Research On Segmentation Method Of Kidney CT Image Based On Deep Learning

With the continuous progress of medical image processing methods,image processing algorithms play an essential role in medical work.CT image analysis of kidney has a vital clinical value in the medical diagnosis and treatment of kidney related diseases and complications through electronic computed tomography.Therefore,kidney CT image segmentation has become one of the most important challenges in the field of medical image processing.There are some complex characteristics in the task of kidney CT image segmentation,which makes the segmentation of kidney and lesion region very difficult,and seriously affects the segmentation effect of the algorithm.In recent years,both traditional image segmentation algorithms and deep learning segmentation methods are difficult to achieve very satisfactory performance.On the other hand,because of the large amount of network structure parameters,the segmentation model based on deep learning is inefficient,the cost of hardware platform is high and the response speed is slow.In order to deal with the above problems,this thesis puts forward the relevant research as follows:（1）In order to solve the problem of low segmentation accuracy of kidney CT image,this thesis proposes an image segmentation model based on multi-scale UNET,optimizes the segmentation algorithm flow and improves the segmentation accuracy.By introducing the Inception module for multi-scale feature extract,the problem of large difference in the size of the feature scale of the segmented target is solved.By introducing the attention mechanism,we can increase the attention weight to the relevant areas and alleviate the complexity of the details of the location of the kidney pelvis.Using the dual stream input network structure,the edge information stream of renal CT image is obtained to supplement the prior knowledge.At the same time,the mathematical morphology method and watershed algorithm are combined in the algorithm flow to realize the post-processing optimization function and further optimize the segmentation effect of the model.（2）For the problem of low efficiency of image segmentation model algorithm,this thesis improves the image segmentation model by lightweight on the premise of ensuring the segmentation effect of the model.Combined with the improvement ideas of relevant lightweight models,the lightweight module is built by using basic operations such as channel split,group convolution,depthwise separable convolution and channel shuffle.The lightweight module is integrated into the backbone network convolution layer of the image segmentation model to minimize the amount of parameters and calculation of the model and improve the inference speed of the model.Combined with a series of technical methods,this thesis proposes a lightweight multi-scale UNET image segmentation model and post-processing optimization process,so that the kidney CT image segmentation task can be completed accurately and efficiently.Using KITS19 kidney CT image dataset,the algorithm experiment has improved the MIOU image segmentation index to a certain extent,reaching 93.44%.At the same time,the model inference speed has also been improved,and the operation on geforce GTX 1080 Ti GPU has reached 9ms/piece.The experimental results show that the kidney CT image segmentation efficiency and efficiency are better than the existing excellent models and benchmark models,which further shows that the algorithm of this thesis is effective and advanced for CT image segmentation.