Precancerous Lesion Recognition Based On Deep Learning And Cervical Images

The incidence of cervical cancer accounts for the first of the female reproductive tract malignancies.Early screening for cervical cancer is important for reducing morbidity and mortality.Acetic acid experiment based on colposcopy is an important technical method for cervical cancer screening.However,this method is subjective and has low precision,which is not conducive to the promotion of grassroots areas without experienced doctors.In view of the above problems,it is particularly important to realize the automation and intelligence of cervical cancer screening.In recent years,due to the advantages of automatic extraction of features in convolutional neural networks in deep learning,it has dominated the field of image classification.Most of the relevant studies that have been reported so far only use the cervical image after acetic acid experiment,the information is relatively simple,and the precision advantage is not obtained more than the artificial extraction feature.To this end,this paper proposes a method of replacing the original image with the cervical ratio image and combining transfer learning.Based on the registration,the method obtained the cervical ratio image before and after the acetic acid experiment,and finally used the pre-trained VGGNet-16 model on the ImageNet dataset to transfer the cervical ratio image to realize the cervical precancerous lesion recognition.In terms of data acquisition,hospitals that have conducted experiments from the site have obtained thousands of colposcopy images,including 110 patients with HSIL(+)data.In order to maintain data balance,a data set containing 110 sets of HSIL(+)and LSIL(-)was constructed,and the sample size was expanded by data enhancement.In this paper,two experimental schemes are adopted.The first one uses the pre-trained VGGNet-16 model as a feature extractor by retraining several layers close to the output and treating the remaining network as a fixed one.The feature extractor is applied to the cervical dataset;the second is to fine-tune its part by fixing the weights of some convolutional layers at the beginning of the model,retraining the subsequent layers,and obtaining new weights.The experimental results showed that the average sensitivity,specificity,and accuracy of the first protocol presented in the cervical ratio image were 81.82%,77.27%,and 79.54%,respectively,compared with the cervix original and cervical segmentation maps.The accuracy rate was improved by 17.31% and 11.36%.The average sensitivity,specificity and accuracy of the second scheme proposed in this paper on the cervical ratio image were 80.73%,88.36% and 84.55%,respectively.The graph and the cervical segmentation map increased the average accuracy by 12.28% and 3.19%.At the same time,the average sensitivity,specificity and accuracy of the current research methods in our dataset were 79.21%,respectively.75.69% and 77.45%.In summary,the effectiveness of the proposed method is proved.