Classification decision tree algorithm assisting in diagnosing solitary pulmonary nodule by SPECT/CT fusion imaging / 西安交通大学学报·英文版

ABSTRACT

Objective: To develop a classification tree algorithm to improve diagnostic performances of 99mTc-MIBI SPECT/CT fusion imaging in differentiating solitary pulmonary nodules (SPNs). Methods: Forty-four SPNs, including 30 malignant cases and 14 benign ones that were eventually pathologically identified, were included in this prospective study. All patients received 99Tcm-MIBI SPECT/CT scanning at an early stage and a delayed stage before operation. Thirty predictor variables, including 11 clinical variables, 4 variables of emission and 15 variables of transmission information from SPECT/CT scanning, were analyzed independently by the classification tree algorithm and radiological residents. Diagnostic rules were demonstrated in tree-topology, and diagnostic performances were compared with Area under Curve (AUC) of Receiver Operating Characteristic Curve (ROC). Results: A classification decision tree with lowest relative cost of 0.340 was developed for 99Tcm-MIBI SPECT/CT scanning in which the value of Target/Normal region of 99Tcm-MIBI uptake in the delayed stage and in the early stage, age, cough and specula sign were five most important contributors. The sensitivity and specificity were 93.33% and 78.57%, respectively, a little higher than those of the expert. The sensitivity and specificity by residents of Grade one were 76.67% and 28.57%, respectively, and AUC of CART and expert was 0.886±0.055 and 0.829±0.062, respectively, and the corresponding AUC of residents was 0.566±0.092. Comparisons of AUCs suggest that performance of CART was similar to that of expert (P = 0.204), but greater than that of residents (P<0.001). Conclusion: Our data mining technique using classification decision tree has a much higher accuracy than residents. It suggests that the application of this algorithm will significantly improve the diagnostic performance of residents.