Cerebral hemorrhage spot size analysis method based on multi-scale power spectrum entropy and singular spectrum entropy

Abstract

The invention relates to the technical field of cerebral apoplexy detection and diagnosis, in particular to a cerebral hemorrhage point size analysis method based on multi-scale power spectrum entropy and singular spectrum entropy, which comprises the following steps: acquiring medical image data containing a cerebral hemorrhage area, and preprocessing the medical image data; analyzing the preprocessed image data by using a multi-scale power spectrum entropy algorithm, and extracting power spectrum entropy characteristics of the bleeding point area; and analyzing the preprocessed image data by using a multi-scale singular spectrum entropy algorithm, and extracting power spectrum entropy characteristics of the bleeding point area. According to the method, the power spectrum entropy and the singular spectrum entropy are extracted and calculated from the microwave signals of different hemorrhagic spot sizes in the cerebral hemorrhagic stroke, the microwave signals are analyzed, the cerebral hemorrhagic spots of different sizes can be effectively distinguished, the multi-scale power spectrum entropy and the singular spectrum entropy are taken as features to be substituted into a machine learning algorithm for training, and the cerebral hemorrhagic spot detection accuracy is improved. And the sizes are quickly classified, so that the aim of analyzing the sizes of different bleeding spots is fulfilled.

Description

Technical Field

[0001] The present invention relates to the technical field of stroke detection and diagnosis, and specifically to a method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectral entropy and singular spectral entropy. Background Art

[0002] A stroke refers to a disease in which the blood supply to the brain is suddenly interrupted due to cerebrovascular lesions, resulting in ischemia (hypoxia) or hemorrhage of brain tissue. Hemorrhagic stroke may be caused by reasons such as aneurysm rupture and artery rupture caused by hypertension. Stroke has become one of the diseases with extremely high mortality. The longer the time of suffering from stroke, the more dangerous it is. Therefore, seeking medical treatment quickly is crucial for the rehabilitation and survival of stroke patients.

[0003] However, in actual use of the existing technology, microwave tomography (MWT), as an emerging non-invasive imaging technology, generates images by measuring the propagation characteristics of microwaves in skin tissue. However, it is sensitive to noise and has a simple model, and cannot comprehensively analyze the sizes of different hemorrhage points. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectral entropy and singular spectral entropy to solve the problem of not being able to comprehensively analyze the sizes of different hemorrhage points.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectral entropy and singular spectral entropy, including the following steps:

[0006] S1: Data acquisition step: Obtain medical image data containing the cerebral hemorrhage area. The medical image data is obtained through a medical imaging device, and the obtained medical image data is preprocessed;

[0007] S2: Multi-scale power spectral entropy calculation: Use the multi-scale power spectral entropy algorithm to analyze the image data preprocessed in step S1, and extract the power spectral entropy features of the hemorrhage point area;

[0008] S3: Multi-scale singular spectral entropy calculation: Use the multi-scale singular spectral entropy algorithm to analyze the image data preprocessed in step S1, and extract the power spectral entropy features of the hemorrhage point area;

[0009] S4: Feature fusion:

[0010] Take the multi-scale power spectral entropy in step S2 and the multi-scale singular spectral entropy in step S3 as feature vectors, and perform feature fusion;

[0011] Use a machine learning algorithm to train and analyze the feature vectors, and establish a mapping relationship model between the size of the cerebral hemorrhage point and the feature vectors;

[0012] S5: Analysis of the size of the cerebral hemorrhage point:

[0013] Input the new ultra-wideband microwave signal data into the model trained in step S4, and calculate its multi-scale power spectral entropy and singular spectral entropy;

[0014] Calculate the size of the cerebral hemorrhage point according to the mapping relationship.

[0015] Preferably, the preprocessing in step S1 includes:

[0016] Edge detection: Use the Canny edge detection algorithm to extract the boundaries of brain tissue structures, including five steps: Gaussian filtering, calculating gradients, non-maximum suppression, double-threshold processing, and boundary tracking and connection;

[0017] Region division: Divide the interior into five regions: white matter, gray matter, dura mater, cerebrospinal fluid, and skull, and finally add the skin layer;

[0018] Numerical model establishment: Use the processed numerical model as the calculation region of FDTD, set the perfectly matched layer, and assign electromagnetic parameters of different brain tissues according to the first-order approximation of the Debye equation.

