Research On Cervical Cytology Image Classification Method Based On Deep Learning

Cervical cancer is a common malignant tumor in women,which has seriously threatened women’s life and health.Early detection,early diagnosis and early treatment are the best ways to reduce the prevalence of cervical cancer.With the popularization of national "two cancers" screening,the increasing number of cervical smears has brought enormous pressure to hospitals.Manual reading is not only labor-intensive,but also prone to misdiagnosis or misdiagnosis due to the subjectivity of interpretation.Therefore,the research on cervical cytology image classification algorithm has important academic value and social significance for cervical cancer screening.Pap smear images have problems such as small differences between classes and difficulty in obtaining labeled images,which brings great challenges to the intelligent classification of cervical smear images.Although some progress has been made in related research,the existing algorithms still have problems such as poor generalization performance and low accuracy.This paper applies the deep learning method to cervical smear image classification,focusing on three aspects: transfer learning,knowledge distillation,and multi-exit network integration.The specific work is summarized as follows:(1)A transfer learning method combining multi-source transfer learning and channel distillation is proposed.Aiming at the problems of insufficient labeled cervical cytology image data and poor model classification performance,an integrated multi-source transfer learning and channel knowledge distillation method is proposed.This method can not only transfer common features between multiple different source domain data,but also utilize different The channel features of the middle layer of the model and the unnormalized probability of the output layer are used as knowledge to realize the knowledge transfer between models.Experiments show that this method can effectively utilize the transfer knowledge,which improves the classification accuracy of Res Net18 by 3.35%.(2)A cervical cell classification network with improved multi-exit ensemble is proposed.Due to the limited "capacity" of the student network,it is difficult to learn all the features used to correctly classify cervical cells,resulting in poor model generalization ability.Aiming at this problem,this paper introduces a multi-exit classification network,and constructs multi-exit networks of different depths to classify cervical cells.At the same time,a global context module and self-distillation algorithm are embedded in each branch network to improve the performance of the exit classifier,and finally the performance difference of multiple exits is averaged using the ensemble strategy,reaching a test accuracy of 97.22%.(3)A novel teacher-student collaborative distillation method is proposed for image classification.Traditional knowledge distillation methods focus on the prior knowledge of the teacher network,while ignoring the self-learning ability of the student network.Based on this,in this paper,based on multi-source transfer learning and channel distillation,the improved multi-exit network is used as a student network,and a cooperative distillation loss function combining knowledge distillation and self-distillation is constructed to realize the collaborative optimization of cervical cell classification network.This method takes full advantage of the performance of the two optimization methods and improves the classification accuracy of the network.To obtain a powerful classifier,the predicted probabilities of multiple sub-models in the student network are fused using an ensemble during the test phase.Finally,extensive experiments are performed on cervical cell and natural image datasets using different teacher-student models to verify the effectiveness and robustness of the method.