Design And Research Of Kidney Tumor CT Image Segmentation System Based On Convolutional Neural Network

Kidney tumors are one of the most common tumors in humans and the main treatment currently is surgical resection.The CT images are usually manually segmented by specialist surgeons as preoperative planning,but this can be influenced by the surgeon’s experience and skill and can be time-consuming.Due to the complex lesions and different morphologies of kidney tumors that make segmentation difficult,this paper proposed a Convolutional Neural Network-based automatic segmentation method for CT images of kidney tumors to address the most common problems of boundary blurring and false positive in tumor segmentation images.The method is highly accurate and reliable,and is used to assist doctors in surgical planning as well as diagnostic treatment,relieving medical pressure to a certain extent.The specific work is as follows.The Efficient Net V2-UNet segmentation model proposed in this paper contains three main components: a feature extractor,a reconfiguration network,and a Bayesian decision algorithm.Firstly,the Efficient Net V2 feature extractor with high training accuracy and high efficiency was selected as the backbone network for the tumor false positive phenomenon,and the shallow features such as location,morphology and texture of the tumor in CT images were extracted using downsampling.Secondly,the reconfiguration network was designed based on the backbone network,which mainly consists of transform blocks,deconvolution blocks,convolution blocks and output block to build an upsampling architecture,gradually recovering the spatial resolution of the feature map to fully identify contextual information,forming a complete encoding-decoding structure,and superimposing the feature map channels at all levels on the left and right sides to achieve multi-scale feature fusion,preventing loss of details and performing accurate segmentation of tumors.Finally,a Bayesian decision algorithm was designed for the edge blurring phenomenon of segmented tumors and cascaded on the output of the reconstruction network,combining the edge features of the original CT images and the segmented images for probabilistic estimation to increase the accuracy of the model edge segmentation.In this paper,the model was subjected to a tri-fold cross-validation experiment using the Ki TS19 kidney tumor dataset,and its performance was compared with seven mainstream models with multiple evaluation metrics in multiple dimensions.The experimental results show that the Dice coefficient of the segmented image similarity index of this model is 0.933 and Jaccard coefficient is 0.861,which are better than the mainstream UNet model by 0.030 and 0.064 respectively;the classification accuracy metric Precision is 0.958 and F0.5-Score is 0.945,outperforming the UNet model by 0.055 and 0.048 respectively;the image structure similarity index SSIM is 0.974,which is better than the UNet model 0.007.Secondly the ROC curve of the model in this paper forms an envelope around the six models,and the AUC area is 0.997,which is close to the optimal value,indicating that it has a good segmentation effect under most thresholds.The model was also validated by step-by-step experiments to improve the blurring of tumor segmentation edge and the false positive phenomenon.In addition,the model inference speed and the number of parameters are17.505 FPS and 21.261 M respectively.In summary,the model has high segmentation accuracy,good robustness and reliability,and meets the segmentation requirements.This paper used the Python3.8 language to develop a Py Qt5 framework-based CT image segmentation software for kidney tumors,to realise model training,generating datasets,and individual or batch image segmentation functions,to deploy theoretical research into practical applications,to achieve automated segmentation of medical images,and to save medical resources.