[0019] Preferably, the specific process of extracting the power spectral entropy feature of the hemorrhage point region in step S2 is as follows:

[0020] (a) Perform multi-scale decomposition on the image data;

[0021] (b) Calculate the power spectral entropy at each scale;

[0022] (c) Extract the power spectral entropy feature.

[0023] Preferably, the specific process of extracting the power spectral entropy feature of the hemorrhage point region in step S3 is as follows:

[0024] (d) Perform singular value decomposition on the image data;

[0025] (e) Calculate the singular spectral entropy;

[0026] (f) Extract the singular spectral entropy feature.

[0027] Preferably, the machine learning algorithms in step S4 include support vector machine, random forest, and neural network algorithms.

[0028] Compared with the prior art, the beneficial effects of the present invention are:

[0029] 1. The present invention extracts and calculates the power spectral entropy and singular spectral entropy of microwave signals at different hemorrhage points of hemorrhagic stroke, analyzes the microwave signals, can effectively distinguish hemorrhage points of different sizes, and trains by taking the multi-scale power spectral entropy and singular spectral entropy as features into a machine learning algorithm for rapid classification of sizes, thereby achieving the purpose of analyzing the sizes of different hemorrhage points. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is the original nuclear magnetic resonance T1 map;

[0031] Figure 2 is the brain tissue boundary map processed by the present invention;

[0032] Figure 3 is the numerical model map of the brain processed by the invention;

[0033] Figure 4 is the schematic diagram of the distribution of the electromagnetic model of the brain without hemorrhage and 16 antennas;

[0034] Figure 5 is the schematic diagram of the two-dimensional electromagnetic model of 8 hemorrhage points of different sizes;

[0035] Figure 6 is the schematic diagram of the single 16*15 signal matrix of the present invention;

[0036] Figure 7 is the schematic diagram of the coarse-grained signal of scale 1-10 of the 10th column and the 3rd signal selected from the signal matrix of the 0mm healthy sample in step S2;

[0037] Figure 8 is the schematic diagram of the line comparison of the power spectral entropy of scale 1-10 of the 10th column and the 3rd signal of 8 hemorrhage points of different sizes;

[0038] Figure 9 is the schematic diagram of the line comparison of the singular spectral entropy of scale 1-10 of the 10th column and the 3rd signal of 8 hemorrhage points of different sizes. DETAILED DESCRIPTION OF THE INVENTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] Please refer to Figures 1-9 , the present invention provides a technical solution: a method for analyzing the size of a cerebral hemorrhage point based on multi-scale power spectral entropy and singular spectral entropy, including the following steps:

[0041] S1: Data acquisition step:

[0042] Obtain medical image data containing the cerebral hemorrhage area. The medical image data is obtained through a medical imaging device, such as Figure 1 shown, where Figure 1 is an original nuclear magnetic resonance T1 map with a resolution of 512×512, that is, there are 512×512 pixel points, which represents a real size of 30cm×30cm, and preprocess the medical image data. The preprocessing is as follows:

[0043] Edge detection: Use the Canny edge detection algorithm to extract the brain tissue structure. There are mainly five steps for the Canny edge detection algorithm to extract the boundary: Gaussian filtering, calculating the gradient, non-maximum suppression, double-threshold processing, boundary tracking and connection, Figure 2 is the processed brain tissue boundary map.

[0044] Region division: Analyze the gray range corresponding to different tissues. It is divided into five regions in total: white matter, gray matter, dura mater, cerebrospinal fluid, and skull. Finally, add a skin layer to the brain according to the outermost contour of the canny operator boundary. Figure 3 is the numerical model diagram of the brain.

