Research On Automatic Segmentation Algorithm Of Medical Image Based On Deep Learning

With the rapid development of computer vision technology,medical images were widely used in the field of medical image analysis,and medical image segmentation had become an intervention that could be used for image guidance.This article took the human oral mucosal micronucleus cell image as the research object,the purpose was to accurately segment the nucleus in the medical image from the pixel level,and provide a basis for the later DNA damage grade evaluation.Due to the complex structure of the pixel-level medical cell image,the adhesion between the cells and the boundary between the background and the foreground were not obvious,this paper proposes a cell image semantic segmentation algorithm based on the combination of Deep Lab V3+ and FCRF,which solved the problem of loss of semantic information and smooth edge of target segmentation image caused by repeated convolution pooling operation of DCNN.In summary,the main research work of this article was as follows:(1)A systematic description of several commonly used deep learning image semantic segmentation algorithms,including CNN,FCN,U-Net and Seg Net,etc.,the characteristics of different algorithms were introduced in detail,which provided theoretical guidance for medical image semantic segmentation.(2)Aiming at the morphology and position of the cell nucleus in the medical microscopic cell image was of great significance to DNA damage detection,in order to improve the quality of cell nucleus image segmentation,a cell image semantic segmentation algorithm based on the combination of Deep Lab V3+ and FCRF was proposed.Firstly,the Atrous Convolution algorithm was used to expand the receptive field of the network feature map without increasing the network parameters.Then,the Dense Atrous Spatial Pyramid Pooling(Dense ASPP)module was used to extract the dense pixel feature of the cell image and expand the receptive field of the feature map for multi-scale feature fusion.Finally,the semantic classification pixels were used as the input of the energy potential function of FCRF.By considering the pixel correlation between neighboring pixels and the entire image,efficient approximate reasoning algorithms and spatial smoothing algorithms were used to refine the edge of the segmented cell image.(3)The traditional image segmentation evaluation index processed the pixels in the segmentation results in the same way,considering the pixel relationship between different positions in the cell image,and the cell image semantic segmentation evaluation standard were established,so that the objective evaluation results were more in line with the human visual system.Human oral mucosal micronucleus cell image comparison experiments confirmed that the algorithm proposed in this paper had higher segmentation accuracy,and the boundary contour was significantly better than other algorithms.The segmentation accuracy PA was 95.34%,and the MIo U was as high as 87.71%.It had high practicability and generalizability.