The Segmentation Of Pathological Tissue Slice Based On Deep Convolution Neural Network

Pathological diagnosis is the "gold standard" for tumor diagnosis.However,professional pathologists need to spend a lot of time on observing huge pathological sections,based on professional knowledge to diagnose the type and classification of tumors.Nowadays,the production of pathological sections is gradually automated,and a large number of pathological sections are saved as digital images,which lays a data foundation for the development of computer-aided diagnosis technology.Accurate automatic segmentation technology is a key prerequisite for subsequent computer-aided diagnostic accuracy.In this paper,based on deep convolutional neural network,taking the pathological section of colon cancer as an example,the accurate automatic segmentation algorithm of pathological tissue section is studied.In the pathological diagnosis of colon cancer,the morphological features of the gland,such as size and contour,are important basis for cancer diagnosis and grading.The difference in morphology of the glands will become larger as the degree of differentiation of the cancer becomes higher.In addition,the glands will be partially close together in the pathological tissue section.Therefore,gland segmentation in pathological sections is a challenging research work.This paper will use the Warwick-QU dataset released by the MICCAI 2015 Gland Segmentation Competition to conduct experiments.In the previous studies on gland segmentation,there were connection problems and dimensional trade-offs,which led to the loss of competitiveness in the calculation of the results of the segmentation.In order to solve the size problem in gland segmentation,this paper proposes a new type of fully convolutional neural network called multi-scale fully convolutional network,which extracts multi-scale features from convolution output of different receptive fields to balance the segmentation of different-size gland.In addition,in order to alleviate the loss of global information of glands caused by pooling layer in the full convolutional network,and to better reshape the finely segmented shape of glands with variable shapes,this paper designed a special high-resolution branch to supplement the global information of the network.For the connection in gland segmentation problem,this paper proposes to use the three-category segmentation method to maximize the division of adjacent glands.The algorithm is collectively referred to as a multi-scale fully convolutional network that combines three-class classifications.Through the vertical comparison experiment of the proposed algorithm,the advantages and effects of each innovation point of the network structure are verified.The comparison with the traditional binary segmentation algorithm proves the superiority of the proposed three-class segmentation method.Compared with other recent gland segmentation algorithms,the proposed algorithm obtained the first overall effect in Warwick-QU dataset,which proves that the proposed algorithm will have a good application in the computer-aided diagnosis of colon cancer.