A Study On Kidney CT Image Segmentation With Deep Learning Algorithm

Kidney segmentation in CT image is an important prerequisite for quantitative analysis of kidney disease,and it supports the accurate diagnosis and treatment of the disease.However,the kidney tissue is complex and diverse in shape,and related diseases can also cause the significant morphology change of kidney.These issues make it difficult to accurately segment the kidney from CT images.Existing traditional methods usually have complicated processing procedures and low performance.In recent years,Deep Learning has been widely used in medical image analysis,showing good performance.Aiming at two types of kidney diseases,such as cystic kidney disease and kidney tumors,we conduct the research on the deep-learning model for kidney segmentation in CT images,specific research results include:Residual dual-attention Unet model for cyst kidney segmentation: Cyst lesions can cause large changes in the shape of the kidneys.To cope with the many challenges of automatic segmentation of cyst kidneys in CT slice images,this paper proposes an improved deep network based on the U-net structure in order to pay more attention to the features of the kidney region,the model is designed with a dual attention module with residual connections.Based on the residual structure,the spatial learning and channel attention mechanism adaptive learning are more effective for feature expression.Experimental results show that the model can accurately segment cyst kidneys in CT slice images,and various segmentation indicators are better than other classical segmentation network models.Cascade-gated shape RDA-3DUnet model for kidney tumor segmentation: During the three-dimensional segmentation of kidney tumors in CT images,there are problems of class imbalance and small targets,which makes it difficult to segment tumors.Two-stage segmentation model for network cascade.In order to fully extract voxel information,the residual dual attention module was extended to three dimensions to build a U-shaped full convolutional neural network,and for the problem of unclear tumor boundaries,a tumor shape branch was designed by introducing a gating mechanism in the network,thereby improving the feature extraction performance of tumors.Experimental results show that the segmentation performance of this model is higher than other comparison algorithms.