[0045] Numerical model establishment: The core principle of the finite-difference time-domain (FDTD) method is to use the discretized Maxwell partial differential equations, that is, the difference equations, to simulate the evolution process of the discrete grid of time and space in the electromagnetic field. Use the processed numerical model as the calculation area of FDTD, and set the perfectly matched layer. The electromagnetic parameters of each different tissue are obtained according to the first-order approximation of the Debye equation. First, establish an electromagnetic model of the brain without bleeding points, and at the same time design a signal transmitter. In order to simulate the antenna position of the stroke detection device, 16 antennas are evenly arranged around the scalp. During the detection process, one antenna emits a signal, and the other antennas receive the signal respectively. In this simulation, the transmitted signal is set as a sixth-order Gaussian pulse signal, and the transmitter adopts the form of single-shot multi-receive. Each antenna takes turns to emit a signal. When one antenna emits, all the other antennas are in the receiving state. As Figure 4 shown is the position of the two-dimensional head model and the antenna array. Each color represents a tissue. The part outside the head in the figure is set as the air layer.

[0046] Among them, common types of cerebral hemorrhage include: intraventricular hemorrhage, parenchymal hemorrhage, subdural hemorrhage, and epidural hemorrhage. In this embodiment, taking intraventricular hemorrhage as an example, a position is fixed in the upper left part of the above model, and bleeding points of different sizes are manually set at the same position to obtain microwave signals. The radii of the bleeding points of different sizes are 8 different sizes of bleeding points with radii of 0 mm (healthy), 1.5 mm, 3 mm, 4.5 mm, 6 mm, 7.5 mm, 9 mm, and 10 mm respectively. The blue points in the red circles are the bleeding points. As Figure 5 shown, the bleeding points of different sizes of 0 mm (healthy), 1.5 mm, 3 mm, 4.5 mm, 6 mm, 7.5 mm, 9 mm, and 10 mm are in sequence from left to right and top to bottom.

[0047] From the above, a total of 8 brain models of different sizes are set. The set emission source is a single - transmit multi - receive mode. The 8 different types of electromagnetic models are calculated by FDTD to obtain 8 signal matrices of 16×15. Since the number of signals in the signal matrix is too large, the 3rd signal in the 10th column of each signal matrix is randomly selected for processing. Among them Figure 6 is a schematic diagram of a single 16*15 signal matrix. As Figure 6 shown in the red part, the power spectral entropy and singular spectral entropy of scale 1 - 10 are extracted and calculated for the 3rd signal in the 10th column of each signal matrix, so as to perform step S2 and step S3.

[0048] Power spectral entropy and singular spectral entropy are often used in the fault diagnosis of rolling bearings. Usually, some operating information of the bearings can be obtained by performing special processing on the obtained bearing signals. Time - domain and frequency - domain analysis of signals are currently common analysis methods. The fault characteristics obtained by processing the obtained rolling bearing operating information using these analysis methods are an important basis for bearing operating state diagnosis. However, for some signals with relatively low signal - to - noise ratio, such as acoustic signals, it is very difficult to obtain their fault characteristics only through simple time - domain and frequency - domain analysis. For this type of signal, it is necessary to combine with other mathematical analysis methods to achieve a better extraction of fault characteristics. The same is true for the above - obtained microwave signals. It is very difficult to directly see the difference between each signal of different bleeding - point sizes. Therefore, inspired by the diagnosis of rolling bearing faults, the signals are processed by extracting multi - scale power spectral entropy and singular spectral entropy.

[0049] S2: Multi - scale power spectral entropy calculation:

[0050] The characteristic signal can be reflected not only in the time domain but also in its frequency domain space. Therefore, the collected signal can be analyzed in the frequency domain. Common frequency domain analysis methods include spectrum, energy spectrum, and power spectrum, etc. Through these methods, the energy distribution of the signal in the frequency domain space can be obtained. Combining this energy distribution with information entropy can achieve the quantitative description of the signal in the frequency domain space. The present invention selects power spectrum entropy to describe the characteristics of the signal in the frequency domain space, and the power spectrum entropy will be introduced in detail below.

[0051] Power spectrum analysis is based on Parseval's theorem, that is, the total energy contained in a signal sequence is always equal to the sum of the energy components of the signal on a complete orthogonal function set. The formula description of Parseval's theorem is:

[0052]

[0053] In Equation (1), X(f) is the continuous Fourier transform, where f represents the frequency component of the signal. Parseval's theorem shows that the total energy of the signal in the time domain is equal to the total energy of X(f) in the frequency domain f after its Fourier transform, that is, the signal can maintain the total energy unchanged after Fourier transform, that is, the time domain energy is always equal to the frequency domain energy.

