Classification Of Cervical Cells Based On Knowledge Distillation

Cervical cancer is a common gynecological malignancy that poses a great threat to women’s life and health.The causes of cervical cancer are clear.Early detection,diagnosis and treatment are important measures to prevent the occurrence of cervical cancer.Currently,cervical cancer screening is mainly performed by manual interpretation of cervical smears by specialized physicians,and the huge volume of cervical smears poses a great challenge to cervical cancer screening.The manual interpretation method not only makes the doctor’s workload huge and time costly,but the screening results are also influenced by the doctor’s expertise and subjective factors.Therefore,it is important to study the cervical cell image classification task and use computer technology to assist in cervical cancer screening.To assist in cervical cancer screening and improve the accuracy of cervical cell image classification,this paper uses the knowledge distillation method for cervical cell image classification research,the main research is as follows:(1)A classification method that combines self-supervised label augmentation and knowledge distillation is proposed.The method changes the definition of "knowledge" in knowledge distillation by means of self-supervised label augmentation,unifies selfsupervised learning tasks and supervised tasks into a single task,enhances the information expression ability of the model,enriches the knowledge reserve of the teacher’s network,better guides the students’ network,and achieves data augmentation effects at the same time.The experimental results showed that the method improved the accuracy of cervical cell classification by 3.09%.And a semi-supervised learning approach was extended on this basis,which is of practical significance in the case of high cost of cervical cell labeling.And based on this,a semi-supervised learning approach is extended,which has practical implications in the case of costly cervical cell annotation.(2)A multiple-stage knowledge distillation method for cervical cell classification is proposed.The method is used to change the learning style of the student network by constructing a multi-exit network architecture to stimulate the learning potential of the student network and improve the model generalization ability.The output of the teacher network is used for supervised training of different stages of the multi-exit network to guide the different stages of learning of the student network.The multiple-stage knowledge distillation method is extended to one-to-one multiple-stage knowledge distillation,and the multi-exit teacher network is used to guide the multi-exit student network one-to-one,and the classification accuracy of 97.65% is finally achieved.The effectiveness of the proposed method is verified by extensive experiments on natural image data sets.(3)A self-supervised label augmentation fused with multiple-stage knowledge distillation is proposed for cervical cell classification.Based on the above study,the method is improved both in terms of the richness of knowledge in knowledge distillation and the way of knowledge transfer,making full use of the advantages of both methods to improve the classification performance of the model,and finally obtaining 98.02% accuracy of cervical cell classification.(4)An intelligent classification system for cervical cell images is designed and implemented.The system can classify cervical cells according to the cervical cell images uploaded by users,predict the types of cervical cells,assist users to judge cervical cell types,and help cervical cancer screening.