Cervical Cell Image Segmentation And Lesion Recognition Based On Deep Learning

As one of the malignant tumors unique to women,cervical cancer is gradually threatening women’s health and even their lives,and the incidence is still increasing.Early diagnosis of cervical cancer is the key to prevention and cure.At present,pathological examination based on cervical smear microscopic images is one of the main methods of cervical cancer diagnosis,that is,cervical cytology screening,but its disadvantage is that the diagnosis result depends too much on the subjective judgment of the pathologist,and the doctor-patient ratio is too large.Diagnose problems such as lengthy time and inefficiency.Machine-assisted image reading based on deep learning has the advantages of high image reading efficiency,high accuracy,and freedom from geographical restrictions of personnel.Because cervical cytology lesions are complex,segmentation of cervical cells can better help physicians understand cell morphology and other information.The identification of lesions can provide physicians with reference and save time for reviewing physicians to judge again.Professional readers need to rely on years of medical clinical experience to be able to accurately diagnose the condition,and at this stage,the various indicators that use computer technology to make the final judgment of the condition cannot be used for medical testing.In view of the above background,the Healev data set is used to conduct research on cervical cell image segmentation and lesion recognition.In the segmentation algorithm research,first use the segmentation algorithm U-Net,which is suitable for small sample medical images,as the basic network to perform semantic segmentation on cervical cell images.Aiming at the serious misrecognition phenomenon in the experimental results,the experiment is carried out on the basis of this experiment.Improved,the attention mechanism SE module is added to the layer jump connection of the U-Net corresponding layer,and the pooling index is retained during downsampling.Through the above improvements,the 2SU-Net for the segmentation of each cell component of the cervical cell image is designed.Each cell component of the cell image is roughly segmented.Since the high frequency details cannot be completely restored at the edge,the fully connected conditional random field is introduced to optimize the edge of the cervical cell,and the edge optimization of each cell component of the cervical cell is realized.The MPA value and MIo U value reached 97.2% and 96.6% respectively.In the study of lesion recognition algorithm,due to the need for high-precision recognition,Res Net was chosen as the basic network to realize the study of lesion recognition on cervical cell images.The experimental results show that although Res Net can better complete the classification and recognition of cervical cell images,it is recalled The rate is low,and it cannot meet the needs of medical testing.On this basis,DRLNet,a network architecture suitable for cervical cell image lesion recognition,is designed.DRLNet continues the residual idea and combines dense connections for improving feature utilization and fine-grained classification in the case of network training.The local discriminant loss function,and the use of grouping convolution to reduce parameters and calculations,greatly improving various evaluation indicators,accuracy,precision,recall and F1-score reached 98.9%,99.2%,98.8% and 99.0respectively %,the study of lesion recognition on cervical cell images has been completed well.