Liling Hao

Segmentation of the left ventricle in cardiac MRI using a hierarchical extreme learning machine model

Segmentation of the left ventricle (LV) from cardiac magnetic resonance imaging (MRI) images is an essential step for calculation of clinical indices such as stroke volume, ejection fraction. In this paper, a new automatic LV segmentation method combines a Hierarchical Extreme Learning Machine (H-ELM) and a new location method is developed. An H-ELM can achieve more compact and meaningful feature representations and learn the segmentation task from the ground truth. A new automatic LV location method is integrated to improve the accuracy of classification and reduce the cost of segmentation. Experimental results (including 30 cases, 10 cases for training, 20 cases for testing) show that the mean absolute deviation of images segmented by our proposed method is about 67.9, 81.3 and 98.7% of those images segmented by the level set, the SVM and Hu’s method, respectively. The mean maximum absolute deviation of images segmented by our proposed method is about 63.5, 77.3 and 98.0% of those images segmented by the level set, the SVM and Hu’s method, respectively. The mean dice similarity coefficient of images segmented by our proposed method is about 13.7, 9.3 and 0.5% higher than that of those images segmented by the level set, the SVM and Hu’s method, respectively. The mean speed of our proposed method is about 38.3, 6.7 and 23.8 times faster than that of the level set, the SVM and Hu’s method, respectively. The standard deviation of our proposed method is the lowest among four methods. The results validate that our proposed method is efficient and satisfactory for the LV segmentation.

Diastolic Augmentation Index Improves Radial Augmentation Index in Assessing Arterial Stiffness

Arterial stiffness is an important risk factor for cardiovascular events. Radial augmentation index (AIr) can be more conveniently measured compared with carotid-femoral pulse wave velocity (cfPWV). However, the performance of AIr in assessing arterial stiffness is limited. This study proposes a novel index AIrd, a combination of AIr and diastolic augmentation index (AId) with a weight α, to achieve better performance over AIr in assessing arterial stiffness. 120 subjects (43 ± 21 years old) were enrolled. The best-fit α is determined by the best correlation coefficient between AIrd and cfPWV. The performance of the method was tested using the 12-fold cross validation method. AIrd (r = 0.68, P < 0.001) shows a stronger correlation with cfPWV and a narrower prediction interval than AIr (r = 0.61, P < 0.001), AId (r = −0.17, P = 0.06), the central augmentation index (AIc) (r = 0.61, P < 0.001) or AIc normalized for heart rate of 75 bpm (r = 0.65, P < 0.001). Compared with AIr (age, P < 0.001; gender, P < 0.001; heart rate, P < 0.001; diastolic blood pressure, P < 0.001; weight, P = 0.001), AIrd has fewer confounding factors (age, P < 0.001; gender, P < 0.001). In conclusion, AIrd derives performance improvement in assessing arterial stiffness, with a stronger correlation with cfPWV and fewer confounding factors.

Synchronous acquisition of multi-channel signals by single-channel ADC based on square wave modulation

Synchronous acquisition of multi-channel biopotential signals, such as electrocardiograph (ECG) and electroencephalograph, has vital significance in health care and clinical diagnosis. In this paper, we proposed a new method which is using single channel ADC to acquire multi-channel biopotential signals modulated by square waves synchronously. In this method, a specific modulate and demodulate method has been investigated without complex signal processing schemes. For each channel, the sampling rate would not decline with the increase of the number of signal channels. More specifically, the signal-to-noise ratio of each channel is n times of the time-division method or an improvement of 3.01×log2n dB, where n represents the number of the signal channels. A numerical simulation shows the feasibility and validity of this method. Besides, a newly developed 8-lead ECG based on the new method has been introduced. These experiments illustrate that the method is practicable and thus is potential for low-cost medical monitors.

Design and implementation of a pulse wave generator based on Windkessel model using field programmable gate array technology

Abstract Purpose Pulse wave contains plenty of physiological and pathological information of cardiovascular system. There have been many commercial products that can analyze pulse wave signals for the quantification of cardiovascular functions. However, their performance often varies from case to case. It is thus necessary to generate typical pulse waveforms in order to quantitatively evaluate these commercial products. Methods A pulse wave generator based on the Windkessel model is designed and implemented in this study because Windkessel model can describe the general features of a pulse wave with physiologically interpretable parameters that can be easily set by users. The numerical solutions are obtained by using the Runge-Kutta method. Results The features of this work include: (1) the critical parameters are flexible to be modulated so that different pulse waveforms representing different states of the cardiovascular system could be simulated; (2) it is possible to add different types of noises with varying signal-to-noise ratios; (3) the software is designed under the System-on-a-Programmable-Chip (SoPC) developing flow; (4) the platform is implemented on Field Programmable Gate Array (FPGA) for efficiency, portability and scalability. Conclusion and significance The pulse waves generated by the designed generator are quite similar to these measured in clinic. The new pulse wave generator is useful to test and evaluate various pulse wave analysis devices. It is also useful for the training of hospital personnel and young students.

Segmentation of the left ventricle in cardiac MRI using an ELM model

In this paper, an automatic left ventricle (LV) segmentation method based on an Extreme Learning Machine (ELM) is presented. Firstly, according to background and foreground, all sample pixels of Magnetic Resonance Imaging (MRI) images are divided into two types, and then 23-dimensional features of each pixel are extracted to generate a feature matrix. Secondly, the feature matrix is input into the ELM to train the ELM. Finally, the LV is segmented by the trained ELM. Experimental results show that the mean speed of LV segmentation based on the ELM is about 25 times faster than that of the level set, about 7 times faster than that of the SVM. The mean values of mad and maxd of image segmentation based on the ELM is about 80 and 83.1 % of that of the level set and the SVM, respectively. The mean value of dice of image segmentation based on the ELM is about 8 and 2 % higher than that of the level set and the SVM, respectively. The standard deviation of the proposed method is the lowest among all three methods. The results prove that the proposed method is efficient and satisfactory for the LV segmentation.

Estimation of carotid artery pressure waveform by transfer function and radial pressure waveform

Carotid pressure waveform is often used to substitute the central aortic pressure waveform, which conveys lots of information regarding cardiovascular system. Thus, a method was proposed to reconstruct the carotid pressure waveform from the radial pressure waveform which can be more conveniently monitored in comparison with the carotid pressure waveform. The reconstruction method was using Finite Impulse Response (FIR) model to calculate mathematical transfer function (TF) between the radial pressure waveform and the carotid pressure waveform. Pulse pressure waveforms of 5 subjects were recorded to test the performance of the TF. Except for some details, the reconstructed carotid pressure waveform is acceptable with the best percent root-mean-square difference (PRD) of 13.60%.

Magnetic detection electrical impedance tomography with total variation regularization

Magnetic detection electrical impedance tomography (MDEIT) is an imaging modality that aims to reconstruct the cross-sectional conductivity distribution of a volume from the magnetic flux density surrounding an object. The MDEIT inverse problem is inherently ill-posed, necessitating the use of regularization. The most commonly used L(2) norm regularizations generate the minimum energy solution, which blurs the sharp variations of the reconstructed image. Consequently, this paper presents the total variation (TV) regularization to preserve discontinuities and piecewise constancy of the MDEIT reconstructed image. The primal dual-interior point method (PD-IPM) is employed for minimizing the TV penalty in this paper. The proposed method is validated by MDEIT simulated data. In comparison with the performance of L(2) norm regularization, the results show that TV regularized algorithm produces sharper images and has better robustness to noise. The TV regularized algorithm preserves local smoothness and piecewise constancy, leading to improvements in the localization of the reconstructed conductivity images in MDEIT.