Research On Key Techniques Of Computer Aided Diagnosis Of OCT Images For Cervical Tissue

Cervical cancer threatens the health of women around the world.In countries with high-income levels,the popularity of screening technology has dramatically reduced the incidence and mortality of cervical cancer.However,existing cervical cancer screening techniques have their shortcomings,and new high-resolution imaging and "non-invasive" screening method is urgently needed.Optical coherence tomography(OCT)is a biomedical imaging technology that has developed rapidly in recent years.It has the advantages of real-time imaging and high resolution and has been widely used in ophthalmology,cardiovascular and other fields.Moreover,some researchers have begun using OCT imaging technology to conduct clinical trials of cervical disease.In this context,considering the shortage of doctors’ manually diagnosis and the rapid development of artificial intelligence technology,this paper intends to combine deep learning with OCT imaging technology to form a new cervical cancer screening technology to assist doctors in making a rapid and accurate diagnosis of cervical cancer patients.In order to achieve the above goals,this paper focuses on the following two issues:1)for the problems of classical convolutional neural networks,propose a classification model more suitable for OCT images;and 2)for the unbalanced data of cervical tissue OCT images,introduce generative adversarial networks to generate a variety of rare category data.Specifically,the main work of this article is as follows(1)Given the shortcomings of classical convolutional neural networks(such as the use of pooling operations,ignoring spatial location information),and the limited ability of capsule networks to process large-scale images,this paper combines capsule networks with VGG16 to propose an OCT image classification model.In order to verify the validity of the model,the classification results of the model were compared with human experts.The experimental results show that the model is superior to the average level of human experts and can be applied to computer-aided diagnosis.At the same time,this paper designs a "five-fold cross-validation" to compare the model with the VGG16 model.The accuracy of the former in five-category task was 84.9%±1.2%,and the accuracy of the second-class task was 88.9%±2.5%,which was better than VGG16 model(80.9%±2.1%and 87.3%±3.9%).It indicates that the model is better and more stable(2)Considering the shortcomings of the original generative adversarial networks(GAN),which are challenging to train and easy to "mode collapse." This paper introduces the WGAN-GP model to generate samples of low-grade lesions in OCT images of cervical tissue.The generated samples have similar resolution and morphological characteristics with the real samples.In order to verify the validity of the generated samples,this paper designs a corresponding comparative experiment,adds the generated samples to the training set,and tand rains the classification model proposed in this paper.Compared with the result of not adding the generated samples before,the sensitivity index was improved by 3%.