The Logistic Regression Analysis Of High-frequency Ultrasound In Differential Diagnosis Of Thyroid Nodules

Objective：to select significant ultrasonographic features from many ultra-sound appearances in differential diagnosis of thyroid nodules by applyingbinary logistic regression analysis and then to establish the regression modelMaterials and methods: before ultrasound-guided Tru-cut needle biopsy,apply high-frequency ultrasound to evaluate357thyroid nodules(245benignones and112malignant ones), including B-mode gray ultrasound (size, shape,longitudinal-transverse ratio, margin, posterior shadow, nodular texture,internal echogenicity, calcification, halo), color doppler appearances (vasculardistribution, resistance index) and ultrasound elastography. The pathologicdiagnosis from Tru-cut needle biopsy was considered standard of diagnosis.The pathologic results were referred to as dependent variable while all of theultrasound features mentioned above were referred to as independent variables.After forward stepwise regression and evaluation of regression coefficientsby wald X~2test, the regression model was established. Use likelihood-ratio testto estimate the goodness-of-fit of the regression model. Assess thedifferentiating ability of the regression model by receiver operatingcharacteristic (ROC). p<0.05was considered to indicate significance.Results: binary logistic regression model was established and fourstatistically significant independent variables were selected after forwardstepwise regression: posterior shadow, internal echogenicity, calcification,ultrasound elastography. The regression model was that [Logit(P)=-3.184+1.904×posterior attenuation-0.700×posterior hyperechoic+2.239×remarkablehypoechogenicity+0.851×slightly hypoechogenicity+2.484×microcalci-fication-1.168×macrocalcification+1.061×strain ratio more than3.855]. Use likelihood-ratio test to estimate the goodness-of-fit of the regression model andX~2=253.113,P<0.01. So the logistic regression model was statisticallysignificant. Use the regression model to evaluate the357thyroid nodules in thisstudy,89.91%could diagnosed correctly with specificity93.88%andsensitivity81.25%if P>0.5was considered malignant and P≤0.5wasconsidered benign. AUC of ROC was0.946(95%confidence interval:0.920-0.972，P<0.01). So this binary logistic regression model was good at different-iating malignant thyroid nodules from benign ones.Conclusion1. ultrasound has an important effect on the examination and differenti-ation of thyroid nodules.2. we can select the significant ultrasonographic features in differentialdiagnosis of thyroid nodules by binary logistic regression analysis.3. compared to single ultrasound appearance, applying logistic regressionmodel is more accurate in differentiating thyroid nodules.