Aiming at the prediction of the size of human cerebral hemorrhage point, a signal processing method based on Resonance Sparse Decomposition (RSSD) algorithm is proposed to decompose and analyze the microwave echo signal. According to the organizational structure of the human brain, a complete human brain model was established, and bleeding points of different sizes were placed at the same position, and 5 antennas were placed around the model (front, back, left, right, and top). RSSD is performed on the obtained echo signal, and Hilbert envelope analysis is performed on the low resonance component obtained by the decomposition, and then the size of the bleeding point is judged. Using CST and MATLAB to conduct simulation analysis and experiments, it is verified that the proposed method can successfully determine the size of the bleeding point, and the effectiveness and feasibility of the method are proved.
Cerebral hemorrhage, also known as stroke, is the third leading cause of morbidity and mortality in many countries. Stroke is mainly divided into two categories, one is ischemic and the other is hemorrhagic. An ischemic stroke is due to loss of blood supply to an area of the brain. This is a common type of stroke. A hemorrhagic stroke is bleeding in the brain due to rupture of a blood vessel [
The size of the bleeding spot in cerebral hemorrhage can vary with individual differences and specific circumstances. In general, the size of the bleeding point of cerebral hemorrhage can vary from tiny to large. Tiny bleeding spots may only affect a small part of the brain and may not cause noticeable symptoms. Larger hemorrhages may involve more extensive brain regions, potentially leading to severe neurological deficits.
The size of the bleeding point is usually determined by neuroimaging studies, such as head CT scan or MRI [
Intracerebral hemorrhage point target detection, in essence, is a kind of fault diagnosis. Resonance sparse decomposition algorithm is a relatively new signal processing method, which is different from the traditional signal processing method based on frequency or size. In the process of realizing the feature extraction of vibration signal, the method considers both the frequency and bandwidth of the signal. The components of vibration signal are understood as periodic harmonics, fault shock and noise, and separated into high and low resonance components and residual components respectively. At present, the resonance sparse decomposition algorithm is mainly used in the fault diagnosis of bearings, and carries out resonance sparse decomposition of bearing signals to obtain high resonance component and low resonance component. The Hilbert envelope spectrum demodulation analysis is carried out on the low resonance component to extract the fault characteristic frequency, and then the fault diagnosis of the rolling bearing is carried out [
Compared with mainstream detection methods, microwave technology can be used for non-destructive detection of stroke, and the price is low. Hemorrhagic stroke will change the dielectric properties of the tissue [
Resonance Sparse Decomposition (RSSD) is a newly proposed signal decomposition algorithm, which is different from the traditional signal decomposition algorithm (wavelet decomposition). RSSD takes both the frequency and bandwidth of the signal into account when decomposing the signal. Tunable Q factor wavelet transform (TQWT) is used to represent the high and low quality factors of the signal according to the different quality factors of the harmonic signal and the fault impact signal. The signals are then separated nonlinearly by morphological component analysis (MCA) [
Q = f c B (1)
In (1), f c is the center frequency of an oscillation of the signal, and B is its bandwidth.
According to the signal quality factor Q, it is decomposed into high resonance component, low resonance component and residual component. The high resonance component has a large Q value and good frequency aggregation, which represents the harmonic component; the low resonance component has a small Q value and poor frequency aggregation. Represents the transient shock classification (fault component).
Assuming that the signal x is formed by a linear combination of K different morphological feature components, each component corresponds to a complete dictionary, so the initial signal x can be expressed as
x = ∑ K = 1 K x K = ∑ K = 1 K S K W K (2)
In (2), S K is an over-complete dictionary, and W K is the coefficient of an over-complete dictionary.
First, TQWT is performed on the signal to obtain two sets of basis function libraries, and the corresponding transformation coefficients are obtained through iterative calculation. Then use MCA to establish the objective function of sparse decomposition, which can be expressed as
J ( w 1 , w 2 ) = ‖ x − S 1 w 1 − S 2 w 2 ‖ 2 2 + λ 1 ‖ w 1 ‖ 1 + λ 2 ‖ w 2 ‖ 1 (3)
In (3), and are the transformation coefficients of high-resonance classification and low-resonance components in S 1 and S 2 ; λ 1 and λ 2 are regularization parameters, and their magnitudes are related to the energy distribution of high- and low-resonance components during decomposition. When the size of λ 1 is fixed, if the value of λ 2 is reduced, the energy distribution of the resonance component corresponding to λ 2 will increase. When the values of λ 1 and λ 2 are reduced at the same time, the energy of the residual signal component will decrease.
