Nuclear Image Segmentation Based On Pixel Classification And Distance Regression

Nuclear image segmentation refers to the extraction of nuclei from the complex background area through a series of computer vision methods.Nuclear image segmentation is a basic premise in digital pathology workflow,which is significant for cancer diagnosis,grading and prediction.In recent years,nuclear image segmentation methods based on deep neural networks have achieved remarkable results.However,due to low contrast of nuclei,large difference in spatial distribution of nuclei,adhesions and overlapping among nuclei,accurate segmentation of nuclear images has become one of challenging problems in medical image analysis.For that,some existing networks and some traditional appoaches for nuclear image segmentation are first summarized in this thesis.Then,a novel nuclear image segmentation method is presentend to address such difficulties in this task.The major works in this thesis are summarized as follows.(1)In order to reduce the impacts of low contrast and overlapping on nuclear image segmentation,we present a novel nuclear image segmentation method,which is based on pixel classification and distance regression.The segmentation method uses pixel classification to obtain pixel classification feature maps with boundary information,and uses distance regression to obtain distance maps with location information.The combination of pixel classification feature maps and distance regression feature maps can effectively improve the segmentation performance of nuclear images.(2)In the skip connections between the encoder and the decoder,a global information attention module based on spatially separable convolution is constructed.This module uses the high-level features to screen the low-level features through the attention mechanism for well guiding the extraction of low-level features and the enhancement of feature contrast between the foreground and the background.Besides,introducing the spatially separable convolution into the global information attention module not only ensures the excellence of features extraction but also reduces the complexity of training.(3)To aim at better utilizing of pixel classification maps and distance regression maps,a feature aggregation module is presented,which aggregates the pixel classification features and the distance regression features from the same feature level.The module not only captures the relevance of the two tasks,but also retains the differences between the two tasks.(4)A post-processing technique based on marker control watershed is introuced,where the distance regression maps are taken as the marker to refine the pixel classification maps,thus achieving the accurate segmentation of nuclear images.The proposed algorithm is implemented by using Tensor Flow deep learning framework under Ubuntu 16.04 system environment.On several available datasets,the training and testing methods for the propoesd nuclear image segmentation algorithm are elaborated in detail,and comparative experiments are designed to compare the performance of the proposed method with other existing algorithms.The experimental results show that the proposed method in this thesis achieves excellent segmenting performance on different kinds and different adhesion types of nuclear images.In addition,the proposed method has better segmentation performance than other methods on seveal public datasets.