[0054] For the discrete-time sequence signal {x(n)|n = 0, 1,... N - 1}, another form of Parseval's theorem can be obtained through discrete Fourier transform as shown in Equation (2):

[0055]

[0056] In the formula, S(k) represents the energy partition of the signal in the frequency domain, and DFT[x n is the discrete Fourier transform of the original signal, and the sample length remains unchanged before and after the transformation. After obtaining the power spectrum of the signal, combining it with information entropy can obtain the power spectrum H f , as shown in Equation (4):

[0057]

[0058] In Equation (4), represents the proportion of the energy of the k-th frequency band of the signal in the total energy.

[0059] Through the above analysis, it can be obtained that the power spectrum entropy is used to represent the uncertainty of the signal energy under the power spectrum partition. When the frequency composition in the signal is simple, the power spectrum is concentrated in some frequency components, and the corresponding frequency spectrum lines will be fewer, and the corresponding component probabilities will be less, resulting in the obtained power spectrum entropy H fThe value will become smaller. On the contrary, if the signal is more complex, the power spectrum corresponding to the signal is more dispersed, the corresponding power spectrum lines will increase, and the obtained power spectrum entropy H f will become larger. Therefore, the power spectrum entropy is a quantitative description of the complexity of the energy distribution of the signal in the frequency domain.

[0060] The basic steps for extracting the power spectrum entropy of a discrete-time sequence signal {x(n)|n = 0, 1,... N - 1} of length N are as follows:

[0061] (1) Perform a discrete Fourier transform on {x(n)|n = 0, 1,... N - 1}.

[0062]

[0063] In Equation (5), k = 1, 2,... N - 1 is the frequency order.

[0064] (2) According to the relationship between signal energy and power, the power spectrum of each order of the spectrum can be obtained, that is, calculated using the power density of Equation (2).

[0065] (3) The information entropy H of the power spectrum f The power spectrum entropy H can be calculated according to Equation (4) from the calculated power spectra S(k) of each frequency band f .

[0066] "Multi-scale" means analyzing the complexity of the signal at different time scales. Specifically, multi-scale analysis generates signal sequences of different scales by "coarse-graining" the original signal, and then calculates the entropy value at each scale. The purpose of multi-scale analysis is to solve the problem that single-scale analysis may ignore important information in the signal, and many signals exhibit different complexities at different time scales.

[0067] The implementation of multi-scale is completed by introducing a "scale factor". In the technical solution, the scale factor is selected as 1 - 10 to calculate the power spectrum entropy of 16 signals of 8 different sizes of cerebral hemorrhage points:

[0068] (1) Scale factor is 1: The processed sequence is equal to the original sequence, with a length of 6000.

[0069] (2) Scale factor is 2: Average the adjacent 2 numbers in the sequence to generate a new sequence with a length of 3000.

[0070] (3) Scale factor is 3: Average the adjacent 3 numbers in the sequence to generate a new sequence with a length of 2000.

[0071]

[0072] (10) Scale factor is 10: Average every 10 adjacent numbers in the sequence to generate a new sequence with a length of 600.

[0073] As the scale factor increases, the sequence length becomes shorter, the signal becomes smoother, the details decrease, and the long-term trend characteristics become more prominent. Taking the 0mm healthy sample as an example, Figure 7 It is the coarse-grained signal diagram of scale 1 - 10 for the 3rd signal in the 10th column selected from the 0mm healthy sample signal matrix in step S2. 1.5mm, 3mm, 4.5mm, 6mm, 7.5mm, 9mm, 10mm show the same trend as the healthy sample.

[0074] Extract the power spectral entropy of scale 1 - 10 for the 3rd signal selected from 8 different sizes of 0mm, 1.5mm, 3mm, 4.5mm, 6mm, 7.5mm, 9mm, 10mm, and plot it as a line chart for comparison, as Figure 8 shown.

[0075] In the 3rd signal, the line charts of the power spectral entropy of scale 1 - 10 for 8 different sizes of bleeding points show the same trend. However, due to the small difference in data size, the images almost overlap. When magnifying each scale, an obvious pattern can be found. The order of the power spectral entropy values of 8 sizes from high to low for all scales is: 0mm, 1.5mm, 3mm, 4.5mm, 6mm, 7.5mm, 9mm, 10mm. That is, the 8 - size line charts at 10 scales do not overlap, can be completely distinguished, and maintain the same size pattern.

