Rearch On Classification Methods For Cervical Cells Based On Convolutional Neural Networks

Effective and extensive cervical cancer screening is one of the core methods to eliminate cervical cancer worldwide.Cervical cells classification is an important part of computer assisted cervical cancer screening,which is of great significance in providing references for early diagnosis and planning of proper treatments.The core purpose of this paper is to improve the classification performance as well as reduce the training cost or hardware requirements of the computer aided cervical screening system.This paper carried out a study on the multi-category classification method of cervical cells based on convolutional neural network,so as to provide some supports for the application of portable equipment or devices in cervical cancer screening.The main research contents of this thesis are as following:Firstly,in order to reduce the dependence of CNNs on the quantity of cervical cell image samples,a cervical cells classification method that integrates model training,data preprocessing with data agumentation and model structure improvement was constructed.The deep transfer learning method of fine-tuning the pre-trained CNNs on the target dataset was adopted for alleviating the over-fitting problem.Then,on the purpose of increasing the diversity of cervical cell samples,a method of cervical cell image preprocessing and data augmentation was designed.Furthermore,the structures of CNNs were modified to enhance the generalization performance.Experiments were carried out on cervical cell datasets,and the effectiveness of the proposed method was verified.Secondly,a method based on snapshot ensemble for cervical cell classification was studied.On the purpose of improving the classification performance under limited amount of cervical cell samples without increasing computing cost,an algorithm called Transfer Learning based Snapshot Ensemble(TLSE)was constructed.The proposed method overcomes the restriction in snapshot ensemble of training the network from scratch.A new training strategy of variable learning rate was designed to ensure the stability of model training.The classification performance based on TLSE method with different CNN architectures on Herlev dataset was studied,and the effectiveness of the proposed TLSE method was verified.Thirdly,a method based on knowledge distillation for lightweight cervical cells classification was studied.In order to explore portable and efficient networks with enhanced performance,a method based on knowledge distillation and lightweight CNNs called LKD was developed.Lightweight CNNs were designed as the basic models for reducing the computer cost.Knowledge distillation method was adopted for improved the representation ability of the lightweight models.Furthermore,in order to ensure as much "dark knowledge" as possible can be transmitted to the lightweight CNNs,the training and selection procedure of the teacher model was designed.The influences of different lightweight CNNs with proposed method were also studies.Experiments results shown that the proposed method can be adopted for developping lightweight CNN model with higher classification accuracy.Fourthly,a method of hybrid loss-constrained lightweight CNNs for cervical cells classification was studied.On the purpose of improving the classification performance of the confuing samples in cervical cell data with low computer cost,a classification framework based on lightweight CNN and hybrid loss function was proposed.The hybrid loss function was built based on triplet loss and cross entropy loss,which alleviate the inter-class similarity and intra-class difference in cervical cell classification task.Furthmore,considering the advantages of Ghostnet in computer cost and feature reusement,a method combined Ghostnet and hybrid loss function was proposed.Experiments were carried out on cervical cell dataset,and the effectiveness of the proposed method was verified by comparing with the existing models and algorithms.At last,the main work of this thesis was summarized and the future research direction was prospected.