Nuclear Image Segmentation Based On Multi-bridge Convolutional Neural Network

With the continuous progress of medical technology,people pay more attention to health issues,which has spawned many medical image processing technologies.However,in the whole medical tissue,the nucleus of the cell carries a large number of genomes,and the recognition of different types of nuclei can also provide morphological information about the nucleus according to the division of the nucleus.Therefore,it plays an important role in the early screening of cancer and the later graded treatment.In order to further analyze the cell image,the nucleus segmentation is usually carried out first.However,in the early nuclear image processing,the artificial segmentation method was time-consuming and laborious,at the same time,it produces interference due to the external environment,resulting in poor segmentation effect.With the development of computer technology,deep learning has gradually become a hot topic in the field of medical image analysis.This paper starts with the analysis of traditional convolutional neural networks,and gradually proposes a novel neural network method,which contains a new convolution block in the architecture.Compared with the traditional neural network,it is confirmed the network segmentation performing well in adhesion cell task,but for subtle image noise and more fuzzy hollow pixels are still unable to predict.In order to solve the above problems,a multi-bridge convolutional neural network is further proposed,which combines the new structure proposed in previous research work and some unique bridge connection modules to achieve excellent results in nuclear image segmentation.The main innovative research methods in this paper are as follows:(1)Inspired by U-Net and ResNet,this paper proposes a convolutional neural network containing residual addition and multi-scale convolution blocks in first,and then it carries out simple preprocessing operations such as normalization,color conversion and we use a series of mathematical morphology enhancement techniques to expand our training dataset and prevent data from overfitting.In order to obtain the results with the same scale as the input in the entire network,an end-to-end parallel symmetrical design is put forward,especially through the redesign of the convolution blocks;our works get a wonderful prediction effect on the premise that the training data and training parameters are consistent.The Jaccard index of the nuclear segmentation results is 0.0421,the F1 index is 0.8389,and the accuracy is 0.9687.Compared with a simple series of convolution accumulations of the same type,our design obtains a good result.(2)In view of some method limitations of previous research work,further theoretical and experimental proof verifies that according to different network layer feature extraction ability puts forward those different convolutional blocks,at the same time the multi-bridge connection model and the learning consolidate middle layer of inverted U-shaped,which increase the network learning ability in a certain extent,and achieve good results in solving weak fuzzy signal prediction.Finally,the results of the multi-bridge convolutional neural network have shown: in the Jaccard index: 0.0384,F1score: 0.8548 and the accuracy: 0.9735,especially in the processing of fuzzy hollow cell nuclear images,these fuzzy pixels can be predicted clearly.The proposed outperforms than state-of-the-arts networks.