| Part 1 Diagnostic value of Spiral CT Multi parameter measurement on tumor deposition in colorectal cancerObjective To evaluate the value of multi-parameter Computed Tomography(CT)in Tumor Deposition(TD)in Colorectal cancer(CRC),and to determine the best diagnostic parameters.Materials and methods Retrospective analysis was performed on 45 pathologically confirmed TD lesions and 45 metastatic Lymph Node(MLN)lesions that could be identified on CT.All patients received plain abdominal spiral CT scan and arteriovenous phase enhanced scan before surgery.Multiple parameters of TD and MLN beside CRC were measured,including the lesion’s long diameter line,short diameter line,average diameter,ratio of long and short diameter line,plain CT value,arterial CT value and venous CT value.The multiparameter values and diagnostic efficacy of the two groups were compared.Results The long diameter line,short diameter line and mean diameter(1.29±0.40cm)(1.11(0.74~1.21)cm)(1.185(0.8~1.28)cm)of the TD group were all > those of the MLN group(1.14 ± 0.64cm)(0.93(0.85~1.11)cm)(0.97(0.88~1.21)cm),and the ratio of length to diameter(1.08(1.07~1.13)was < those of the MLN group(1.20(1.15~1.20).TD group plain CT value,arterial enhanced CT value,venous enhanced CT value(25.64±16.8 Hu)(28.53±19.9 Hu)(47.4±23.04 Hu)were all >MLN group(15.44±6.95 Hu)(20.08±11.85 Hu)(43.66±16.18Hu),and the differences of plain CT value,length to diameter ratio and arterial enhanced CT value were statistically significant(P < 0.05).The AUC of the subject operating characteristic(ROC)curve was > 0.716,among which the CT value of plain scan,the ratio of length to diameter,and the enhanced CT value of arterial phase were the best,with the AUC of 0.893,0.825,0.716,the optimal cut-off point of 18.5HU,1.134,31.50 HU,the sensitivity of 84.44%,75.56%,51%,and the specificity of 77.78%,80%,and100%,respectively.Conclusion There was no significant difference between TD group and MLN group in terms of multi-parameter measurement in terms of medium long diameter line,short diameter line,average diameter and venous phase enhancement value.There were statistically significant differences in the CT value,the ratio of length to diameter and the enhancement value of arterial CT.The AUC of CT value,the ratio of length to diameter and the enhancement value of arterial CT were all > 0.716,indicating a better diagnostic efficacy,suggesting that multi-parameter measurement of spiral CT is of great value in in identifying CRC para-TD and MLN.Part 2Preliminary analysis of the value of CT texture analysis and machine learningin differentiating tumor deposition and metastasis lymph node in colorectal cancerObjective To explore the diagnostic value of CT image 3D texture analysis(TA)and Machine Learning(ML)in discriminating CRC TD and MLN.Materials and methods A retrospective analysis of 45 CRC para-crc TD lesions and 45 MLN lesions confirmed by surgical pathology and identifiable on CT images was conducted.All abdominal CT plain scan and dual-phase enhanced arteriovenous scan were performed within one week before the operation.Ma Zda(Version 4.6)software was used for 3D TA of iv phase enhanced axial images,and 3D texture parameters of each lesion were extracted,including 794 texture feature parameter values in six categories:3D histogram,3D GLCM,RUN,absolute gradient(GRA),ARM and WAV.Texture feature selection method using software built-in fair coefficient(Fisher),the joint probability of classification error average correlation coefficient(POE + ACC)and mutual information(MI)and the combination of the above three methods(FPM),and by the built-in bl1 module of raw data analysis(RDA),principal component analysis(PCA),linear classification analysis(LDA)and nonlinear classification analysis(NDA)four classification methods to classify these texture parameters statistical analysis,the texture feature parameters of diagnosis in the form of misjudgment rate shows.The above were selected 30 of the optimal method of FPM texture parameters of principal component analysis(PCA)feature value dimension reduction,extracting characteristic root > 1 choice corresponding principal components,the characteristics of the selected are worth to postoperative pathological diagnosis,and 90(TD lesion45,45 MLN lesions)samples were randomly divided into training set and test set,70% of the samples for training set training machine learning model,30% of the sample do the test set,characteristics after the use of