| Objective:At present,Adrenal Venous Sampling(AVS)is the gold standard for the lateral diagnosis of primary aldosteronism(PA),but there are many shortcomings in AVS.The purpose of this study is as follows.(1)To seek a new index by analyzing the general imaging features and imaging omics features of PA patients’CT data using AVS as the gold standard for lateral diagnosis of PA and provide evidence for a new lateral diagnosis of PA based on CT.(2)To build a PA lateral diagnosis model by analyzing of general imaging features and imaging omics features using AVS combined with the treatment effect after surgical resection of the adrenal gland on the dominant side determined by AVS as the gold standard of PA lateral diagnosis.(3)To construct the adrenal automatic segmentation model in CT based on improved U-Net.Methods:(1)253 PA patients who successfully underwent AVS in West China Hospital of Sichuan University before June 2019 were enrolled in this study.177 cases were enrolled in the training group and 76 cases were enrolled in the test group.All of these patients underwent adrenal thin-slice CT,and their adrenal limb width and volume were measured respectively,and the Left-versus-right adrenal width ratio(L/Rw)and Left-versus-right adrenal volume ratio(L/Rv)were calculated.L/Rv receiver operating curves(ROC)for confirmed Left-PA and Right-PA were drawn,and cut-off values with 100%specificity were obtained,which were then validated in the test group.The adrenal imaging omics features of these patients was extracted by the software Pyradiomic,and the classified models were trained with Kernel Support Vector Machine(KSVM)and Random Forest(RF)machine learning algorithms and verified in the test group.Confuse matrix and ROC were used to evaluate the predictive performance of models,and the evaluation indexes included accuracy,Kappa and Area Under Curve(AUC).(2)In this study,77 patients with unilateral adrenalectomy were selected from253 patients in partⅠwho were clearly indicated for unilateral adrenalectomy according to AVS results and followed up for more than 6 months after lesion resection.Among them,54 cases were enrolled into the training group and 23 cases into the test group.All of these patients underwent adrenal thin-slice CT,and the adrenal Regions Of Interest(ROI)were manually delineated.The software Pyradiomics was used to extract the adrenal imaging omics features of these patients,and the classified models were trained with KSVM and RF machine learning algorithms and verified in the test group.Confuse matrix and ROC were used to evaluate the predictive performance of models,and the evaluation indexes included accuracy,Kappa and AUC.(3)In this study,adrenal CT imaging data of 500 patients(included PA,hypertension or essential hypertension complicated with adrenal nonfunctional tumor/adrenal thickening,Cushing’s syndrome,pheochromocytoma,aldosterone and cortisol co-secretory tumor,adrenocortical hypofunction,secondary aldosteronism,multiple endocrine adenomatosis)admitted to West China Hospital before June 2019 were randomly selected.Firstly,format conversion and manual segmentation of data images were carried out.Format conversion was to convert Digital Imaging and Communications in Medicine(DICOM)to Neuroimaging Informatics Technology Initiative(NIFTI)format by using converted DCM2NIIX(Web);ITK-SNAP(www.itksnap.org)software was used to manually delineate the adrenal boundary layer by layer and semi-automatically synthesize the adrenal gland in the venous phase image of NIFTI thin slice CT.In the computer with 64-bit Linux operating system installed,the image preprocessing and uu Net model training were completed based on the Py Torch open source neural network library under Python3.8.3,and the model verification was further completed.Visual comparison and Dice similarity coefficient were used to evaluate the automatic segmentation performance of the model.Results:(1)1).The left adrenal limb width and right adrenal limb width of PA patients were 0.45±0.11cm and 0.39±0.10cm,respectively,and the total limb width was0.84±0.19cm;The left and right volumes were 5.31±2.87 and 4.22±2.44 cm~3,respectively,and the total volume was 9.53±4.24 cm~3.2).The optimal cut-off value of L/Rv of 2.211 can predict partial Left-PA without misdiagnosis,and the optimal cut-off value of L/Rv of 0.656 can predict partial Right-PA without misdiagnosis.3).Based on AVS results,the texture feature model of adrenal CT image trained by KSVM algorithm after Laplacian of Gaussian(Lo G)processing had a better prediction performance for Left-PA,with an accuracy of 88.2%,a Kappa value of0.754,and an AUC of 0.930.The accuracy,Kappa,and AUC of the optimal left PA prediction model constructed by RF algorithm were 68.4%,0.322,0.659,respectively.4).The texture feature model of original adrenal CT image trained by KSVM algorithm had a better prediction performance for Right-PA,with an accuracy of88.2%,a Kappa value of 0.709,and an AUC of 0.964.The accuracy,Kappa,and AUC of the optimal prediction model for right PA constructed by RF algorithm were 80.3%,0.382 and 0.657,respectively.5.Based on the original adrenal CT image texture feature model trained by KSVM algorithm,the accuracy and Kappa values of left,right and bilateral PA(triclassification)prediction were 63.2%and 0.429,respectively.The accuracy and Kappa values of the optimal prediction model of the triclassification PA constructed by RF algorithm were 51.3%and 0.25,respectively.(2)1).Among PA patients who underwent unilateral adrenalectomy according to AVS results,the complete biochemical success rate was 88.3%(68 cases),the partial biochemical success rate was 9%(7 cases),and the absent biochemical success rate was 2.7%(2 cases).2).Using AVS and follow-up results as the lateral diagnostic criteria for PA,the prediction accuracy and Kappa value of the adrenal CT texture feature model trained by KSVM for the three categories of PA(left,right and bilateral)were 69.6%and 0.474,respectively.The accuracy and Kappa of the best prediction model of the triclassification PA constructed by RF algorithm were 65.2%and 0.395,respectively.The prediction model accuracy,Kappa value and AUC of optimal Left-PA constructed by KSVM were 78.3%,0.563 and 0.871,respectively.The accuracy,Kappa,and AUC of the optimal left PA prediction model constructed by RF algorithm were 69.6%,0.388,and 0.693,respectively.The prediction model accuracy,Kappa value and AUC of optimal Right-PA constructed by KSVM were 87.0%,0.704 and0.893,respectively.The accuracy,Kappa,and AUC of the optimal prediction model for right PA constructed by RF algorithm were 78.3%,0.506,and 0.763,respectively.(3)From the perspective of quantitative analysis,the adrenal automatic segmentation model based on nnU-Net has better segmentation accuracy,and Dice similarity coefficient was 0.891.From the perspective of qualitative analysis,visual observation of manually segmented and automatically segmented images had a high degree of similarity.The boundary between adrenal glands and surrounding tissues could be found more accurately.In terms of efficiency,the average time of manually segmenting bilateral adrenal glands in one patient was 2.5 hours,while the average time of automatically segmenting bilateral adrenal glands in one patient was 36 seconds.Conclusion:(1)Unilateral PA prediction model derived from adrenal CT texture features has better prediction performance.The classification performance of KSVM is better than that of RF.(2)With AVS and the results of follow-up after adrenal surgery on the dominant side determined by AVS as the lateral diagnostic criteria for PA,the prediction model of Right-PA derived from adrenal CT texture features has better predictive performance.The predictive models of Left-PA and tripartite PA have limited predictive power.The classification performance of KSVM is better than that of RF.(3)The adrenal automatic segmentation model based on nnU-Net has good segmentation performance and can be applied to the adrenal automatic segmentation in CT in the future. |
PDF Full Text Request