Research On Urinary Tumor Segmentation Based On Deep Learning

In today’s medical system,medical image segmentation technology has high clinical application value.For surgical resections and quantitative analysis of lesions,it is necessary to accurately grasp the important information such as the size,three-dimensional structure and specific position of organs and tumors.Medical image segmentation is an important means to provide this information.In recent years,the incidence of urinary tumors has been increasing year by year,and bladder cancer and kidney cancer are two types of urinary system cancer with high incidence.Therefore,research on automatic segmentation algorithms for bladder tumor images and kidney tumor images is a practical need for disease treatment.Due to the complex noise and artifacts in medical images,also the small difference of gray value between different tissues,traditional image segmentation algorithms that depend on manually designed features of prior knowledge are difficult to obtain accurate segmentation results.Therefore,this paper proposes a method based on deep learning to segment the target,and the research work is mainly focused on the target segmentation about MRI images of the bladder cancer and CT images of the kidney cancer.The main research work of this paper is as follows:1)In order to use the features of different receptive fields to distinguish bladder tumor and bladder wall,this paper proposes a multi-scale network(DMC-Unet)based on dilated convolution.Firstly,the network uses dilated convolutions with different dilation rates to obtain different receptive fields,and then realizes the structure of branch down-sampling to obtain features of different scales.Secondly,the residual structure is used to extract feature as the sub-module of DMC-Unet,which effectively alleviates the disappearance of the gradient.Then,a data augmentation strategy combining Gaussian noise,Gamma transform,and spatial geometric transformation is proposed,which avoids the phenomenon of over-fitting caused by the limited data.Finally,the effectiveness of our methods is verified by experiments,and high accuracy has been obtained for bladder wall and bladder tumor segmentation aiming at two-dimensional MRI images.2)In order to solve the problem of category imbalance caused by kidney tumor voxels being far less than background voxels in 3D CT images,this paper proposes a two-step segmentation framework to segment targets from coarse to fine.In the coarse segmentation stage,a training method assisted by priori-contour is proposed to train DMC-Unet,which improves the ability of model to learn features of edge contours.In the fine segmentation stage,a kidney tumor segmentation model(3D Res-Unet)of improved U-net based on 3D convolution is firstly proposed to effectively obtain the stereo spatial information for 3D images.Then based on the cross-entropy function,a new loss function(cmv-loss)is proposed,using the mean and variance of the gray level in the target area to guide the gradient update,which reduces the misjudgment rate of prediction results.Next,a method based on a pre-trained model of the abdominal image database is proposed.The target model 3D Res-Unet is trained by fine-tuning the pre-training model,making full use of the general features of the abdominal image that have been learned in the pre-training model.Finally,a post-processing algorithm based on morphological operations is proposed to further process the prediction results and effectively improve the segmentation results.Experiments show that the two-step segmentation framework proposed in this paper can achieve better segmentation of kidneys and kidney tumors.3)Aiming at the close relationship between the slices of the 3D CT image similar to the time-series features,an improved U-net network(LC-Unet)that combines Conv LSTM and 3D CNN is proposed.First,each slice of the three-dimensional image is regarded as a piece of information in time series.In the down-sampling phase,the features between the slices are extracted using the ConvLSTM structure,and then in the up-sampling phase,the 3D convolution structure is used to extract the three-dimensional spatial features of the data.Experiments show that the network achieves higher accuracy in segmenting kidney tumors in 3D CT images.