Auxiliary Diagnosis Of Cervical Cancer Based On Multi-state Colposcopy Image Fusion

Cervical cancer ranks the fourth among female malignant tumors world,but colposcopy screening and timely treatment of cervical intraepithelial neoplasia(CIN)can greatly improve the cure rate of early cervical lesions.The colposcopy can be divided into acetic acid test,green filter test and iodine test.Each patient will get five acetic acid images,one green filter image and one iodine test image during this process.Based on these three types of images(seven pictures)obtained during this examination,the gynecologist will make a pre-diagnosis and perform a biopsy or cervical curettage if necessary.But even for experienced gynecologists,the specificity of their clinical diagnosis is still less than 70%.Based on the multi-state colposcopy image dataset of 673 patients provided by the First Affiliated Hospital of University of Science and Technology of China(USTC),seven backbone networks are constructed first,and then the green filter image and six other state images are fused at the feature level and trained to obtain six feature encoding networks.Finally,the features of all state images are xonstracted and connected to a two-dimensional vector by a multi-layer perceptron for diagnosis.A multi-state image diagnosis convolutional neural network(MS-CNN)is constructed to fuse all image features to diagnose whether a case is negative(normal and CIN1)or positive(CIN2/3 and cancer).Since our data set is relatively small,and the ratio of negative cases is 53.96% more than that of positive cases,we use a “ logit adjustment ”loss function with label prior information for network training and solve the overfitting problem that often occurs in the training process of small data sets.We used real clinical data to evaluate this framework,and finally realized that the accuracy rate,recall rate and AUC value of classified diagnosis are 82.58%,79.85%and 87.89%,respectively.This framework has improved the diagnosis and treatment level and efficiency of cervical cancer,can meet the needs of clinical diagnosis in remote areas with low resources,and is helpful for the reasonable allocation of medical resources.In order to assist gynecologists in paying more attention to cervical abnormalities during diagnosis process,and assist doctors in performing cervical biopsy.We design a "cervical biopsy point recommendation system",which mainly uses color threshold segmentation technology to locate the lesions of iodine test images obtained during colposcopy,and generates biopsy site markers to help doctors to locate the biopsy and improve the accuracy of the biopsy.