| Bladder cancer is a common malignant tumor of the urinary system,mostly commonly found in middle-aged and elderly men.Lymph node metastasis is the process by which cancer cells proliferate and invade the surrounding tissues and disseminate to lymph nodes via lymphatic vessels and blood vessels,which is one of the key steps in the development of bladder cancer into malignant tumor.Clinically,lymph node metastasis is an important basis for assessing the stage of bladder cancer,choosing treatment plan and prognosis.Imaging is currently the main tool for bladder cancer lymph node diagnosis,including computed tomography(CT)imaging,magnetic resonance imaging(MRI)and ultrasound imaging.Contrast-Enhanced Computed Tomography(CECT)can effectively differentiate between normal and tumor tissues by enhancing the image of tumor location through injection of contrast agent into the patient,and is recommended by bladder cancer guidelines as the first choice for examination of lymph node metastasis in bladder cancer patients.However,due to the influence of imaging physicians’ experience and subjective visual perception,as well as objective factors such as noise and artifacts of imaging equipment,preoperative imaging is prone to false positive,which exposes some patients to the risk of insufficient staging and affects the formulation of optimal surgical strategies.There is an urgent clinical need for sensitive and accurate preoperative lymph node metastasis prediction tools to assist physicians in decision-making.With the continuous development of artificial intelligence in the field of cancer diagnosis,the great potential shown by radiomics technology and deep learning methods in the preoperative diagnosis of bladder cancer.The project team and the team from Sun Yat-sen University and other teams have carried out preliminary research on bladder cancer grading and staging,recurrence prediction and tumor segmentation,and preliminary research on lymph node metastasis prediction models based on radiomics methods,but the existing research only considered the tumor area,and used specific manually selection image features,failed to consider the changes in the surrounding tumor and lymph nodes regions,and did not fully explore the deep information in the image.To address the challenges and problems of preoperative diagnosis of lymph node metastasis in bladder cancer,this thesis uses preoperative CECT data and corresponding messenger Ribonucleic Acid(m RNA)data and clinical diagnostic information of 80 bladder cancer patients obtained from the Cancer Imaging Archive database and the Cancer Genome Atlas Project database to conduct the following three studies:(1)To address the problem of low sensitivity prediction of lymph node metastasis in preoperative bladder cancer,an intratumoral-peritumoral-based local imaging-omics prediction model was constructed.The peritumoral tissue of bladder cancer patients reflects the tumor microenvironment,which contains a large number of capillaries and lymphatic vessels as well as immune cells and various signaling molecules,etc.,and may contain key features associated with lymph node metastasis.In this thesis,based on the extraction of image features in the tumor region,we further extracted the image features in the peritumor tissues.A combination of variance method and recursive feature elimination method based on support vector machine is used for feature screening to reduce the risk of model overfitting while removing redundant features.Finally,a lymph node metastasis prediction model for bladder cancer was constructed based on support vector machine,Gaussian plain Bayesian and logistic regression algorithms.The experimental results showed that the area under the curve(AUC)of the working features of the average subject for the 5-fold cross-validation could be improved from 84.7% to 89.3%,and the sensitivity could be improved from 70.0% to 85.8%,with substantial improvement in both AUC and sensitivity,indicating that the imaging features in the peri-tumor tissue are important for improving the sensitivity of preoperative prediction of bladder cancer lymph node metastasis The AUC and sensitivity were significantly improved,indicating that the imaging features in peritumoral tissues are important for improving the sensitivity of preoperative prediction of bladder cancer lymph node metastasis.