Segmentation And Classification Of Cervical Cancer Based On Deep Learning

Cervical cancer is a common gynecological malignancy.Early screening and diagnosis of cervical cancer are of great significance for the prevention and treatment of the disease.The traditional pathological diagnosis of cervical cancer is mainly based on observing the characteristic morphology of the nucleus under the microscope,determining the degree of lesion and its pathological grading,limited observation field,visual fatigue and subjective judgment,which can easily lead to missed diagnosis and misdiagnosis of the disease.With the development of digital pathological diagnosis technology,it can automatically provide repeatable and accurate disease features for cell analysis,reduce the difference of diagnosis,and improve the efficiency and accuracy of diagnosis.To this end,this paper mainly carries out the automatic analysis and diagnosis process research of digital pathological image of cervical cancer full section,completes the research work of suspected lesion area detection,nuclei segmentation of cervical cancer,cervical cancer lesion tissue classification and so on,and achieves the following research results:1)Because only a small part of the suspected lesion areas play a decisive role in the diagnosis of cervical cancer in the super-large pathological digital slice images,this paper proposes an improved U-Net network framework based on ResNet for automatic localization and segmentation of suspected lesion areas,which enhances the feature extraction ability of the basic network for cervical cancer cells.Through cooperation with Shanghai National Gynecology and Infant Pathology Department,we constructed a whole slide image data set of cervical cancer,and carried out quantitative analysis and evaluation.A comparativeexperiment was designed and carried out for the data sample labeling method.The accuracy of localization and segmentation of the network reached more than 80%,which met the needs of rapid localization of suspected lesions,and laid a solid foundation for subsequent nuclear segmentation of cervical cancer cells and classification of cervical cancer lesions.2)Nuclei segmentation is very important for the diagnosis of cervical cancer.In view of the low accuracy of existing algorithms for nuclei segmentation with uneven internal gray levels and severe overlapping adhesion,an end-to-end nuclei segmentation method of cervical cancer based on multi-scale and multi-task cascade network is proposed.Drawing lessons from the idea of post-segmentation in traditional image segmentation,we combine foreground segmentation with edge segmentation,add noise reduction learning in the training process to reduce the influence of internal gray level inequality,add edge enhancement learning to improve the segmentation effect of network on adherent overlapping cells,and make full use of the upper and lower information and detail information of pathological image to carry out end-to-end training.By comparing with the existing models,the accuracy of the segmentation algorithm in this paper is improved by nearly 3percentage points,which provides a technical guarantee for further accurate classification of cervical cancer tissues.3)The existing methods only consider some features of the image in the recognition and classification of pathological diseases,ignoring the multi-level and multi-scale features of the image,resulting in low classification accuracy and poor generalization ability.In order to satisfy the different occasions and functions of cervical pathological diagnosis,on the one hand,we classify cells in a coarse-grained way based on migration learning;on the other hand,we combine nuclear regions and nuclei features to achieve fine-grained classification of tissues and cells end-to-end.Compared with the existing models,the accuracy rate ofcoarse-grained classification is more than 91%,and that of fine-grained classification is more than 85%.It provides an important pathological diagnosis auxiliary basis for pathologists to diagnose diseases correctly.On the basis of the above research,the automatic analysis and diagnosis process of pathological image of cervical cancer in whole slide image is also proposed in this paper.By comparing with the manual diagnosis of doctors,the process can provide important basis and means for the diagnosis of pathologists.