[0076] It is verified that the line charts of 8 different sizes of bleeding points for 240 signals (16 * 15 = 240) in the signal matrix can be completely separated, without intersection points, and can be arranged in different high - low orders. Each signal maintains the same size pattern at 10 scales.

[0077] S3. Multi - scale singular spectrum entropy calculation:

[0078] In step S2, an analysis method in the signal frequency - domain space is used. In this step, we return to the analysis of the signal time - domain space. Commonly used analysis methods for signals in the time - domain space include time - domain analysis based on correlation, time - domain analysis based on statistics, and singular spectrum entropy. These different analysis methods have a commonality that they describe the time - domain characteristics of signals based on different calculation parameters or corresponding model estimation methods. There are also some differences among them, and each analysis method has its own application conditions and characteristics. Singular spectrum entropy is an analysis method that combines the singular spectrum of the dynamic analysis method and information entropy. It is applicable to application scenarios with fewer sampling points and signals containing noise.

[0079] The definition of singular value decomposition is as follows: If matrix A ∈ R m×n , then there must be an orthogonal matrix A ∈ R m×n , V ∈ R m×n , such that equation (6) holds.

[0080] U T AV = diag(δ1, δ2,... δ p ) (6)

[0081] Where p = min(m, n); the singular values of matrix A are δ1, δ2,... δ p , and δ1 ≥ δ2 ≥... ≥ δ p > 0; the eigenvectors of matrix A T A or A T A are the column vectors of the orthogonal matrices U and V.

[0082] When applying singular value decomposition to the field of signal processing, it is first necessary to construct a matrix A using the collected discrete time series. Assume that the collected discrete time series is {x i | i = 1, 2... N}, select the window length m to reconstruct x i in the phase space, and the obtained trajectory matrix A is as follows:

[0083]

[0084] According to the trajectory matrix A obtained in equation (7) m×n , where n = N - m + 1. According to the singular value decomposition principle, for the trajectory matrix A m×n there must exist a diagonal matrix Λ and two orthogonal matrices U m×l and V n×l , such that equation (8) holds:

[0085]

[0086] Where Λ = diag(λ1, λ2,... λ l ), and (λ1 ≥ λ2 ≥,..., ≥ λ l ≥ 0), for the values on the diagonal of the diagonal matrix Λ are the singular values of the trajectory matrix A m×n . The number of eigenvalues of the trajectory matrix is related to the frequency components in the discrete time series, that is, if its frequency composition is more complex, there are more non-zero elements on the diagonal of the diagonal matrix Λ. Conversely, if the frequency components of the time series are simple, the corresponding number of non-zero elements on the diagonal of Λ is less. For a time series with a relatively simple frequency composition, the diagonal matrix obtained by performing singular value decomposition on its trajectory matrix A m×n can be expressed by equation (9):

[0087] Λ = diag(λ1, λ2,... λ k ,..., 0), (k < l; λ i ≠ 0, i = 1, 2,..., k) (9)

[0088] From the above analysis, the following conclusions can be drawn. The number k (k < l) of non - zero eigenvalues in the diagonal matrix Λ is closely related to the frequency components in the time series. The richer the frequency composition in the time series, the larger the number k of eigenvalues obtained. The more single the frequency composition in the time series, the smaller the number k of eigenvalues obtained. Therefore, the value of k can be used to represent the number of different patterns contained in A m×n , and the proportion of each pattern in the total patterns can be reflected by the magnitude of its eigenvalue λ i .

[0089] After obtaining the singular value spectrum {λ i} of the signal, combining it with the information entropy can obtain the singular value spectrum entropy H t , and its calculation process is as follows:

[0090]

[0091] In Equation (11), p i represents the proportion of the i - th singular value λ i in the entire singular value spectrum, and can also be used to represent the component probability. From the above analysis, it can be seen that the singular spectrum entropy H t is a quantitative description of the uncertainty degree of the time series after singular spectrum decomposition, and can also reflect the distribution of signal energy. For a more complex time series signal, its corresponding time - domain energy distribution is more dispersed, and the obtained singular spectrum entropy value is often larger. Similarly, for a simpler time series, the obtained singular spectrum entropy value is smaller.

