Research On 3D Segmentation Depth Algorithm Of Renal Tumors In Plain CT Images

Automatic segmentation of the kidney and tumor targets in the CT image sequence can provide doctors with a quantitative diagnosis basis in clinical diagnosis.The current 3D segmentation technology based on deep learning has been extensively studied.However,due to the complex shapes of the kidney and tumor targets in the CT image diversity,and there are problems such as small tumor targets and few samples.The existing three-dimensional segmentation network cannot segment tumors and other small targets well.How to accurately segment kidney tumors is still a challenging problem.At the same time,the current three-dimensional depth segmentation network model is large,has many parameters,and requires high computer hardware conditions,which also brings great challenges to the actual deployment of the model.In response to the above problems,this thesis carries out in-depth research on the three-dimensional segmentation of kidney tumors in plain scan CT images and the compression and application of three-dimensional network models.The specific work includes:(1)Three-dimensional dual attention-driven CT image kidney tumor cascade segmentation network: In order to cope with the many challenges in the three-dimensional segmentation task of CT image kidney tumor,this thesis proposes a three-dimensional dual attention-driven cascaded segmentation network,two-level network both use a multi-scale codec structure,through the learning of high and low resolution CT image sequences,to finely segment the kidney tumor area,and deploy a three-dimensional dual attention module in the network.This module can capture the complementary information of the main path through the dual path.Thereby extracting complete key area feature information is conducive to more accurate positioning of small-scale tumor targets.The experimental results show that the model can accurately segment the kidney tumor in the CT image,and the segmentation index is better than other three-dimensional segmentation networks.(2)Model pruning of the ‘soft filter standard' of the 3D segmentation network: In order to improve the computational efficiency of the 3D segmentation network and solve the problem of difficulty in actual deployment caused by the complexity of the 3D segmentation model,this thesis further constructs a ‘soft filter for the 3D segmentation network'.The ‘filter pruning' method further extends to the three-dimensional segmentation network.This method uses the norm as the pruning standard,further sets the corresponding pruning rate,and adopts the strategy of pruning while training.The key point is that this method will continue to update the parameters of the pruned filter in the next cycle of training,so that the network model can learn more information.The experimental results show that this method can effectively eliminate the redundant parameters in the 3D network model while maintaining the current 3D network segmentation accuracy,reduce the complexity of the network structure,and realize the compression and acceleration of the 3D network model.(3)Three-dimensional segmentation software for CT medical images: Using the lightweight network after pruning,an integrated software for network model training and sample testing and evaluation for three-dimensional segmentation tasks of CT medical images such as kidney tumors has been further developed.The batch evaluation of CT data improves the efficiency of doctors in clinical diagnosis and realizes the deployment of deep learning three-dimensional segmentation models in actual application scenarios.