Research On Renal Tumor Image Segmentation Based On Deep Attribute Representation Learnin

Medical image segmentation is a technique for segmenting structures and tissues in medical images into different regions,which can help doctors or researchers better understand medical images for more accurate diagnosis and treatment.Accurately segmenting diverse shapes,sizes,and distributions of tissues such as the kidney and tumors from CT scans is a crucial step in patient diagnosis and treatment planning.Traditional manual screening and segmentation methods for tumors require experiential support and consume a lot of time and effort,while segmentation techniques using deep learning can effectively improve the efficiency of diagnosis and treatment by doctors.However,integrating the spatial relationships and dependencies between multiple different scales of the kidney and tumor remains a challenge.Therefore,segmentation techniques based on 3D CT images that consider the similarity and spatial relationships between slices are a valuable research direction.This technique can help doctors more accurately determine the boundary between the kidney and tumor,improve the accuracy and efficiency of segmentation,and provide better assistance for doctors.This article proposes three deep learning-based methods for segmentation of kidney tumor CT images.Among them,the first method is the image segmentation based on graph construction and graph convolutional autoencoder(GAE-Seg).This method maps the three-dimensional deep structural feature information into graph structure data and uses graph convolution to propagate information between nodes.In order to capture the long-distance semantic and spatial relationships between the kidney and tumor,this method calculates the similarity between image feature vectors and uses it as the weight on the edges connecting nodes in the graph structure.Then,the graph structure is used as the input of the graph convolutional autoencoder,and information is propagated along edges to capture non-local features,thereby learning semantic information between the kidney and tumor globally.The second method introduced in this article is the multi-scale topological and attention-based kidney tumor segmentation(MGR-Seg).Firstly,the topological structure between image regions is constructed from deep feature representations,and topological properties are assigned to each node of the image region through multi-step random walks and neighborhood updates.Then,using the multi-scale graph convolutional autoencoder(MGC),deep multi-scale topological representations of nodes are extracted,and knowledge is propagated along the graph edges during convolution and optimization to supplement feature representations.On this basis,the scale-level attention module is used to learn the adaptive weights of multiple scale topological representations,thus achieving adaptive fusion.Evaluation results on the common kidney and tumor CT segmentation dataset show that this method outperforms other segmentation methods.Moreover,the ablation experiments using different backbone networks also demonstrate the main technical innovations and generalization capabilities of this method.The third method is the masked autoencoder-based kidney tumor image segmentation(MAE-Seg)using Transformer.This method uses self-attention mechanism to reconstruct the masked image features for better segmentation of kidney and tumor.Firstly,the input image is divided into small blocks using the masking mechanism,and the uncovered blocks are used as input for the autoencoder in a uniformly distributed manner across the entire region.Then,the Transformer is used to learn the global dependency between features,where multi-head attention can focus on spatial relationships of different information to better capture the dependency between the kidney and tumor at different positions.During the decoding process,the hidden representation of the encoder is first expanded,and the masked part of the features is added and supplemented with positional information as input to the decoder,and the missing area is reconstructed using the Transformer.The masking strategy can compress data,extract features,and remove noise.Experiments with different masking rates demonstrate different segmentation effects,indicating that the masking rate has an important impact on the segmentation results.MAE-Seg method achieves excellent performance on kidney and tumor CT segmentation datasets,and through ablation experiments and comparison with other methods,the effectiveness of MAE-Seg for kidney tumor segmentation is also demonstrated.