Using artificial neural network analysis of global ventilation-perfusion scan morphometry as a diagnostic tool.

OBJECTIVE The purpose of this study was to determine whether global statistical data from radionuclide ventilation-perfusion scans could predict the likelihood of pulmonary embolism. MATERIALS AND METHODS Digital data were obtained from 161 patients undergoing both radionuclide ventilation-perfusion scanning and subsequent pulmonary angiography. Morphometric data characterizing whole-lung perfusion and ventilation parameters were input into artificial neural networks in an attempt to predict the likelihood of pulmonary embolism. RESULTS The performance of artificial neural networks using only automated global region of interest-based data was superior to that of clinicians in predicting the likelihood of acute pulmonary embolism in patients with normal findings on chest radiographs with segmental or larger emboli (p < .005) and in patients with normal findings on chest radiographs and emboli of any size (p < .01). Network performance did not significantly differ from clinician performance in patients with abnormal findings on chest radiographs. CONCLUSION The adjunctive use of artificial neural networks using only user-independent, standard image statistics can significantly improve accuracy in the diagnosis of pulmonary embolism in patients with normal findings on chest radiographs.