Application Of Bayesian Analysis In The CT Diagnosis Of Solitary Pulmonary Nodules

Objective:To explore the value of Bayesian analysis in distinguishing between benign and malignant solitary pulmonary nodules (SPNs)with CT,and hope to improve the management of SPNs with Bayesian analysis.Methods:Three hundred and fifty-two consecutive SPN cases (malignancy n=135,benignity n=217)were collected retrospectively to form the training set.Utilizing Bayesian analysis,the probability of malignancy in each SPN was calculated from the prior odds of malignant SPNs and the likelihood ratios of clinical and CT findings which were derived from the training set.SPNs with≥50%calculated probability were judged as malignancy and those with＜50%calculated probability was judged as benignity.On the test set(malignancy n=61,benignity n=71),Bayesian analysis was tested prospectively for its diagnostic validation and precision of predictive probability,compared with the performance of the two chest radiologists and two radiologic residents using routine diagnostic method.Results:(1)The prior odds of malignant SPNs is 0.61; (2)Deduced from the hierarchy of likelihood ratios,the more significant features for malignant SPNs were vacuole sign,short spiculation,deep lobulation etc.,and those for benign SPNs were a benign pattern of calcification,net nodule enhancement＜20HU,polygonal contour,and so on;(3)The sensitivity,specificity,accuracy of Bayesian analysis for the training samples were 88.9%,93.1%and 91.5%respectively.On the test set,the sensitivity,specificity,accuracy,positive predictive value, negative predictive value of Bayesian analysis were 88.5%,85.9%, 87.1%,84.4%and 89.7%respectively,its accuracy showed no statistically significant with chest radiologist A(80.3%,x~2=2.37,P=0.122) and B(79.5%,x~2=3.12,P=0.076),and was higher than radiological residents C(74.2%,x~2=7.05,P=0.012)and D(74.2%,x~2=6.56,P=0.009); (4)As for non-metastatic SPNs,the area under the receiver operating characteristic curve(Az)of Bayesian analysis was 0.957,which is higher than chest radiologists(Az=0.886,P=0.003)and radiological residents (Az=0.845,P=0.000);(5)The Brier score was 0.099 for Bayesian analysis,0.140 for chest radiologist A,0.137 for chest radiologist B, 0.154 for radiological resident C,and 0.179 for radiological resident D; (6)Excluding the solitary metastasis(n=11)misclassified,the false negative rate of Bayesian analysis was 1.0%(5/484)when SPNs with＜20%estimated probability of malignancy were considered.Conclusion:(1)The likelihood ratios of clinical and CT findings associated with SPNs can be used to guide the interpretation of CT images for physicians in daily work;(2)Bayesian analysis is an effective diagnostic aid which can enhance physicians' capacity of differentiating benign from malignant SPNs,especially for less experienced physicians; (3)Bayesian analysis has high precision in forecasting the malignant probability of SPNs,and it is expected to provide a quantitative reference indicator in clinical decision-making for SPNs.