| [Objectives]To evaluate the feasibility and application value of CT radiomic model in differential diagnosis of abdominal lymph node lesions caused by different causes(lymphoma,metastatic lymph nodes and inflammatory reactive lymph nodes)and classification of benign and malignant lymph nodes.[Materials and Methods]The CT images and pathological data of 146 patients were retrospectively collected.Among them,27 patients(70 lymph nodes in total)underwent surgical resection due to abdominal visceral malignant tumor,and all lymph nodes around the tumor were proved to have metastasis by pathology after surgery;The control group consisted of 36 cases(78 lymph nodes in total)with no metastasis of lymph nodes around the tumor confirmed by postoperative pathology;33 cases of lymphoma were confirmed by biopsy of lymph nodes,94 lymph nodes in total;50 cases(91 lymph nodes in total)were confirmed as inflammatory bowel disease by surgery or gastrointestinal endoscopy,without virus and other pathogen infection,and with reactive enlargement of peri-intestinal lymph nodes.Lymphoma and metastatic lymph nodes were regarded as malignant lymph nodes group,169 in total;The control group and inflammatory reactive lymph nodes were regarded as benign lymph nodes group,totaling 164.After uploading the patient’s CT venous phase original image and thin-layer reconstruction image to the Radiology Cloud Platform(Radcloud platform,Huiying Medical Technology Co.,Ltd.),the area of interest(ROI)is delineated along the edge of the lymph nodes on the whole layer of the thin-layer image.First,the benign and malignant pathological classification is labeled,and then the four(lymphoma,metastatic lymph nodes,control group,inflammatory reactive lymph nodes)pathological classification is labeled,A total of 333 lymph nodes from 146 patients were randomly divided into a training set and a verification set at a ratio of 7:3.The features with the highest correlation were selected by using the variance threshold method,the SelectKBest method,and the minimum absolute contraction and selection operator(LASSO)method in turn.The best features obtained after screening were used for machine learning,and k-nearest neighbor(KNN)Support vector machine(SVM),extreme gradient boosting(XGBoost),logistic regression(LR)and random forest(RF)algorithms are used to construct classification models for the above pathological labels.Each classification model is verified and compared separately,and the area under the operating characteristics curve(AUC)and sensitivity of the subject are calculated The specificity and other parameters were used to evaluate the diagnostic efficacy of the radiomics model.[Results]Regarding the differential diagnosis of benign and malignant lymph nodes with the label of benign and malignant,the radiomics model built by LR classifier has the best diagnostic effect.A total of 23 histological features have been selected.The area under the ROC curve(AUC value)of the training set is 0.895(95%CI:0.841～0.949),sensitivity is 80%,and specificity is 78%;The AUC value of the validation set is 0.792(95%CI:0.704～0.880),the sensitivity is 69%,and the specificity is 74%.Regarding the differential diagnosis of lymphoma and metastatic lymph nodes with the label of pathological type,we used LR classifier to construct a radiomics model based on venous phase images to achieve the best diagnostic effect.A total of 23 histological features were screened out.The area under the ROC curve(AUC value)of the training set was 0.986(95%CI:0.942～1.000),sensitivity 94%,and specificity 94%;The AUC value of the validation set is 0.924(95%CI:0.819～1.000),the sensitivity is 86%,and the specificity is 79%.With regard to the differential diagnosis of lymphoma and inflammatory reactive lymph nodes,the radiomics model constructed by SVM classifier achieved the best diagnostic effect.A total of 25 histological features were selected.The area under the ROC curve(AUC value)of the training set was 0.993(95%CI:0.954～1.000),sensitivity 95%,specificity 95%;The AUC value of the validation set is 0.983(95%CI:0.907～1.000),the sensitivity is 90%,and the specificity is 91%.About the differential diagnosis of metastatic lymph nodes and the control group,the radiomics model constructed by KNN classifier achieved the best diagnostic effect.A total of 14 histological features were selected.The area under the ROC curve(AUC value)of the training set was 0.857(95%CI:0.774～0.940),sensitivity 70%,specificity 78%;The AUC value of the validation set is 0.760(95%CI:0.634～0.886),the sensitivity is 71%,and the specificity is 76%.About the differential diagnosis of metastatic lymph nodes and inflammatory reactive lymph nodes,the radiomics model built by LR classifier achieved the best diagnostic effect.A total of 10 histological features were selected.The area under the ROC curve(AUC value)of the training set was 0.865(95%CI:0.797～0.933),sensitivity 74%,specificity 76%;The AUC value of the validation set is 0.836(95%CI:0.767～0.905),the sensitivity is 74%,and the specificity is 77%.[Conclusion]The radiomic model based on CT can effectively differentiate benign and malignant abdominal lymph nodes,as well as lymphoma,metastatic lymph nodes and their control,inflammatory reactive lymph nodes,etc.,which provide accurate information for clinical treatment. |
