The maximum frequency waveforms of Doppler ultrasound blood flow signals were analyzed using a multi-scale wavelet transform. The variation of maxima of wavelet transform modulus under various scales was extracted from the time-scale representation. This novel approach was applied to the analysis of Doppler signals from carotid blood flow. It was found that the shape of this variation from cases with normal cerebral vessels differed from those associated with abnormal cases. The curve was fitted by a polynomial, and its coefficients were put into a back-propagation (BP) neural network to make a classification. The clinical experiments showed that this approach got good performance and could be a new means in the clinical diagnosis of cerebral vascular disease.