Deep Learning-based Cell Segmentation Methods For Cervical Clinical LBC Images

Cervical cancer is one of the malignant tumors that endanger women’s health.Liquid-based cytology(LBC)test is the most common screening test for the prevention and early detection of cervical cancer,which aims to detect atypical cells that may be cancerous through the slide of cervical exfoliated cells.In clinical,cervical LBC images can present non-cancerous lesions such as infection,inflammation,and cancerous cells.Cervical LBC images provide the basis for pathological analysis of women’s cervical health.At present,the diagnosis of the cervical cytological image is mainly based on the diagnosis of pathologists physicians.But the diagnosis process is often affected by objective factors such as image staining differences,background complexity,and noise interference.It is also affected by subjective factors such as doctors’ personal experience judgment.The diverse cell types,complex background,and high cell overlap in clinical cervical cell images bring significant challenges for cell segmentation.This thesis conducts more profound studies on the challenges of segmenting overlapping cells in cervical clinical fluid-based images and overlapping cell nuclei in cell clusters.The paper’s primary work is as follows.(1)For the problems of diverse cell types,complex background information,high cell overlap,and difficulty in accurate cell segmentation in clinical cervical cell images,this thesis proposes a generative adversarial overlapping cell segmentation method LCBranch-GAN based on localization constraints.First,we view the cell image segmentation problem as a process of filtering invalid image information,with the localization nucleus as a localization constraint branch input and its localization information as a generative constraint for generating an adversarial network,prompting the network to focus only on the region of the cell image specified by the localization nucleus.Secondly,using the power-generating ability of the LCBranch-GAN network,through the adversarial training of its generator and discriminator,the network can generate a fake single-cell image.Finally,the generated single-cell image extracts the cell from the original image.The resulting cell image retains the features in the original image and has rich and complete boundary information.The visual experimental results show that the segmentation results of the LCBranch-GAN on overlapping cells are consistent with the contour of human eye recognition cells.And the segmentation results of the LCBranch-GAN are not significantly affected in the severely overlapping cell even if the cell edge information is blurred or indistinguishable by the human eyes.We present a quantitative comparison with four classical segmentation methods.The result shows that the LCBranch-GAN performs optimally on self-owned and ISBI2015 datasets.We are achieving a segmentation task of overlapping cell regions in clinical cervical cell images.(2)Because clinical cervical cell images contain many high aggregation cell clusters,their internal cells are severely deformed and overlapping,and the image contrast is low—the above situation poses significant challenges for nucleus segmentation.We propose GCP-Net,a nuclear segmentation method based on gated context-aware information.This method is based on an encoding-decoding structure network,whose encoding path adopts the Resnet-34 pre-training model to optimize the network training speed.The feature extraction module uses the multi-scale context gating and the global context attention module to extract the feature map details and realize the weighted attention on the critical information from the encoding module.The decoding path adopts the decoding module with residual fusion connection to build remote dependence and global context interaction on the low-level semantic information from skip connection to refine the prediction mask and achieve a good nucleus segmentation effect.We performed sufficient comparative experiments with seven existing excellent segmentation models on the Clustered Cell dataset and three typical medical histology image public datasets.The experimental results show that GCP-Net improved in AJI,Dice,and PQ compared with some state-of-the-art segmentation networks,reflecting the superiority of GCP-Net in the self-owned dataset and the generality of the GCP-Net network for other types of cell image.We achieved the nuclear segmentation task of dividing cell clusters in clinical cervical cell images.(3)In the case of the severely overlapping nucleus in highly crowded cell clusters,the segmentation results will have mask connectivity,which will bring false-positive cases for the subsequent pathological diagnosis.This thesis combines the curvature variation characteristics of the intersection of overlapping cell nucleus boundaries and proposes a bending loss-based dual-decoding network overlapping nucleus segmentation method BLoss-DDNet.First,to increase the constraints on the cell contour,a dual-task decoding branch network based on the encoding-decoding structure is constructed,with the encoding module incorporating residual learning and SE Attention module for feature extraction,and the decoding module using mask segmentation task branch and contour segmentation task branch to form a dual-task structure.Secondly,the network architecture search method optimizes the decoding block in the dual-task branch.Obtain the mask segmentation task decoding module and the contour segmentation task branch decoding module,respectively.Subsequently,the feature fusion model is utilized between the dual-task branches to enhance contour sensitivity.Finally,the fusion loss function constrains the network training process,and the bending loss function acts on the contour segmentation branch to penalize the images with significant image bending loss.This thesis conducted extensive ablation experiments and then performed contrast experiments with seven other advanced segmentation networks on our own dataset Clustered Cell and the public dataset Mo Nu Seg.The experimental results proved that the proposed BLoss-DDNet method achieved the best performance in AJI,Dice and PQ,successfully solved the segmentation problem of overlapping nuclei of cell clusters in clinical cervical cell images and provided a strong guarantee for subsequent clinical diagnosis.