Imaging-Aided Diagnosis Of Renal Diseases Based On Machine Learning

The precise treatment of diseases in modern medicine requires doctors to fully extract the individual characteristics of the disease before treatment.For kidney stones and kidney tumors,which are two common kidney diseases,the benefits of extracting individual characteristics before treatment are as follows:1)In-vivo determination of kidney stones is fundamental in directing urologists to select appropriate medical management techniques and improve the prognosis;2)The accurate segmentation of kidney tumors is of great significance in optimizing the surgical path and resecting the lesion precisely,and the classification of kidney tumors helps to provide support for doctors’ diagnosis and treatment,which will avoid the patients from receiving unnecessary surgery or drug treatment.In recent years,with the development of computer-aided diagnosis,analysis of medical images with machine learning can provide better diagnosis information for doctors.Therefore,the research work in this thesis is mainly focused on the application of machine learning in composition analysis of kidney stones and the segmentation and classification of kidney tumors based on CT images.The main research work and innovations of this thesis are as follows:(1)Multi-label classification fusion modeling for in-vivo determination of kidney stones based on machine learning.In this study,we retrospectively collected 252 patients with kidney stones,and sought to establish a novel fusion-based multi-label classification framework to flag each composition in stones.First,four categories of ESA(energy spectrum analysis)parameters extracted from a spectral CT analysis were utilized for modeling.Then,we employed 10 feature selection methods and therefore their combinations with the 25 multi-label classifiers finally resulted in 250 multi-label base classification models.The ESA parameters were fed into the miscellaneous multilabel base classification models,and performances of the base models were ranked.Finally,top-10 models among all the 250 base models were integrated to reach a consensus prediction via a multi criterion weighted fusion algorithm(MCWF)method.We compared the proposed MCWF with the top-10 base models and three benchmark model fusion methods.In addition,the ESA parameters were further analyzed in our study,from which the top-10 parameters were identified.Experiments and results show that our proposed fusion-based multi-label classification framework has demonstrated a satisfactory Accexam of 81.2%and 76.4%on the independent testing cohort 1 and 2.The top-10 ESA parameters associated with the stone composition identification included seven Zeff histograms and three material densities.It is observed that the top-10 ESA parameters have a good discriminating effect on stone composition after t-SNE visualization.(2)Research on kidney tumor segmentation and classification based on 3D U-Net.We sought to establish a 3D U-Net-based kidney tumor segmentation and classification model with CT images from multiple centers and multiple image preprocessing methods.First,the dataset was established by using the CT images of unenhanced phase(UP),corticomedullary phase(CMP),nephrographic phase(NP)and excretory phase(EP).Three methods were then used for image preprocessing,including crop,resample,and normalization.Finally,the preprocessed images were fed into 3D U-Net for training and evaluation.We compared the segmentation model with:1)model trained with CMP images only,2)models with different image preprocessing methods,respectively;and evaluated the tumor classification results in the training and test dataset.Preliminary results of the experiment verified the reliability of the model.In the future,we will study the input methods of CT images and the selection of basis network.