Segmentation Of Cervical Nuclei Based On Deep Learning And A Conditional Random Field

Cervical cancer is a common disease with high morbidity and mortality.So far,the main screening method is the Pap smear based cytological screening,due to the manual screening by cytologists and pathologists is a task with highly repetitive,time-consuming and error-prone,the development of automatic computer-assisted cytology screening and diagnosis systems for cervical cancer is of great significance,thus the accurate segmentation of the nuclei has become a subject explored by many researchers.In this paper,a coarse to fine strategy is applied to segment the nuclei,in the stage of coarse segmentation,two deep learning methods were employed to obtain the coarse segmentation of nuclei,in the refinement stage,the coarse segmentation as well as the intensity and position information of all pixel in the nuclei RoI are integrated to construct a local fully connected conditional random field,and the refined nuclei segmentation is achieved by minimizing the energy function of the conditional random field.The main innovations of this research are as follows:(1)In order to make full use of the prior information of cytological images to gain accurate and stable nuclear localization,a tiny FCN(T-FCN)with receptive field matching with the resolution of the nuclei in cytological images is proposed based on fully convolutional network(FCN),while this will lead to a reduction in network depth,for the compensation to the loss of network feature representation,T-FCN expands the channel of convolutional layers to twice that of the original FCN.The experimental results show that T-FCN yields a fine coarse segmentation with the precision,recall and Zijdenbos similarity index of 0.84,0.91,and 0.85 respectively,this means that the nuclear segmentation is well matched to its ground truth,however the segmentation is rough in the high-frequency of the nuclear contour,and exists some mis-segmentation in complex cytological images such as those with too pool contrast and too blurred nuclear contour.(2)To overcome the mis-segmentation existed in the results of T-FCN,based on Mask R-CNN,the basic convolutional layers of the original ResNet is improved according to the high nonlinear ability requirement to the basic feature extraction due to the rich basic features of the cytological images in Herlev dataset,meanwhile,on the basis of meeting the extraction of high level features of the nuclei,the high level feature extraction layers of the ResNet are simplified to prevent over-fitting,then a feature pyramid network based on the refined ResNet is employed to extract the multi-scale information of the cytological images to generate the nuclear RoI by a region proposal network,then utilize the ground truth of the nuclear bounding box and the segmentation as the supervised information to achieve the training of the network.The results indicate that the Mask R-CNN based algorithm can address most of the mis-segmentation in some complex cytological images by the T-FCN,and the performance of precision,recall and ZSI are 0.985,0.918,and 0.946 respectively,but still have some problems such as the loss of nuclear contour details and the contour retraction.(3)For the loss of nuclear contour details in the stage of coarse segmentation and the contour retraction in the result of the Mask R-CNN based algorithm,a local fully connected conditional random field is constructed for the stage of refinement,the coarse segmentation is used as the unary potential function of the energy function in the conditional random field,and the pairwise potential function is constructed by integrating the position and the intensity information of all pixels in the nuclear RoI,then minimizing the energy function of the conditional random field to get the maximum posterior probability,and the output of the random field is the refined segmentation.The subjective and objective evaluation shows that the local fully connected conditional random field makes the nuclear contour closer to the ground truth,and balances the precision and recall to address the problem of contour retraction existed in the coarse segmentation of the Mask R-CNN based algorithm,besides,the ZSI of the Mask R-CNN based algorithm is improved nearly one percent with low standard deviations,demonstrating that our method enables more accurate and stable cervical nucleus segmentation than the current state-of-the-art nucleus segmentation methods in Herlev dataset.