REAL-TIME IDENTIFICATION OF CEREBRAL MICROEMBOLI WITH US FEATURE DETECTION BY A NEURAL-NETWORK

PURPOSE: Abnormal transcranial Doppler ultrasonographic (US) signals indicating cerebral microembolism have characteristic but complex features. The authors wanted to assess the agreement among human observers and test the feasibility of an automated detection system. MATERIALS AND METHODS: Automated on-line detection of cerebral microemboli was accomplished by employing real-time overlapping Fourier transform and artificial neural network technology. By using long-term transcranial Doppler US recordings of the middle cerebral artery in consecutive cerebrovascular and cardiac patients, the method was evaluated in a clinical setting. RESULTS: The proportion of specific agreement (p(s)) among four experienced investigators identifying cerebral microemboli was high (mean p(s), 0.91). Agreement among the neural network and the human observers was only slightly less (mean p(s) 0.77). CONCLUSION: The technique allows highly reliable on-line evaluation of transcranial Doppler US recordings across multiple centers. It obviates time-consuming analyses by human observers.