[0092] Summarizing the above process, the solution process of H t is as follows:

[0093] (1) Obtain the signal trajectory matrix A m×n : Select an appropriate window length m to perform phase - space reconstruction on the one - dimensional time - series signal collected to obtain the trajectory matrix A m×n .

[0094] (2) Perform singular value decomposition on the trajectory matrix. Perform singular value decomposition on the A m×n calculated in the first step to obtain the corresponding singular value spectrum λ i (i = 1, 2... k).

[0095] (3) Calculate the singular value spectrum entropy H t Combine the singular value spectrum λ obtained in the second stepi Substituting (i = 1, 2... k) into equations (10) and (11) can obtain the singular spectrum entropy H t .

[0096] In the present invention, the selected window length of singular value decomposition is 500. Similarly, as in step S2, when calculating the singular spectrum entropy of the signal, the "scale factor" selected is 1 - 10. The coarse-grained signal diagrams for scales 1 - 10 (taking 0 mm as an example) are still as Figure 7 shown. Extract the singular spectrum entropy of the 3rd signal in the 10th column with 8 different sizes of 0 mm, 1.5 mm, 3 mm, 4.5 mm, 6 mm, 7.5 mm, 9 mm, and 10 mm for scales 1 - 10. Also, as in step S2, plot a line graph for comparison, as Figure 9 shown.

[0097] In the 3rd signal, the line graphs of the singular spectrum entropy of 8 different-sized bleeding points at scales 1 - 10 show the same trend. However, due to the small difference in data sizes, the images almost overlap. When magnifying each scale, a more obvious pattern can be found. The order of the singular spectrum entropy values of 8 sizes at all scales from high to low is: 10 mm, 9 mm, 7.5 mm, 6 mm, 4.5 mm, 3 mm, 1.5 mm, 0 mm. That is, the 8-sized line graphs at 10 scales do not overlap and can be completely distinguished, and maintain the same size pattern.

[0098] It is verified that the line graphs of 8 different-sized bleeding points of 240 signals (16 * 15 = 240) in the signal matrix can be completely separated without intersection points, and can be arranged in different high-low orders. Each signal maintains the same size pattern at 10 scales.

[0099] S4: Feature fusion:

[0100] Take the multi-scale power spectrum entropy in step S2 and the multi-scale singular spectrum entropy in step S3 as feature vectors, and perform feature fusion through feature splicing;

[0101] Use the neural network algorithm to train and analyze the feature vectors. Use the dataset with known cerebral hemorrhage point sizes to train the model. These datasets should include the composite features of the cerebral hemorrhage area and the corresponding cerebral hemorrhage point sizes, and establish a mapping relationship model between the cerebral hemorrhage point sizes and the feature vectors. During the training process, adjust and optimize the parameters of the model to improve the prediction accuracy and generalization ability of the model;

[0102] S5: Analysis of cerebral hemorrhage point size:

[0103] According to the model trained in step S4, input new ultra-wideband microwave signal data, and calculate its multi-scale power spectrum entropy and singular spectrum entropy;

[0104] According to the mapping relationship, calculate the size of the cerebral hemorrhage point, and the prediction result can provide valuable reference information for doctors, which helps to formulate more accurate and effective treatment plans.

[0105] Extract and calculate the power spectrum entropy and singular spectrum entropy of microwave signals with different hemorrhage point sizes in hemorrhagic stroke at scales 1 - 10. It is found that the line graphs generated by hemorrhage points of 8 sizes have no intersection points and can be completely separated and distinguished. This lays the foundation for accurately judging and distinguishing 8 sizes of hemorrhage points through a neural network later. The power spectrum entropy and singular spectrum entropy calculated for each scale can be used as features and brought into a suitable neural network model for training to quickly classify the sizes. Compared with the existing technical solutions mentioned above, the present invention makes a contribution in analyzing different hemorrhage point sizes in hemorrhagic stroke and provides a preliminary theoretical basis.

