Cervical Cell Image Segmentation Method Based On Multi-Task Learning

With the rapid development of artificial intelligence,computer-aided diagnosis systems for cervical cancer have become available.This system can effectively assist doctors in detecting cervical cancer precursors,improving diagnostic accuracy,and reducing the workload of diagnosis.The segmentation of cervical cell images is a crucial technique in computer-aided diagnosis systems for cervical cancer.This method can be used to segment single nuclei and cytoplasmic regions from the image,which is the premise of calculating cell morphology and texture features,thus classifying these cells.However,there are many kinds of cells in the image.Their uneven staining and pathological changes will lead to similar colors between some nuclei,cytoplasm,and background,resulting in blurred boundaries.In addition,many adhesive or overlapping cells will further increase the difficulty of image segmentation.Currently,image segmentation based on deep learning is widely used in nuclear and cytoplasmic segmentation tasks.The deep learning method based on a single task only realizes each task’s training and prediction,ignoring the shared benefits that similar tasks may bring.The multi-task learning method can improve each task’s learning efficiency and quality by learning the connections and differences between different tasks and improving the overall performance of the model.In summary,cervical cell image segmentation is challenging,and the multi-task learning method will be more suitable for the study of the method in this paper.Therefore,multi-task learning is studied in-depth to achieve the segmentation of the nucleus and cytoplasm.The main work of this paper is as follows:(1)An Attention Learning Network(AL-Net)based on multi-task learning is proposed to address the problem that the current cell nucleus segmentation methods do not fully utilize the edge and shape prior information of the nucleus.The network includes a main task to predict the nucleus regions and an auxiliary task to predict the nucleus edges.The auxiliary task is used to enhance the feature extraction ability of the main task.An attention learning module is designed to obtain salient and focused attention in the nucleus segmentation task.In addition,a context encoding layer is proposed to extract the contextual features of cell nuclei.AL-Net outperforms the mainstream nuclei segmentation methods on datasets such as 2014 ISBI,BNS,and nuclues Seg.(2)A multi-task style-aware two-stage overlapping nuclei segmentation method is proposed to address the influence of unknown style images and overlapping cells on the segmentation algorithm.First,a style-aware Multi-style Standardization Network(MSNet)is designed for image style standardization.Then a two-stage segmentation method is used to segment the overlapping cell nuclei,and finally,the overlapping region pixels are reconstructed.Furthermore,two datasets are established in this paper to evaluate the proposed method comprehensively.The MSTransfer dataset covers most image styles,and the CNSeg dataset contains three subsets.The proposed method can effectively improve the generalization ability of the segmentation model,and its performance is better than the current mainstream overlapping nucleus segmentation methods.(3)A weakly supervised cytoplasmic segmentation method of multi-task collaborative learning is proposed to address the problem that current segmentation methods require a large amount of pixel-level labeled data for training and do not adequately consider the severe overlapping between cytoplasms.A Deformable Convolution(DC)branch is added to the one-stage instance segmentation model Cond Inst to segment overlapping cytoplasms.An unsupervised loss based on region consistency is designed for training DC branches to overcome the problem of sparse pixel-level labeled data.The consistency loss is used to realize the collaborative training of the DC branch task and main task.In addition,MS-Net is used to standardize the style of cell image before training to improve the robustness of the model further.The proposed method outperforms the current mainstream weakly supervised instance segmentation methods on self-built datasets and achieves 95% of the performance of fully supervised methods on publicly available datasets.In summary,a series of solutions are proposed to address the challenges in cervical cell image segmentation in this paper.The proposed methods’ effectiveness is verified using many experiments.The results show that the multi-task learning method in this paper can significantly improve the performance of the nuclear and cytoplasmic segmentation models.The attention learning module can focus on specific regions and reduce the prediction errors of the nucleus segmentation model.Adding category branches to the generator can effectively improve the effect of cell image style normalization.The weakly supervised method of multi-task collaborative learning for segmenting cytoplasm can achieve performance close to the fully supervised method.