Research Of Kidney Tumor Segmentation Based On Deep Learning

Kidney tumors are mostly malignant tumors.It is one of the most common tumors of the urinary system.There are more than 400,000 new cases of kidney tumors every year,which seriously endangers people’s health.In clinical practice,surgical resection is often used to remove kidney tumors.Nephron Sparing Surgery is increasingly becoming an option for patients with kidney tumors.Nephron-preserving surgery refers to the removal of all kidney tumors while preserving part of the patient’s normal kidney as much as possible,which requires the location of the lesion to be segmented as accurately as possible.In traditional segmentation methods,doctors manually mark the location of kidneys and kidney tumors on CT images with the help of corresponding annotation software based on pathological knowledge and related experience,which consumes a lot of time and energy.In addition,since the accuracy of the labeling depends on the doctor’s subjective judgment and professional level,there is a lot of uncertainty.The kidneys and kidney tumors between different individuals are very different.For example,the pathological complexity and clinical manifestations of kidney tumors are different.Traditional segmentation methods have great limitations and are difficult to meet clinical requirements.Therefore,the study of automatic segmentation of kidney tumors has important application value.In recent years,with the development of artificial intelligence,deep learning has been widely used in image classification,segmentation,target detection and other fields.Taking medical image segmentation as an example,the deep learning algorithm uses bountiful data to train the model to learn the characteristics of the target adaptively without too much human intervention,and the process is simpler than traditional algorithm,which can effectively improve the accuracy and speed of medical image segmentation.Therefore,deep learning is more suitable for medical images with higher complexity and widely used in medical image segmentation.According to the above problems and background,the thesis has completed the following work:Firstly,this paper proposes a U-Net segmentation model based on hybrid of inter-layer information,which is applied to kidney and kidney tumor segmentation.The traditional U-Net model inputs a single CT slice for training,and there is no correlation between each slice.Therefore,the inter-layer information of the 3D image is ignored,which causes information loss and reduces the segmentation accuracy.The U-Net segmentation model that aggregates inter-layer information proposed in this paper performs 3D convolution on the model encoding end,and makes full use of the inter-layer information between adjacent slices to predict the middle-layer slices in a targeted manner.Secondly,this paper proposes an image segmentation model based on multi-scale fusion,which is applied to kidney and kidney tumor segmentation.Considering that when the up-sampling part of the decoding end is restored from low resolution to high resolution,some detailed information will definitely be lost,which will affect the final segmentation result.Therefore,this paper adds a multi-scale fusion unit to the model decoding end to fuse feature information of different scales..In addition,this model adds a residual unit at the encoding end,on the one hand,it combines the features of different levels of the network and reuses the learned shallow structure features;on the other hand,it suppresses the network overfitting.The image segmentation model based on multi-scale fusion fully mines detailed information,effectively solves the problem of small kidney tumors that are difficult to segment,and significantly improves the segmentation performance.