Research On Nuclei Segmentation Algorithm Of Digital Pathological Image Based On Deep Learning

Currently,many diagnoses of diseases rely on pathological tests,and morphological and structural changes of cells provide important criteria for disease detection,making cell analysis an important pre-requisite for pathological diagnosis.However,traditional microscopic observation methods rely strongly on manual work,which is not only timeconsuming and labor-intensive but also subjective.The advent of full section scanners has enabled the digitization of histopathology images and further enabled the process of computer-aided pathology diagnosis by combining information technologies such as machine vision and artificial intelligence.However,the task of nuclear segmentation on digital pathology images still faces many challenges.The large variation in the morphological structure of cell nuclei and the presence of overlapping adhesions make it difficult to accurately segment the nuclei.Many current approaches are trying to solve this problem using deep learning,but it is difficult to achieve good results with only limited data due to the production cost of image annotation.To cope with these problems,the following studies based on deep learning method are carried out in this paper:(1)For the problem of non-uniform staining caused by the production environment during scanning of cytopathological images,a pre-processing algorithm based on stain separation is implemented in this paper.By means of the blind source separation idea,the staining matrix is estimated from the original image to obtain the contribution of each dye to each channel of the image,and finally the image normalization under various staining conditions is achieved.(2)For the cell nucleus foreground segmentation task,a cell nucleus segmentation algorithm based on U-Net convolutional neural network is designed in this paper.The model is prompted to encode different semantic details by introducing a multi-scale attention mechanism in the network to integrate semantic information at different scales and generate global features.Later,the SE module is used at the decoder side to filter the redundant information.This enables the model to utilize richer contextual features,thus improving the performance of the neural network.The effectiveness of this model is finally demonstrated on two datasets.Finally,the Dice coefficients reach 0.831 and 0.825 in the Mo Nu Seg and Co NSe P datasets respectively,proving the effectiveness of the method.(3)For the cell nucleus overlap problem,a multi-task learning neural network is designed in this paper,which divides the decoder side into two task branches by hard parameter sharing.The first task branch is used to predict the cell nucleus foreground,and in the second branch,the cell nucleus pixels are assigned multiple ordinal labels using Euclidean distance and the ordinal rank is predicted.Finally,the individual cell nuclei are then separated using a post-processing algorithm with labeled watershed,which improves the segmentation of overlapping cells.And under the condition of small data volume samples,the self training strategy based on pseudo-labeling is used to alleviate the contradiction between data labeling cost and labeling demand to a certain extent.Finally,the AJI indicators on the two datasets reached 0.658 and 0.592,showing the advantage in separating overlapping nuclei.In summary,this paper combines machine learning techniques and traditional image processing methods to apply convolutional neural networks to the cell nucleus segmentation task,combining their respective advantages to achieve improved segmentation performance,and verifies the effectiveness of the method through ablation experiments,further exploring the feasibility of the segmentation scheme under small data volume.