| ● Part I Exploration of new ideas and markers in the diagnosis of renal cell carcinoma● Chapter Ⅰ The clinical significance of the ratio of maximum to minimum tumor diameter in prediction of the pathology type of renal cell carcinoma before surgeryObjectives:Several imaging studies have described and investigated various methods used to distinguish pathological subtypes of renal cell carcinoma(RCC).However,there is no report about the value and significance of primary tumor morphology in identifying the pathological types of RCC.In order to improve the accuracy of preoperative imaging diagnosis and better guide clinical decision-making,this study aimed to evaluate the utility and value of the ratio of maximum to minimum tumor diameter(ROD)in predicting pathological subtypes of RCC,Methods:Data from 1661 patients with RCC at Cancer Hospital,Chinese Academy of Medical Sciences Cancer Hospital Chinese Academy of Medical Sciences(CHCAMS),who underwent surgery between January 2015 and December 2019,were retrospectively reviewed.The cut-off value for ROD was calculated using receiver operating characteristic(ROC)curve analysis.According to the ROD threshold,all groups were compared.Clinicopathological factors were also collected and compared.The Cox proportional hazard models were applied to identify predictive factors.Results:In the clear cell RCC(ccRCC)(n=1477)and non-ccRCC(n=184)groups,the optimal ROD cut-off value to predict ccRCC was determined to be of 1.201(sensitivity,90.7%;specificity,76.1%;and area under the ROC curve[AUC]0.827;p<0.001).In the non-ccRCC group,the cut-off value for ROD in predicting papillary RCC was 1.092(sensitivity,87.9%;specificity,40.5%;and AUC 0.637;p=0.003).Compared to patients with ROD<1.201,more in the ccRCC group exhibited tumors with an ROD≥1.201(14.2%versus 85.8%,respectively;p<0.001).Multivariate analysis of preoperative features revealed that ROD≥1.201 was an independent predictive factor for ccRCC(Odds ratio[OR]:3.061,95%CI[Confidence interval]:2.179-4.300;p<0.001).In addition,patients with ROD≥1.201 had higher percentages of Fuhrman grade III/IV(91.2%versus 8.8%;p=0.014),tumor necrosis(86.7%versus 13.3%;p=0.012),and sarcomatoid differentiation(90.6%versus 9.4%;p<0.001).Conclusions:ROD could be a novel indicator for preoperatively predicting histological type in patients with RCC and is an auxiliary method for imaging diagnosis.The ROD cut-off values of 1.201 and 1.092 were the most discriminative for ccRCC and papillary RCC,respectively.Moreover,ROD≥1.201 was associated with high Fuhrman grade,sarcomatoid features,and tumor necrosis.However,these findings need to be further confirmed by subsequent prospective studies.● Chapter Ⅱ Pre-operative risk factors analysis and scoring model establishment for the diagnosis of missed diagnosis of renal vein tumor thrombus in renal cell carcinomaObjectives:Previously,we have investigated long-term survival of RCC patients with renal vein tumor thrombus(RVTT)and found that some patients with RVTT were missed preoperatively.To improve the accuracy of preoperative diagnosis of RVTT,better guide clinical decision-making,and ensure the safety of perioperative period in RCC,we explore whether clinical features could be predictive indicators of missed diagnosis of RVTT.It is also the further in-depth and extension of the previous research.Methods:We retrospectively reviewed 128 RCC patients with RVTT between January 2000 and December 2015 at NCC/CHCAMS.Patients were classified and assigned to missed diagnosis group and unmissed diagnosis group according to whether the RVTT was missed preoperatively,and sex-and age-matched patients followed 1:1 statistical matching were selected as no tumor thrombus group which were performed nephrectomy in the same continuous period.Kaplan-Meier method was used to evaluate the survival outcome.The Cox proportional hazard models were applied to identify risk factors.Results:The missed diagnosis rate of RVTT in RCC was 30.5%(39/128).Missed diagnosis group patients were prone to have bigger proportion of tumor located in the middle pole(56.4%vs 28.2%,p=0.012),renal vein contrast agents filling not well(46.2%vs 23.1%,p=0.032)and collateral vessels(33.3%vs 7.7%,p=0.005)compared with no tumor thrombus group,while these features had no difference between missed diagnosis group and unmissed diagnosis group.In multivariate analysis,maximal tumor diameter,tumor located in the middle part,renal vein contrast agents filling insufficiently,and tumor with collateral vessels(odds ratio=1.22,1.35,1.25,1.22;and p=0.034,0.003,0.015,and 0.037,respectively)were independent predictors of missed RVTT diagnosis.Based on the final multivariable model,a missed diagnosis score was calculated.A missed-diagnosis score was presented as area under curve of 0.852(95%Cl:0.77-0.94,p<0.001).The sensitivity and specificity were 74.4%and 84.6%for patients with 3 score.Moreover,the missed diagnosis group had favorable prognosis,and tumor with collateral vessels was an independent prognostic indicator of poor overall survival time(hazard ratio[HR]:1.15,95%CI:1.02-1.47;p=0.025).Conclusions:The possibility of complicating tumor thrombus should be considered when there is pre-operative presence of tumor with large diameter,renal tumor in the middle part,renal tumor with collateral vessels,and renal vein contrast agents filling insufficiently.Patients with 3 points in missed-diagnosis scoring suggested a high possibility of missed RVTT diagnosis.