Research On Cervical Cell Image Segmentation Algorithm Based On Deep Learning

Cervical cancer is one of the most common malignancies in women.With the popularity of screening,the large number of cervical cell smears has put tremendous pressure on pathologists.The shortage of pathologists,the large volume of reading work,the long reading time,and the large variation in the reading experience and subjectivity will all lead to a decrease in diagnostic accuracy.Early screening by medical means and timely detection of cervical cytopathic lesions has become crucial task.Cervical cytology segmentation is the basis for quantitative and qualitative analysis of cells and is the most important part of automated cervical cytology screening.To address the current problems of high background interference,weak feature extraction,poor boundary segmentation,and difficulty in locating a large number of impurities in the real cervical cell images in the cervical cell segmentation task,this paper investigates cervical cell localization as well as segmentation.The main work is summarized as follows:(1)A single-stage detector-based cervical cell localization algorithm is constructed.The current detection algorithms applied in the cervical cell task have problems such as long detection time and no real-time detection.In this paper,a real-time cervical cell localization algorithm based on single-stage decoupled detection is constructed to complete the deep decoupling of cervical cell localization and classification to achieve real-time cervical cell detection and improve localization accuracy.The detection m AP of cervical smear image can reach 0.635,and the detection speed can reach 89 FPS.(2)An ultra-lightweight cervical cell detector is constructed.To further localize the region of interest in complex real cervical cell slices more rapidly,this paper constructs a more lightweight detector by using a lightweight backbone network,Anchor-free,and single detection head to save time for subsequent accurate segmentation of cervical cells,and finally improves the model accuracy by 4.8 percentage points by using the migration learning method.(3)A model of cervical cell segmentation algorithm based on global dependence and local attention is constructed.To address the problems of existing segmentation algorithm models that are easily disturbed by background color,poor boundary segmentation,and lack of global semantic information.In this paper,an embeddable module for feature extraction and efficient utilization of the cervical cell segmentation task is first constructed.It mainly consists of three parts,which are the global context module,channel attention module,and spatial attention module.And it is embedded into the segmentation network to capture the global information dependency to improve the model’s ability to extract features and to complete the model’s self-adaptation to features and the effective information transfer between codecs.An end-to-end network model for accurate segmentation of cervical cell nuclei and cytoplasm is constructed using this embedded module,and the performance of segmentation of cervical cell nuclei and cytoplasm is improved while ensuring that the number of parameters does not increase.(4)A nucleus and cytoplasm segmentation system was designed and implemented for cervical cell pictures,which can effectively extract the nucleus and cytoplasm of cervical cells and can segment cervical cells more accurately.