Research On Medical Image Segmentation Algorithm Based On Deep Learning

With the continuous improvement of medical level,people pay more and more attention to their own health condition,which promotes the rapid development of medical imaging and recognition technology.Compared with simple language and text description,images have unique advantages in the expression of information,and images can carry and contain more information elements.Medical images can provide people with more abundant and more intuitive detection information and play an increasingly important role in medical diagnosis.Automatic medical image segmentation can help doctors quickly identify and locate the lesions in the patient’s body,observe the changes before and after the lesions,and qualitatively evaluate the effect of patients before and after treatment.The liver is one of the body’s vital organs and can protect the body from many kinds of damage.Localization and identification of liver tumors are important research directions in the medical field,and automatic segmentation of liver tumors has great significance of clinical treatment.It is helpful to accurately calculate the number of diseased cells and grasp the situation of cytopathic changes by accurately dividing the region and location of biological cell nuclei.In this thesis,medical image segmentation methods are studied.The research content mainly includes the following points:Study on segmentation method of liver tumor,in this thesis,Global Attention Upsample(GAU)was introduced on UNet network to segmentation method of liver tumors,and it was called GAUNet.GAU attention characteristics of input,a senior figure in the global execution module average pooling and convolution operation,each channel is a low-level Feature Map provides weight information,enhance useful information channel,inhibition of useless information channel,which can strengthen figure contains important low-level feature space and detail information,improving the capacity of network segmentation.The residual block in the Residual Network makes the extracted feature information flow more easily between network layers,which can not only increase the network depth to improve the accuracy,but also alleviate the gradient disappearance during the back propagation.In this thesis,Residual Module of Res Net is used to replace the convolution module in GAUNet network decoder to further improve the network performance.Finally,the combined RGAUNet network has a relatively good segmentation effect on liver tumors.Compared with FCN,UNet++ and other deep networks,higher segmentation performance indexes are obtained.For nuclear segmentation,the nuclear structure is relatively complex and the edge information is abundant.The Feature Aggregation Module FAM(Feature Aggregation Module)is introduced into the basic network using Deep Lab V3+ as the basic network skeleton to better compensate for the information lost in spatial location and enhance the network’s ability to extract high-resolution Feature Map information.In the decoder part,the low resolution high-level Feature Map is fused with the enhanced high resolution low-level Feature Map,so that the fused Feature Map retains more detail information.The common Convolution in ASPP module is converted into deeply separable Convolution,which makes it better to extract the information of Multi-scale Feature Map and capture more spatial and detailed information.The Multi-scale Feature Map obtained by processing are connected together,which is beneficial to the segmentation of objects of different sizes.The DSFDeep Lab V3+ network proposed in the thesis realizes the optimization of Deep Lab V3+ network.Compared with Deep Lab V3+ network,the indexes of MIo U,Dice and Accuracy of DSFDeep Lab V3+network are improved by 3.43%,3.58% and 3.44% respectively.Compared with other deep learning methods,the segmentation index is relatively high,which is 1.86%,1.65%and 3.14% higher than DANet segmentation performance index of MIo U,Dice and Accuracy respectively.DSFDeep Lab V3+ showed good results in nuclear segmentation.