(2)A lymph node metastasis prediction model based on joint local-global image feature analysis was proposed to address the problems of insufficient image information incorporated in the small data set model.The model features include: 1)using the powerful learning capability of deep residual networks to obtain deep image features related to bladder cancer lymph node metastasis from the global by fine-tuning pre-training weights and model parameters and structure;2)using Gradient-weighted Class Activation Mapping(Grad-CAM)to extract the gradient information of the last convolutional layer and display the regions in the image that have the greatest influence on the classification result,to improve the decision process and the interpretability of the deep features;3)two specialized fully connected layers are designed to compress the output dimension of the deep features and reduce feature redundancy;4)the intra-and peri-tumor local image histology features are fused with the deep learning global image features to form a new joint feature matrix,and multiple machine learning methods are used to Multiple machine learning methods are used for joint prediction analysis.The experimental results based on deep global image features using 10-fold cross-validation showed that the mean AUC of the model was 72.5%(95% CI: 0.667 ～ 0.774);the results of 5-fold cross-validation based on combined local-global image features showed that the mean AUC was further increased to 96.0% and the sensitivity was improved to 92.3%.In addition,Grad-CAM visualization analysis showed that the deep learning network focused mainly on tumor and lymph node suspicious regions,which was consistent with our hypothesis.The above study suggests that both local-global imaging features can complement each other and have greater potential for preoperative prediction of lymph node metastasis in bladder cancer.(3)To address the problems of unclear biological significance of imaging features and weak clinical interpretability of models,this thesis investigated the mapping relationship between driver pathways and imaging features related to lymph node metastasis in bladder cancer based on paired m RNA sequencing data.Genomics can explain the underlying mechanisms of tumorigenesis,progression and lymph node metastasis,but it is difficult to characterize the evolutionary trends and changes of tumors at the macroscopic level.Imaging genomics studies are expected to establish the correlation between the obtained imaging features and genomics data,and the main contents and methods include: 1)performing weighted gene co-expression network analysis on m RNA data to identify highly synergistic gene sets;2)performing gene set variation analysis to view the driver pathways associated with imaging features;3)performing correlation analysis between modules and imaging features to identify modules that are highly correlated with bladder cancer lymph node metastasis with highly correlated modules to obtain key driver pathways.The analysis results showed that the imaging features could be regrouped into six categories based on the correlation with signaling pathways,including metabolic response,immunoinflammatory response,cell function,tumor transformation and lymphatic vessel metastasis,angiogenesis and biosynthesis.The experimental results showed that the local-based first-order statistical and texture features and the global-based DL features were closely correlated with the eight most significantly associated LNM biological pathways in BCa patients.First-order,GLSZM,GLCM and DL were closely correlated with metabolic pathways,in which the intratumoral image features(f12,f17,f24)were closely correlated with metabolic pathways(lipid metabolism,The DL features in the global images(f47,f48,f65)were negatively correlated with metabolic pathways,and f40,f51,f66 were negatively correlated with metabolic pathways.GLSZM,NGTDM,GLDM,First-order,GLSZM and DL are closely associated with immune inflammatory response,enhanced or diminished with up/down regulation of related immune inflammatory response pathways;GLCM,GLRLM,GLCM,GLSZM and DL are closely associated with tumor transformation and lymphatic metastasis related pathways GLCM,GLRLM,GLCM,GLSZM and DL are closely associated with tumor transformation and lymphatic metastasis-related pathways,where global-based DL signatures are positively correlated with this pathway,and enhanced pathway activity is followed by enhanced signaling of these signatures.f22,f31,f36,f56 and f60 are associated with vascular and lymphatic vessel generation,such as TGF-β,KRAS,VEGF signaling that resist apoptosis and invasion pathways are associated.Imaging features associated with biosynthesis include GLDM,NGTDM,GLSZM,GLCM,First-order,GLRLM,and DL.With the above study,this thesis substantially improved the comprehensive performance of the preoperative bladder cancer lymph node metastasis prediction model,especially the sensitivity of prediction.Considering the prevalence of CECT clinical examination and the consistency of imaging protocols,it is expected to be clinically applied after subsequent multicenter validation.The image-gene mapping relationship obtained from the study also lays the foundation for further research on the relationship between macroscopic imaging features and microscopic molecules of bladder cancer lymph node metastasis. |
PDF Full Text Request