Renal Cancer Classification Of PET/CT Imaging Based On Deep Learning

Renal cancer is one of the common disease of the urinary system with the increase of the morbidity and mortality year by year,which seriously threatens the human health.The improvement of the diagnosis accuracy of renal cancer is helpful for the accurate pre-treatment staging,which is very important for the formulation of treatment plan and judgment of prognosis.Positron emission tomography/computed tomography(PET/CT)is a commonly used clinical diagnosis method of renal cancer.It can not only locate the lesions accurately through CT imaging,but also obtain the metabolic capacity and function information of lesions from PET imaging.Clinical analysis and judgment on PET/CT imaging is affected by various factors.With the continuous development and progress of computer technology,it receives growing attention to design a renal cancer classification network with high accuracy and robustness,which realizes classification and recognition of renal cancer accurately and efficiently.This paper conducts a study on renal cancer classification algorithms comprehensively and thoroughly.The powerful feature extraction capabilities of convolutional neural network(CNN)is used to fully excavate highly abstract features of complex PET/CT imaging.The transition from the general features in the source domain to the target domain is achieved by transferring the parameters of the pretrained CNN of the source domain dataset while reducing the need for training data and time.The feature fusion method is proposed to fuse the features extracted by CNNs of PET and CT imaging,which make full use of features and further boost renal cancer classification performance.The main research contents in this paper are listed as follows:Firstly,the basic theory of deep convolutional neural networks especially the theory,training and optimization process of the structure are researched thoroughly.After that,the difficulties and limitations of deep learning in medical imaging classification are analyzed and the characteristics and advantages of transfer learning are subsequently introduced in detail.For limitation of the small sample size,the regularization method to strengthen the ability of model to resist overfitting is analyzed from three aspects of data,model and label,which provides theoretical basis for building renal cancer classification models based on deep learning for PET/CT imaging in the following parts.Secondly,renal cancer classification of PET/CT imaging based on feature fusion is completed.The characteristics of PET/CT imaging especially the significance in imaging diagnosis are studied thoroughly.The DICOM file storage format and general preprocessing methods of PET/CT medical imaging are further analyzed.The traditional medical imaging classification methods are studied in detail.Then,by transferring the parameters of the pre-trained CNN of the Image Net dataset.After that,the features of PET imaging and the CT imaging are fused,and a classification model is constructed to classify the features of the integrated expression.Finally,renal cancer classification of PET/CT imaging based on 3D CNN is studied.The spatial information between slices as part of the diagnosis basis during clinical diagnosis of radiologists is taken into account.A 3D CNN is designed in order to take advantage of the information between adjacent layers of PET/CT imaging sequence.Features of sequence level extracted by 3D CNN are more robust and more suitable for medical image classification tasks.In this paper,more accurate classification results are obtained on three datasets.