NEURAL-NETWORK ANALYSIS OF DOPPLER ULTRASOUND BLOOD-FLOW SIGNALS - A PILOT-STUDY

It has been hypothesised that each artery in the human body has its ow n characteristic ''signature''-a unique Doppler flow profile which can identify the artery and which may also be modified by the presence of disease, To test this hypothesis an artificial neural network (ANN) w as trained to recognise three groups of maximum frequency envelopes de rived from Doppler ultrasound spectrograms; these were the common caro tid, common femoral and popliteal arteries, Data mere collected from 2 4 subjects known to have no significant atheromatous disease, The maxi mum frequency envelopes were used to create sets of training and testi ng vectors for a backpropagation ANN, The ANN demonstrated a high succ ess rate for appropriate classification of the test vectors: 100% for the carotid; 92% for the femoral; and 96% for the popliteal artery, Th is work has demonstrated the ability of the ANN to differentiate accur ately between different and similar flow profiles, outlining the poten tial of this technology to identify subtle changes induced hy the onse t of arterial disease within a specific vessel, It should be noted tha t the ANN not only models the maximum frequency envelope but also, unl ike standard indices, makes a decision as to which artery the maximum frequency envelope belongs to, thus providing the potential to obviate human subjective classification, (C) 1997 World Federation for Ultras ound in Medicine and Biology.