ACUTE PULMONARY-EMBOLISM - ARTIFICIAL NEURAL-NETWORK APPROACH FOR DIAGNOSIS

PURPOSE: To investigate use of an artificial neural network (ANN) as a computer-aided diagnostic (CAD) tool for predicting pulmonary embolism (PE) from ventilation-perfusion lung scans and chest radiographs. MATERIALS AND METHODS: The data base consisted of cases extracted from the collaborative study of the Prospective Investigation of Pulmonary Embolism Diagnosis (PIOPED). Initially, scan findings from 1,064 patients (383 with PE, 681 without PE) were used to train and test the network by using the ''jackknife'' method. Then, a receiver-operating-characteristic analysis was applied to compare the performance of the network with that of the physicians involved in the PIOPED study. RESULTS: The ANN significantly outperformed the physicians involved in the PIOPED study (two-tailed P value = .01). CONCLUSION: The findings suggest that an ANN can form the basis of a CAD system to assist physicians with the diagnosis of PE.