[0106] The method of the present invention for extracting features using multi-scale power spectrum entropy and singular spectrum entropy has the following advantages:

[0107] ① Capturing complexity: The multi-scale power spectrum entropy (MPE) energy feature can analyze the spectral energy distribution of the signal, while the multi-scale singular spectrum entropy (SSE) captures the complexity and structural features of the signal through singular value decomposition.

[0108] ② Multi-scale analysis: It can analyze signal features at different scales, is suitable for processing non-stationary signals, and is also very suitable for subsequent experimental signals.

[0109] ③ Robustness: It has strong anti-interference ability to noise, can perform stably in different noise environments, and also has certain advantages for subsequent processing of experimental signals.

[0110] By extracting the power spectrum entropy and singular spectrum entropy of microwave signals with different hemorrhage point sizes in hemorrhagic stroke, different sizes of cerebral hemorrhage points can be effectively distinguished. The present invention uses multi-scale power spectrum entropy and singular spectrum entropy as feature vectors, trains and analyzes through machine learning algorithms, can establish a mapping relationship model between the size of the cerebral hemorrhage point and the feature vector, realize the rapid classification of the size of the cerebral hemorrhage point, provide valuable reference information for doctors, and help to formulate more accurate and effective treatment plans. In addition, this method has strong anti-interference ability to noise, performs stably in different noise environments, and has certain robustness. In summary, this method has obvious beneficial effects in improving the accuracy, comprehensiveness and stability of the diagnosis of the size of the cerebral hemorrhage point.

[0111] It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0112] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectrum entropy and singular spectrum entropy, characterized by: The following steps are involved: S1: Data acquisition step: acquiring medical image data containing the cerebral hemorrhage area, wherein the medical image data is acquired by a medical imaging device, and preprocessing the acquired medical image data; S2: Multi-scale power spectrum entropy calculation: Use the multi-scale power spectrum entropy algorithm to analyze the image data pre-processed in step S1 and extract the power spectrum entropy characteristics of the bleeding point area; S3: multi-scale singular spectrum entropy calculation: using the multi-scale singular spectrum entropy algorithm to analyze the image data pre-processed in step S1, and extract the power spectrum entropy characteristics of the bleeding point area; S4: Feature Fusion: The multi-scale power spectrum entropy in step S2 and the multi-scale singular spectrum entropy in step S3 are used as feature vectors to perform feature fusion; The feature vectors were trained and analyzed using machine learning algorithms to establish a mapping relationship model between the size of the cerebral hemorrhage point and the feature vectors; S5: Analysis of the size of cerebral hemorrhage points: According to the model trained in step S4, new ultra-wideband microwave signal data is input to calculate its multi-scale power spectrum entropy and singular spectrum entropy; According to the mapping relationship, the size of the cerebral hemorrhage point is calculated.

2. The method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectrum entropy and singular spectrum entropy according to claim 1, characterized in that: The pre-processing in step S1 includes: Edge detection: Use the Canny edge detection algorithm to extract the boundaries of brain tissue structures, including five steps: Gaussian filtering, gradient calculation, non-maximum suppression, double threshold processing, and boundary tracking and connection; Regional division: The internal part is divided into five areas: white matter, gray matter, dura mater, cerebrospinal fluid and skull, and the skin layer is added at the end; Numerical model establishment: The processed numerical model is used as the calculation area of ​​FDTD, a perfect matching layer is set, and the electromagnetic parameters of different brain tissues are assigned according to the first-order approximation of the Debye equation.

3. The method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectrum entropy and singular spectrum entropy according to claim 1, characterized in that: The power spectrum entropy feature of the bleeding point area extracted in step S2 is specifically: (a) Perform multi-scale decomposition on image data; (b) Calculate the power spectrum entropy at each scale; (c) Extract power spectrum entropy features.

4. The method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectrum entropy and singular spectrum entropy according to claim 1, characterized in that: The power spectrum entropy feature of the bleeding point area extracted in step S3 is specifically: (d) performing singular value decomposition on the image data; (e) calculating singular spectral entropy; (f) Extract singular spectral entropy features.

5. The method for analyzing the size of cerebral hemorrhage points based on multi-scale power spectrum entropy and singular spectrum entropy according to claim 1, characterized in that: The machine learning algorithms in step S4 include support vector machine, random forest and neural network algorithms.

Images