● Chapter Ⅲ Research on the diagnosis of renal clear cell carcinoma by using artificial intelligence technology to identify pathological slicesObjectives:The use of computer science in medicine has become more and more extensive,in which artificial intelligence(AI)is the most prominent.AI has made breakthrough progress in the pathological research of solid tumors with the advent of digital pathology.However,according to reports in the literature,the application of AI in the diagnosis of RCC is still in its infancy.We first tried AI on the pathology of RCC to gain primary experience in AI-assisted RCC diagnosis.In this study,we aimed to establish an AI model for the diagnosis of clear cell RCC by collecting pathological slices,building digital whole slide images(WSI),artificial labeling,computer-aided diagnosis.We tried to explore the feasibility of diagnosing clear cell RCC with AI,and improve the efficiency of clinicians and possibly to use this technology to predict the prognosis of RCC in the future.Methods:We retrospectively collected pathological biopsy data of 95 patients with clear cell RCC receiving surgery who were admitted to CHCAMS from January 2016 to December 2016.All pathological slices conforming to the standards of the AI model were manually annotated first,and then the WSI was obtained through the scanner.The WSI was preprocessed to extract the region of interest(ROI).Pathological slices were divided into training set and test set,and the ratio of tumor slices and normal tissue slices in training set and test set was 3:1.In the training set,WSI with the highest pixel was divided into small images of fixed size for extraction(resolution:256×256).Positive and negative samples were randomly extracted.Two hundred and seventy small images were extracted from each tumor slice in the positive sample,and 550 small images were extracted from each pathological slice in the negative sample,which were used to train the model.Model training used convolutional neural network(CNN)and random forest model.The Receiver operating characteristic(ROC)curve was constructed to evaluate the accuracy of the model.Results:A total of 663 pathological slices from 95 patients with clear cell RCC were collected in this study.The mean number of slices for per patient was 7.6±2.7(range:3-17),including 506 tumor slices and 157 normal tissue slices.There were 200 tumor slices and 74 normal slices in the training set,and a total of 200,870 small images were extracted;there were 250 tumor slices and 63 normal slices in the test set,and a total of 39,211 small images were extracted.The test set was identified at the slice level.According to the CNN model and random forest model trained by the training set,11 pathological slices in the test set were identified as the false normal slices,and 6 pathological slices were identified as the false tumor slices.The total accuracy was 94.6%(296/313),the precision rate was 97.6%(239/245),and the recall rate was 95.6%(239/250).The generated probabilistic heat maps were in good consistent with the manually annotated pathological images.The ROC curve results showed that the AUC reached 0.9658(95%confidence interval:0.9603-0.9713),the specificity was 90.5%,and the sensitivity was 95.6%.Conclusions:Using AI in the diagnosis of clear cell RCC is feasible.The AI model of clear cell RCC established in this study has high accuracy.Preliminary results showed that this technique is worthy of further study.● Part Ⅱ Research on intelligent staging and prognostic learning model of renal cell carcinoma● Chapter Ⅰ Feasibility study on the diagnosis of T stage of renal cell carcinoma by using artificial intelligence technology to identify medical record text data and auxiliary software developmentObjectives:Artificial intelligence technology is booming in the medical field,but there are few research reports in the field of kidney cancer.Natural language,as a branch of artificial intelligence,can transform human language into a form of computer expression.The main purpose of this research is to explore the feasibility and accuracy of using natural language to diagnose T stage of renal cell carcinoma automatically,and develop an intelligent staging software to verify its accuracy.Thus,the pathological staging of RCC could be standardized,and it may be also conducive to the promotion at the grassroots level.Methods:This study retrospectively collected 