Research On Feature Extraction And Classification Of Doppler Ultrasonic Blood Flow Signals Of Wrist Radial Artery

Theories of traditional Chinese medicine insist that pulses of wrist radial arterial include the most pathological information of human body, doctors can get state of sick persons through putting their fingers on human wrist to feel vascular pulsation. Ultrasonic medical diagnosis which has become an important part of clinical diagnosis can use ultrasound instruments to detect human blood flow and movement of organizations. However, most research on Doppler signals often focus on its application in diagnosing and preventing cardiovascular diseases. In this paper, ultrasound instruments can supplant putting their fingers on human wrist; some feature extraction methods are studyed on Doppler ultrasound blood flow signals of wrist radial arterial and make analysis of its reactions to certain pathological changes of human body, for the purpose of computer-aided diagnosis.During the data collection process, ultrasonic blood flow signals can be saved as Doppler sonograms. So some extraction algorithms should be designed to extract maximum velocity waveform and mean velocity waveform from Doppler sonograms. These two kind waveforms reflect the information of blood flow velocity, certain pathological changes of human body influence the movement of flood and this infection can be reflected in the shape of these two kind waveforms. During preprocessing process, because Doppler signal is relatively weak and has low SNR, a wavelet packet threshold denoising method is proposed to remove noise on the maximum velocity waveform, which not only remove the noise farthest, but also remain details features in the high-frequency of signal.During feature extraction process, two feature extraction methods are proposed in this paper: feature extraction method based on Hilbert-Huang transform and feature extraction method based on Local wave Approximate Entropy (LAE).Biomedical signals is usually non-stationary. Hilbert-Huang transform is a new kind of time-frequency analysis method for non-stationary signals. One precise time-frequency distribution can be showed through Hilbert-Huang transform which is applied on the maximum velocity waveform after removing noise, and then four dimensions features which have promising discriminating ability are created. LAE analysis the distribution of Approximate Entropy of original signal and its every component, components can get through Local wave decomposition. So the features are composed of Approximate Entropy of original signal and every component.In the classification phase, this paper used some common pattern recognition algorithms to perform the classification experiments on data set with the two feature extraction methods. In the experiments base on Hilbert-Huang transform, a support vector classifier could efficiently discriminate the healthy persons and patients, and the result reaches as high as 90%. In the experiments base on LAE, experimental results showed that this feature extraction method also provide a wonderful discriminating capability between patients and healthy persons. At last, a new algorithm that combined SVM with KNN is presented in this paper, and this classifier is used in the classification experiments of some types of patients. Experimental results showed that not only result of this classifier is higher than SVM, but also feature extraction method based on LAE extracts pathological information of different diseases efficiently.