Research On Segmentation Method Of Kidney Tumor Based On Deep Learning

Kidney cancer is a type of malignant tumor with very high morbidity and mortality,early cancer screening can improve the survival rate of cancer patients.CT scan is the most important method for clinical diagnosis of kidney cancer,and it can qualitatively diagnose most kidney tumors.The accurate segmentation of kidney and kidney tumor plays an extremely important guiding role for clinicians in disease diagnosis and surgical planning.Manual segmentation of kidneys and kidney tumors will bring great burden to doctors and is affected by subjective factors of doctors.Each kidney and kidney tumor has a different shape and the pathological characteristics of kidney tumors are also different,Traditional image segmentation methods have limited application scope.Most biomedical images are three-dimensional data.Convolutional neural networks for three-dimensional data make full use of the relevant information between medical image slices to achieve excellent performance in medical image segmentation tasks.For the above reasons,this thesis applies the deep learning method to the diagnosis of kidney cancer,and uses the three-dimensional convolutional neural network to automatically segment the kidney and kidney tumors.The main content of this article is as follows:(1)This thesis introduces the data set used in the experiment.Because the data set is small,the information extracted from the limited data is limited.Therefore,the CT image is enhanced by training data by horizontal flipping,adding noise,and re-sampling the area containing the kidney and kidney tumors to increase the diversity of the training data.(2)The use of three-dimensional convolutional neural networks 3D-Unet and V-Net to achieve simultaneous segmentation of kidney and kidney tumor,and simultaneous kidney segmentation and kidney tumor segmentation,that is,multi-task learning can improve efficiency.Because different tasks have different learning difficulties and convergence speeds,as training progresses,tasks sometimes no longer benefit from each other.Therefore,two-label segmentation of kidney and two-label segmentation of kidney tumor are proposed.The experimental results show that the proposed method has better segmentation effect on this data set.(3)The use of depthwise separable convolution to replace the conventional convolution at a specific position in the V-Net network model greatly reduces the number of parameters of the network model without changing the segmentation effect.A mixed convolution kernel is introduced,and the stacked small convolution kernel and large convolution kernel are combined in a structure similar to Inception to enhance the information extraction ability of the model.(4)This thesis introduces attention mechanism,adding the channel domain attention mechanism SENet and the mixed domain attention mechanism CBAM to each stage of the network model.It was found that adding SE module and CBAM did not improve the segmentation effect,improve the SE module and CBAM,experiments show that the network model with improved SE module and the network model with improved CBAM can effectively improve the segmentation effect.