Then, use the SALSA algorithm to perform iterative operations on the above equations, and update the transformation coefficients to obtain the minimum objective function. When the dissipation function is minimized, the optimal coefficient matrices w 1 * and w 2 * are obtained. Finally, the sparse expressions S 1 w 1 * and S 2 w 2 * of the high and low resonance components in the original signal are reconstructed. The flowchart of RSSD is shown in
The human brain is the center of the human nervous system. It has complex structure and functions and is one of the most important organs of the human body. It is responsible for controlling and regulating almost all physical and mental activities of the human body, including perception, thinking, memory, movement, emotion, etc. In order to study the physical phenomena inside the human brain more accurately, such as the reflection and absorption of microwaves by bleeding spots, a complete electromagnetic model of the human brain is established. The established model is shown in
The anatomical structure of the human brain is very complex. According to the anatomy diagram of the human brain, it can be concluded that the human brain is composed of multiple parts. During simulations and experiments, the dielectric constant is one of the indicators that measures the ability of a medium to respond to an electric field. Therefore, selecting brain tissues with similar or equal dielectric constants for similar sensing can ensure that simulations and experiments are easy to conduct and can provide accurate and reliable results.
In the human brain model, the brain is divided into different tissues, including cerebrospinal fluid, gray matter, white matter, skull, and skin. These tissues differ in structure and properties and respond differently to electromagnetic fields. The corresponding models of different tissues in the human brain are shown in
It can be seen from
| Brain tissue | Dielectric constant ε r | Conductivity σ (S/m) |
|---|---|---|
| Cerebrospinal fluid | 68 | 2.5 |
| Eyeball | 5.447 | 0.053 |
| Gray matter | 52.282 | 0.985 |
| Brain parenchyma | 42 | 0.7 |
| White matter | 46 | 0.82 |
| Skull | 12 | 0.2 |
| Skin | 40.936 | 0.089 |
| Blood | 61 | 1.6 |
The antenna used in this experiment is “dodecagon Archimedean spiral antenna”. It is a small size ultra-wideband antenna. The antenna operates at a frequency of 0.6 - 6.5 GHz. The gain in the working range is 2 - 5 dBi, which basically meets the experimental requirements. The main radiant area of the antenna consists of a 60 mm × 60 mm dielectric substrate and a dodecagon Archimedean spiral radiator. The feed mode of the antenna adopts Barron feed structure, which realizes the impedance matching between coaxial line and radiator and the conversion of unbalanced signal to balanced signal.
A bleeding spot with a radius of 5 mm was placed on the upper left of the complete electromagnetic model of human brain. The location of the bleeding point is shown in
The model was simulated after the hemorrhagic sites were placed. The input signal is a single Gaussian pulse signal, and five sets of output signals (o1, 1) are obtained after simulation, as shown in
From the five groups of o1,1 waveform diagrams obtained above, it is not difficult to see that there is little difference between the signals obtained from bleeding points of different sizes. This is because the bleeding point model is not an order of magnitude in size or complexity compared with the whole human brain model, so the difference reflected in the signal is very small. This also reflects the importance of extracting features by processing the signal.
The obtained simulation result data is exported. Considering that the original excitation signal can only be a monopulse signal, the resulting o1,1 waveform is also the result of a monopulse signal. In order to simulate the actual signal waveform, input a multi-pulse signal. The o1,1 waveform graph is now periodically extended by 602 units points. Then the first 4096 points are taken for RSSD processing. The final waveform is shown in
The resonance sparse decomposition algorithm mainly includes three parameters: quality factor Q, wavelet redundancy r, and decomposition layer number J. According to the decomposed signal waveform diagram and experience, quality factors Q1 = 6, Q2 = 1 and redundancy r1 = 12, r2 = 25 can be preliminarily selected. The maximum value of the number of decomposition layers J is determined as
J m a x = ⌊ l o g ( β N / 8 ) l o g ( 1 / α ) ⌋ (1)
In (4), ⌊ ∗ ⌋ represents the largest integer that does not exceed the number in parentheses, and N is the signal length.
After the parameters are determined, the signal to be decomposed is subjected to resonance sparse decomposition, and the decomposition results of the five signals to be decomposed are shown in
It can be seen from the above figure that there is almost no difference between the high resonance components in the decomposition results of the five groups, and there are obvious differences in the low resonance components. Next, the five groups of low resonance components are further processed and analyzed.
First of all, in order to limit the preprocessed data to a certain range, so as to eliminate the adverse effects caused by singular sample data. The low resonance component is normalized, and the result is shown in
The obtained normalized data is subjected to envelope extraction (Hilbert transform), and then fast Fourier transform (FFT) is performed on it. The obtained results are shown in
To see the trend more clearly. The second highest amplitude point of each picture in
It is not difficult to see from the result graph that with the increase of the bleeding point, the value of the second highest amplitude point of the real part of the envelope spectrum of the low resonance component of different bleeding point signals is getting larger and larger, and there is a tendency to gradually delay.
This experiment is based on the signal analysis of cerebral hemorrhage points of different sizes based on the resonance sparse decomposition algorithm. Firstly, the cycle extension is carried out on the modeling and simulation results obtained, and then the signal is decomposed on the basis of selecting important parameters. The decomposed signals of low resonance components with obvious differences are processed by normalization, envelope extraction and FFT. Finally, with the increase of the bleeding point, the value of the second highest amplitude point of the frequency domain signal is getting larger and gradually moving backward. An important conclusion can be drawn that the larger the bleeding point, the higher the value of the second highest amplitude point in the frequency domain.
The authors declare no conflicts of interest regarding the publication of this paper.