Design And Implementation Of A Kidney Tumor Image Segmentation System Based On Deep Learning

Automatic segmentation of the kidney and tumor from computed tomography(CT)images is an essential step in precision oncology and personalized treatment planning.Some previous methods use 2D models to segment the kidney and tumor on a single slice of the patient,ignoring the spatial information between slices.There are also some methods that use simple 3D models to segment kidneys and tumors,without fully mining the information in the feature map,and the correlation between channels is not fully utilized.In addition,some of the previous methods only used the characteristics of local areas,and did not correlate the relationship between different local areas.However,it is a valuable research to use the information of the feature map itself from the 3D CT image,the correlation information between the channels,and the correlation information between the regions to segment the kidney and tumor.This paper proposes two methods using deep learing to automatically segment the kidney and tumor from the tomographic image of the kidney tumor.The first is a kidney tumor image segmentation method based on pixel adaptive convolution and depth separable convolution(PD-UNet).PD-UNet uses pixel-adaptive convolution and depth separable convolution to convolve feature maps containing rich semantic information to capture more feature information of kidneys and tumors.First,pixel adaptive convolution is used to adaptively learn the weights of different positions according to the feature information of different positions.Then,the depth separable convolution is used to further extract the information of different channels,while also reducing the convolution kernel parameters.In order to verify the degree of contribution of different types of convolutions of this method,we conducted ablation experiments for comparison.Comparisons with the most advanced methods on public dataset show improved performance for tumor and kidney segmentation.The second method is to integrate channel context attention and region-associated attention for kidney and tumor segmentation(CR-UNet).CR-UNet,to extract,encode and adaptively integrate multiple layers of relevant features.Since the semantic features of different channels contribute differently to the segmentation of kidney and tumor,we introduce semantic attention mechanism of channels.We design a feature map selection strategy to extract detailed information about the kidney and tumor.The regional association attention mechanism is established to integrate the semantic and positional connections between the different regions.In the end,we put forward a multi-angle feature fusion strategy,which fully integrates the semantic channel features,original detail information and regional association features.Ablation experiments demonstrate the contributions of semantic associations between learning channels,detailed information extraction,and regional relation modelling.Comparison results with the most sophistic methods over public dataset indicated improved tumor and kidney segmentation performance.Finally,using the pytorch deep learning framework,the above two segmentation algorithms are implemented,and the kidney tumor segmentation system is designed and implemented.Using this system,two segmentation algorithms can be selected to obtain the segmentation results.At the same time,the system has functions such as image segmentation and image visualization.