dimension reduction and Python program set up random forest,decision tree,bayesian,logistic regression of four kinds of commonly used machine learning model,The accuracy of the model was verified,and the ROC curve and AUC were obtained.SPSS 22.0 software was used to statistically compare the statistical differences between 30 optimal texture parameters of TD and MLN,and Med Calc 15.8 software was used to analyze the ROC curve of texture parameters with statistically significant differences and calculate AUC,analyze the values of lesion texture parameters in the two groups and compare their diagnostic efficacy.Results In the selection of texture features in TD group and MLN group,the misdiagnosis rates of Fisher coefficient,POE+ACC,MI and FPM in the identification of the two types of lesions were 7.78% ~ 22.22%,10.00% ~ 31.11%,6.67% ~ 16.67%and 4.44% ~ 18.89%,respectively.Among the characteristic statistical methods,the misdiagnosis rate of NDA for the two lesions(4.44% ~ 10.00%)was lower than that of RDA(15.5% ~ 26.67%),PCA(16.67% ~ 31.11%)and LDA(10.00 ~30.00%).Based on the three-dimensional texture features of spiral CT vein enhanced images,the misjudgment rate of nonlinear classification analysis(NDA)was the lowest,among which the misjudgment rate of FPM combined with nonlinear classification analysis(NDA)was the lowest,which was 4.44%.Among the 30 optimal texture parameters of TD group and MLN group,5 3D histogram parameters,10 3D grayscale symbiosis matrix,10 3D run-length matrix,3 absolute gradient and 2autoregressive models were statistically significant(P<0.05).The accuracy of the random forest method established by extracting the principal components of 30 optimal texture parameters,decision tree,bayesian and logistic regression ML models was 81.48% ~ 93.33%,with > 81.48% on average.AUC was 0.76 ~ 0.87,>0.76.Bayes’ accuracy rate was 93.33%(25/27)and AUC was 0.87.In the TD group and MLN group,30 optimal texture feature parameters were selected by FPM method,which were statistically significant.In the 30 optimal texture parameters for the subject working characteristic curve analysis,there are 15 features of AUC>0.70,among which,there are 10 AUC in the range of 0.70-0.799,4 in the range of 0.8-0.899,and 1 in the range of >0.9.The AUC in 3D histogram is at the highest of 90%grayscale percentile 3D,with AUC=0.947,specificity of 84.4% and sensitivity of97.8%.In the gray histogram,the highest AUC was S(0,5,5)entropy,with AUC=0.777,specificity of 68.9% and sensitivity of 75.6%.In the run-length matrix,the highest AUC was 135° long run-length compensation,with AUC=0.878,specificity of 91.1% and sensitivity of 77.8%.In the autoregression model,the highest AUC was Teta2,with AUC=0.756,specificity 73.3% and sensitivity 77.8%.The AUC of all texture parameters within the absolute gradient was lower than 0.35,indicating poor diagnostic efficiency.Among the 794 texture feature parameters in six categories,the highest AUC is the 90% grayscale percentile 3D in 3D histogram,with AUC=0.947,specificity of 84.4% and sensitivity of 97.8%.Conclusion1.Based on the 3D texture features of spiral CT venous phase enhancement images,nonlinear classification analysis(NDA)was adopted in the statistical methods of texture feature selection and feature classification in TD group and MLN group with the lowest misjudgment rate,among which FPM combined with NDA had the lowest misjudgment rate,with a misjudgment rate of 4.44%.2.Among the 4 mainstream ML models established based on the 30 optimal texture feature principal components selected by the extraction FPM method,the accuracy was 81.48% ~ 93.33%,all> 81.48%.AUC was 0.76 ~ 0.87,all> 0.76,indicating good accuracy and diagnostic efficiency.The accuracy rate of the bayesian algorithm is 93.33%(25/27),and the AUC is 0.87 at most,which is higher than the fitting effect of other three models,indicating that the bayesian algorithm has more excellent classification ability than other models.3.Based on spiral CT venous phase 3D texture method can be improved by the CRC TD and MLN diagnosis and differential diagnosis to provide a good quantitative objective basis,six categories a total of 794 texture feature parameter values,the highest AUC for 90% of the 3 d histogram gray percentile 3 d,the AUC = 0.947,84.4%,sensitivity was 97.8%,the prompt has great value in clinical application,help to improve N staging of CRC and provides more accurate imaging diagnostic information. |
PDF Full Text Request