.To sum up,through the research and application of deep learning,this paper combines deep neural networks and colposcopy image analysis to explore the classification method of cervical lesions based on deep neural networks.It mainly solves the classification and diagnosis of cervical lesions from image segmentation,image classification,and strategy improvement based on deep learning.Starting from the direction of clinical diagnosis,it continuously overcomes the defects of manual diagnosis and the problems of optimizing computer methods.Finally achieves A better deep neural network classification method for image data in this paper.It provides a theoretical basis for computer-aided diagnosis in the field of colposcopy and cervical lesion classification.It lays a research foundation for the further realization of smart medicine.