200 RCC patients at CHCAMS from January 2018 to January 2020 as the training group,and selected 200 RCC patients with statistical matching gender,age,and pathological stage from January 2015 to December 2017 as the test group.Two artificial intelligence natural language processing methods including rule based template matching and conditional random fields were used to extract and analyze pathological text data.Python=3.6 and sklearn_crfsuite=0.3.6 were used to develop information extraction algorithms,and compare the prediction effects of the two methods.Microsoft Visual Studio Enterprise 2017(version 15.9.21)was used to write program code and develop software.Two-hundred RCC patients with complete pathology reports at CHCAMS from January 2020 to January 2021 were selected for verification of software results.Results:The accuracy rates of rule-based template matching and conditional random fields were 99.0%and 95.5%,respectively.In the method performance evaluation of the test group,the accuracy rate of the rule-based template matching was 99.0%,the recall rate was 99.0%,and the F1-score was 99.0%.The conditional random field method had an accuracy of 97.1%,a recall rate of 95.5%,and an Fl-score of 96.3%.Based on the T stage elements of RCC as the core code of the software including tumor diameter,perirenal fat,renal sinus fat,cancer thrombus and perirenal invasion,a software V 1.0 of automatic diagnosis of T stage of renal cell carcinoma was developed(registration number 2020SR1527729).The software for automatic diagnosis of T stage of RCC had an accuracy of 100%after 200 cases of testing.Conclusions:It is feasible for artificial intelligence to automatically diagnose the T stage of RCC by natural language processing,and the algorithm of rule-based template matching had high accuracy.The T stage software for automatic diagnosis of RCC is highly accurate and can be used in clinical guidance,standardization of RCC staging,which could be promoted.● Chapter Ⅱ Machine learning algorithms of artificial intelligence development to predict the prognosis of patients with renal cell carcinoma using the SEER databaseObjectives:Machine learning is an important subfield of artificial intelligence technology,while there are few literatures focusing on the research of machine learning in RCC.The main purpose of this study is to evaluate the applicability of machine learning algorithms for survival prediction of kidney cancer patients,and to compare the differences of machine learning methods,so as to provide theoretical support for future prognostic research on different databases or big data.Methods:RCC patients with the number of 55,334 met the inclusion were obtained from the Surveillance,Epidemiology,and End Results(SEER)database from 2004 to 2015.The data used three preprocessing methods including Standard,Normalise and Min Max Scaler.Six machine learning algorithms including support vector machine(SVM),Bayesian method,decision tree,random forest,neural network and extreme gradient boosting were applied to predict 5-year overall survival(OS).The cross-validation method evaluates the stability of the model,and the ROC curve and C index calibration curve are used to evaluate the accuracy of the model.Survival analysis were evaluated by the Kaplan-Meier method.Results:This study collected a total of 192,912 patients from the SEER database.Min Max Scaler is more conducive to SVM model training,and Bayesian method,decision tree,random forest and XGBoost were based on a single decision tree or multiple decision trees or data frequency enhancement methods,which were not sensitive to data preprocessing.Standard is suitable for neural network models.After deleting the missing data,each model had a poor data recognition effect.The accuracy of neural network and XGBoost is relatively higher,and the AUC is 66.6%and 67.0%,respectively.In the data set with deleted missing data and patients with survival time less than 5 years,the accuracy of random forest,neural network and XGBoost were high,with AUC of 80.8%,81.5%and 81.8%,respectively.After deleting only the missing tumor diameter and filling the missing data set with missForest,the accuracy of neural network and XGBoost are high,with AUC of 69.8%and 71.4%,respectively.After deleting the missing tumor diameter and patients with survival time less than 5 years,and filling the missing data with missForest,Random forest,neural network and XGBoost had high accuracy,with AUC of 84.1%,84.7%and 84.8%,respectively.Conclusions:Artificial intelligence machine learning algorithms could be used to predict the prognosis of RCC.In the machine learning model we established to predict the 5-year survival rate of patients with RCC,the neural network and XGBoost model had high accuracy.Given the limitations and complexity of data sets,machine learning can be used as an auxiliary tool to analyze and process larger data sets and the complex data. |
PDF Full Text Request