Research On Cervical Cell Image Segmentation Based On Deep Learning

With the advancements in computer technology,computer-aided diagnosis has become a widely used tool in medical research,particularly in cervical cell segmentation.However,the ideal segmentation effect for overlapping cervical nuclei with complex backgrounds has yet to be achieved.In this thesis,we propose a deep coding network,RAD-Net,based on the residual structure and the improved watershed algorithm to address this problem.Firstly,RAD-Net is developed based on U-Net,with Res Net34 used as the network encoder to extract deep features and obtain multi-level semantic information.A multi-scale residual module is added to the deepest layer to perform multi-scale feature extraction and retain more spatial information,improving the network’s feature reuse.The skip connection part of the network helps to fuse deep and shallow features,reducing information loss in the pooling process.Experiments show that RAD-Net performs well in segmenting cervical nuclei and nuclei of other tissues.Secondly,we propose a new method based on the POA watershed to solve the overlapping adhesion nuclear segmentation problem.The segmentation results of RAD-Net are optimized by morphological reconstruction to remove isolated points and fine adhesions in the image.After the distance transformation of the optimized nuclear image,the POA optimization algorithm is used to optimize OTSU for multithreshold segmentation to obtain the foreground information of the nuclear image.The background image uses the optimized nuclear image directly.Finally,mark points are generated through the connected domain identification of the foreground and background images,and the watershed transformation is carried out based on these mark points.This method can solve the over-segmentation problem of the traditional watershed algorithm,showing better results in overlapping nuclear segmentation.Finally,we develop an additional diagnosis software using the Py Qt platform to enable the practical application.The software can perform image preprocessing,rough nuclear segmentation,and fine nuclear segmentation on the produced cervical nuclear picture.The segmented nuclear picture provides a reference for the radioreading physician’s diagnosis,facilitates the observation of the nucleus size,and helps to screen cervical cancer cells.