Context Modeling And Vector - Scalar Quantizer Of The Ecg Signal Compression

Electrocardiography (ECG) is the synthetic reflection of the heart electricity on the body surface over a unite time, which has the great clinical significance in diagnosing heart diseases. The multi-lead ECG recording system has been extensively used in many situations such as the 24-hour portable Holter monitor, the clinical ECG workstation and telemedicine applications, due to its merites of noninvasiveness, low cost and real-time representation. However, the amount of digital ECG data grows with the increase of the sampling rate, the sample resolution, the recording time, the number of channels and the number of subjects, etc. It is necessary to compress ECG data while to maintain clinically acceptable signal quality when the storage space and bandwidth are very limited. Up to now, there are three main compression categories: direct methods, transform methods and parameter methods. Among them, the one based on the discrete wavelet transform achieves the outstanding performance.Actually, as a quasi-periodic physiological signal, the multi-lead ECG has sample-to-sample, beat-to-beat and channel-to-channel correlations. The target of the dissertation is to reduce those redundancies as much as possible. Our studies focuse on the ECG compression method based on context model and vector-scale quantization, which have been carried out in following aspects.The first part mainly discusses three QRS detection methods for multi-lead ECG signals. It solves the low detection accuracy problem of the single lead ECG signal, which induced by the interference and the poor contact between the electrode and the body surface. It also could provide the accurate heartbeat information for the following data compression methods.A QRS detection method based on the combined wavelet entropy (CWS) is proposed for two-channel ECG signals. The raw ECG signals are firstly transformed by the continuous wavelet transform (CWT) within a selected frequency interval. It can effectively suppress the interference and only focus on the QRS component. The wavelet entropy based method is employed to analyze one channel of the ECG signal based on an automated threshold determination. When the difference between the current RR interval and the average RR interval is significant, the CWS based detection method is employed to enhance detection results by fusing the QRS information coming from both channels. Two QRS detection methods based on the blind source separation are proposed for 12-lead ECG signals. In the method one, the principal component analysis is employed to separate the ventricular signal and atrial signal from the interference noise. Then, the quasi-period sorting method is proposed to reorder principal components (PCs), which may help the following combined wavelet entropy based method detecting the QRS complex in the lower sorted PCs easily. In the method two, the independent component analysis is employed to extract independent components (ICs) corresponding to the ventricular activity from filtered multi-lead ECG signals. Each IC is transformed by the CWT and the phase space of the CWT coefficient is reconstructed. The IC has been reordered based on the QRS information in the phase space. Finally, the CWS based method is employed to detect the QRS of multi-lead ECG signals.The second part mainly discusses the ECG compression method. The context modeling entropy coding is effectively studied from several aspects, which are the lossless compression and lossy compression, the scalar quantization and the vector quantization, one dimensional (1-D) compression and two dimensional (2-D) compression, and the single-lead compression and two-lead compression.In the lossless compression method, the raw ECG signal is transformed by the lifting wavelet transform with well-known 9/7 filters which could map integers to integers. Then, wavelet coefficients are decomposed into the significance map, the sign stream, the position of the most significant bit stream and the residual bit stream. An adaptive arithmetic coder with different context models is employed for the conditional entropy coding of these streams.Two lossy compression methods are proposed for single-lead ECG signal. The method one is a 2-D compression based on scalar quantiztion. ECG signals are firstly cut and aligned to form an image by the QRS information. After the period sorting and mean removal, the image is transformed by 1-D discrete wavelet transform (DWT) and quantized by uniform scalar dead-zone quantizer (USDZQ). After decomposition of quantized wavelet coefficients, the coefficient streams are coded by context modeling entorpy coding. This method needs to collect periods of ECG data to form an image before each compression, which will lead to time latency. The method two is a 1-D compression based on vector-scalar quantization (VSQ). DWT coefficients in a hierarchical tree order are taken as the component of a vector named tree vector (TV). Then, the TV is quantized with a VSQ, which is composed of a dynamic learning vector quantizer (VQ) and a USDZQ. All quantization index and decomposed wavelet coefficients are also compressed by context modeling entropy coding. This method could be employed as an on-line compression schme.Taken as a specific example of the multi-lead ECG signals compression, a modified 1-D compression based on VSQ is applied to 2-lead ECG signals. Here, the VQ index from each channel is collected to form a new vector. The vector is then vector quanztied losslessly by using one additional codebook. Meanwhile, the original dynamic learning mechanism of the code book replenishment is modified by a static learning mechanism. It could squeeze more redundancy of ECG signals between each channel, which will improve the compression performance.Since the proposed methods greatly utilize the correlations between adjacent samples, heartbeats and channels, they successfully increase the compression ratio while maintain the quality of ECG signals, which outperform other methods based on JPEG2000 or set partitioning in hierarchical trees (SPIHT).