Differential Diagnosis Of Adrenal Lipid-poor Adenoma And Non-adenoma

Objective The purpose of this study was to use the uAI scientific research platform to segment the area of interest(ROI)for adrenal lipid-poor adenomas,quantitatively shrink the ROI of the maximum area and total volume obtained,and explore the diagnostic value of different area ROIs in CT imaging radiomics studies for lipid-poor adenomas.Methods According to inclusion and exclusion criteria,CT imaging data of 219 adrenal lipid-poor adenomas and 215 non-adenomas from medical center A and B were analyzed.In Medical Center A,a total of 44 cases of lipid-poor adenomas and 42 non-adenoma cases were matched based on their baseline characteristics utilizing propensity score matching.To obtain the maximum slice ROI(MSS-ROI)and total volume ROI(FV-ROI),the tumors were segmented using uAI.Further,four ROIs were automatically derived by contracting each ROI by 0～3mm.From each ROI,first-order texture features were extracted for analysis.Recursive feature elimination and Spearman correlation coefficient were employed to feature selection,while logistic regression was used to develop an image-based radiomics model using various ROIs.Comparative analysis was carried out to determine the most optimal ROI.The data in medical center A were randomly divided into training and internal validation sets with an 8:2 ratio,while the data in medical center B served as an independent external validation set.The best ROI was obtained by automatic shrinkage of MSS-ROI and FV-ROI using the uAI system,and the radiomics features of each ROI were extracted.The inter-rater correlation coefficients(ICC)were used to evaluate the repeatability of the radiomics features,and features with an ICC less than 0.75 were excluded.Recursive feature elimination and LASSO regression were utilized for reducing feature dimensionality,and logistic regression models were constructed based on different ROI segmentation methods.Using the area under the curve(AUC)of the receiver operating characteristic(ROC),along with sensitivity at 95%specificity,the diagnostic performance of the radiomics model was assessed.The sensitivity between models at 95%specificity was compared using the McNemar test.SHapley Additive exPlanations(SHAP)values were employed to interpret the contributions of radiomics features to the model.Results After propensity score matching,18 first-order features were extracted for each of the 8 ROIs of each tumor in Medical Center A.After feature selection,MSS-ROI0,MSS-ROI1,MSS-ROI2,and MSS-ROI3 retained 7,7,6,and 6 features,respectively,and the AUC range of the various radiomics models was 0.735～0.803.The sensitivity range at 95%specificity was 0.386～0.546,and the MSS-ROI3 radiomics model had the highest AUC and sensitivity at 95%specificity.In the total volume group,FV-ROI0,FV-ROI1,FV-ROI2,and FV-ROI3 retained 6,6,7,and 6 features,respectively,and the AUC range was 0.797～0.814,with a sensitivity range at 95%specificity of 0.477～0.523.The FV-ROI3 radiomics model had the highest AUC and sensitivity at 95%specificity.At both centers,each ROI extracted 104 unfiltered radiomic features.After feature selection,MSS-ROI0,MSS-ROI3,FV-ROI0,and FV-ROI3 were selected with 6,5,5,and 6 radiomic features,respectively.The corresponding AUC for each radiomics model ranged from 0.797～0.814,with FV-ROI3 having the highest AUC.Based on a 95%specificity threshold,the sensitivities for the training,internal validation,and external validation sets were in the range of 0.567～0.813,0.460～0.778,and 0.282～0.603,respectively,with the highest sensitivity achieved by the FV-ROI3 model in the external validation set.The McNemar test showed that only FV-ROI3 had statistically significant differences in sensitivity compared to MSS-ROI0,MSS-ROI3,and FV-ROI0(P<0.05).The SHAP analysis indicated that the shape_Maximum3DDiameter and first-order texture features had negative contributions to the diagnosis of adenoma,while high-order texture features and shape_sphericity had positive contributions.Conclusion The radiomics model utilizing MSS-ROI3 and FV-ROI3 demonstrates superior diagnostic performance for lipid-poor adenoma.Particularly,FV-ROI3 significantly enhances the model’s accuracy and specificity,underscoring its potential clinical value in early detection of lipid-poor adenoma.The application of SHAP values facilitates the quantification and visualization of the contribution of radiomics features to the model,enabling the identification of positive and negative effects of each feature and improving the model’s interpretability.