Convolutional Neural Networks For Classification And Segmentation Of Renal Tumors In CT Images

Kidney is an important part of human urinary system.It is mainly responsible for removing metabolites and absorbing useful substances such as water,protein and metal ions.Renal tumor is the second most common tumor in the urology department.There are many subtypes of renal tumors,most of which are malignant tumors.Because renal tumors are insensitivity to radiotherapy,resection is often used clinically.However,traditional radical nephrectomy has great damage to renal function.Recently,laparoscopic partial nephrectomy(LPN)has become the first choice in clinical practice.Although LPN can preserve the renal function of patients to the greatest extent,the feasibility of this surgical method mainly depends on the diagnosis of doctors.That means that there is no unified gold standard.The subtypes of tumors,the shape and location of tumors and other factors have important reference.For segmentation,most of the existing methods are based on the segmentation of kidney regions.For the classification of renal tumors,traditional machine learning methods are widely used.The aim of this thesis to construct a simultaneous segmentation and classification model of renal tumors based on convolutional neural network to assist doctors in diagnosis.Convolutional Neural Networks(CNNs)is a popular technology in recent years.It has been widely used in the field of computer vision,especially in classification,segmentation,detection and other tasks.Different from natural images,the gray value of each point in medical images,such as CT imaging,has physical significance.Medical images also have various data forms,for instance,three-dimensional,four-dimensional and sequence.When neural network is applied to medical image processing,it is necessary to take into account the characteristics of medical images for improvement.For three-dimensional medical images,there are two overwhelming processing methods.The first one is based on two-dimensional CNNs and threedimensional results are integrated from two-dimensional results.The second one use threedimensional CNNs to get three-dimensional results directly.2D CNNs can have a deep network structure while the layer of 3D CNNs are limited by graphics memory.However,3D CNNs can learn the context information of the third dimension,which is more suitable for 3D data.In this thesis,both methods are used to segment and classify renal tumors,while three-dimensional CT images of renal tumors are also converted into two-dimensional images or threedimensional images according to the needs of two methods.Firstly,based on 2D CNN,a dual-task neural network named 2D SCNet is proposed in this thesis,which combines the segmentation and classification of renal tumors together.For the start of 2D SCNet,the segmentation network and classification network share the feature extraction layer.Then,the features obtained from the feature extraction layer are sent to the segmentation and classification network for training together.The combination of segmentation and classification not only improves the effect,but also is more suitable for our task.Besides,in the segmentation network,a two-step segmentation strategy is used to further improve the results.Finally,the result of 2D SCNet is better than that of single task network.Because of the limitation of 2D CNNs,a small part of segmentation results of 2D SCNet are discontinuous in the third dimension,which can be improved by 3D CNNs.Based on 3D CNN,because of the limited GPU memory,this thesis adopts a different network structure from 2D SCNet.Imitating the traditional method of kidney tumor segmentation,3D SCNet segmented the area of kidney and the tumor,obtained region of interest(ROI)from the original image through the segmented mask,and sent ROI into a shallow classification network for training.This method eliminates the background,obtains the foreground,simplifies the classification process,achieves the end-to-end multi-task network,and ensures the accuracy of the results.Compared with a single task segmentation network,the segmentation region of 3D SCNet is more complete.The classification results of a single classification network based on real ROI are excellent,while the classification based on ROI generated by 3D SCNet achieves the same effect as a single classification network does.Compared with 2D SCNet,3D SCNet solves the problem of missing tumor and classifying the region of tumor wrong in 2D segmentation.Meanwhile,the segmentation results of 3D SCNet are more continuous.However,due to the limited number of layers in the network,3D SCNet performs worse than 2D SCNet in classification.