Method for detecting arterial plaque based on thermoacoustic imaging technology

Abstract

The application discloses an arterial plaque detection method based on thermoacoustic imaging technology. The method processes the collected ultrasonic signals, and the results are reflected on two different channels in software. After processing by an image reconstruction module, a thermoacoustic imaging graph and a photoacoustic imaging graph are reconstructed, and a vulnerable plaque part is automatically identified. Through mutual verification and self-examination of the processing results of the two kinds of imaging graphs, a suspected vulnerable plaque part is marked and position inversion is performed, necessary basic information is provided for classification and treatment of the vulnerable plaque; the combination of the two imaging modes can provide physiological indexes; the transmitted data is processed in the image reconstruction module through a network, and the application in health prevention and disease diagnosis is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, and particularly relates to an arterial plaque detection method based on thermoacoustic imaging technology. BACKGROUND

[0002] Stroke is the second leading cause of death in the world, and the prevalence of stroke in China is increasing year by year, with an annual growth rate of about 8.7%. Its mortality, morbidity and recurrence rate are also high, which not only endangers human health, but also greatly reduces the quality of life of patients, and brings a heavy burden to families and society. Therefore, early detection and effective prevention are very important. The enlargement or shedding of carotid artery plaques leading to lumen stenosis and occlusion is the most important cause of stroke formation. Research has found that the formation of vulnerable carotid artery plaques is an important cause of carotid atherosclerotic disease leading to cerebrovascular events, and the rupture and shedding of vulnerable carotid artery plaques is the primary factor for the occurrence of stroke. Therefore, the study of the vulnerability of carotid artery plaques has become a hot spot in the prevention and treatment of stroke.

[0003] Although great progress has been made in the prevention and treatment of cardiovascular diseases in recent years, coronary heart disease is still a major cause of harm to human health. Acute coronary syndrome patients have a high risk of death, and the main pathological basis is the rupture of unstable plaques or erosion of plaque surface followed by thrombosis. Unstable plaques are closely related to plaque rupture and major adverse cardiovascular events (MACE), also known as vulnerable plaques or high-risk plaques. Therefore, early identification of vulnerable plaques and intensive intervention accordingly are of great significance in reducing the occurrence of MACE.

[0004] Medical imaging is an indispensable diagnostic technique in the clinical field. Among the existing medical imaging methods, the techniques used in clinical examination of carotid atherosclerotic plaques mainly include ultrasound, high-resolution magnetic resonance imaging (HR-MRI), CT angiography (CTA), etc. Ultrasound is widely used in clinical practice, and its intuitive and non-invasive advantages make it widely used in clinical practice. However, it should be noted that the results of conventional ultrasound examination are highly subjective, and clinicians often prefer to have experienced doctors operate or jointly use other new technologies to assist in diagnosis. In addition, as the only non-invasive technique that can realize in vivo imaging of intracranial arterial wall structure, HR-MRI has great potential and application prospects, but due to its high cost, time-consuming, high noise, and certain contraindications (such as installation of a cardiac pacemaker), its clinical application is also limited. Finally, CTA is inexpensive, fast, and has high resolution, and it has unique advantages in the examination of small intracranial vessels and deep vascular plaques in the internal and external carotid arteries, but it should be noted that there is overlap between the CT density values of lipid cores, fibrous connective tissue, and intraplaque hemorrhage components, which often cannot be accurately distinguished, and the reliability is low. Moreover, it is an invasive detection method. Therefore, developing a new imaging technology to overcome these shortcomings has become a focus in the field of medical imaging.

[0005] A microwave thermoacoustic imaging device and method based on compressed sensing are described in patent document CN102058416A, which solves the problems of complex thermoacoustic imaging process, significant difference in imaging effect, low image resolution, serious image artifact phenomenon, and high hardware cost of the system. However, this patent document does not involve the medical field, and the technical problem solved is different from the technical problem solved by the present patent in the medical field. Patent document CN114943727A describes a carotid plaque echo classification method based on key point detection, which alleviates the problem of difficulty in extracting plaque features due to small plaque size in ultrasound images, and can effectively utilize the information of the plaque region and the entire ultrasound image. It can be used for single-plaque carotid ultrasound image echo classification. However, compared with the thermoacoustic imaging method, this method has low precision and strong subjectivity. SUMMARY

[0006] To solve the above technical problems, the present application proposes an arterial plaque detection method based on thermoacoustic imaging technology. The excitation source system and data acquisition system are arranged to realize the collection and storage of thermoacoustic signals and photoacoustic signals. Further, the image is reconstructed by the image reconstruction module. The neural convolution network is used to intelligently process and analyze the two-channel reconstructed images to feedback the arterial plaque information of the tissue part, and a two-channel self-checking mechanism is established based on historical data, including transfer learning and artificial judgment. The detailed classification formed by the classification database of the verification module is used to predict the development trend of the arterial plaque. The physiological index parameters are reconstructed using the collected thermoacoustic signals and photoacoustic signals to realize real-time monitoring in the boot stage.

[0007] In a preferred scheme, after the intelligent processing and analysis operation is processed by the image reconstruction module, the reconstructed thermoacoustic imaging graph and photoacoustic imaging graph are used as the input of this module, and then based on further algorithm optimization, the vulnerable plaque part is automatically identified;

[0008] The Gaussian mixture model algorithm is combined with deep transfer learning data to automatically extract the region of interest, i.e. ROI;

[0009] The known image is converted into one-dimensional data points as the elements of clustering, and the GMM algorithm is used to classify all points in the image;

[0010] The known image is calibrated, and the ROI with plaque is calibrated as "1" and the rest is calibrated as "0". The calibrated image is trained and features are extracted;

[0011] The five-fold cross-validation method is used for training, the expectation maximization algorithm is used to estimate the GMM parameters, and the EA algorithm is used to calculate the GMM parameters;

[0012] After the iteration is stable, the clustered arterial image is graded according to the artery and its related characteristics;

[0013] Image features are extracted by using a neural convolution network-based plaque detection algorithm;

[0014] All features are connected through a fully connected layer to generate an output value, which is input to a classifier to obtain a final classification result;

[0015] An unknown image is input into an arterial plaque detection model to obtain a preliminary judgment result;

[0016] Thermoacoustic imaging images are used for auxiliary verification, and mutual verification and self-examination of the thermoacoustic imaging images and photoacoustic imaging image processing results are performed;

[0017] If the results of the two groups of images processed by deep learning are different, thermoacoustic imaging is used for self-examination, and a diagnosis opinion is given or actively marked for the next verification;

[0018] The suspected vulnerable plaque part is marked and the position is reversed, and the data processing module matches and marks the inspection tissue position, ultrasonic probe position and data sequence position, thereby providing necessary basic information for vulnerable plaque classification and treatment.

[0019] In the preferred scheme, deep transfer learning is used to apply the information learned from histology and MRI images to the analysis and processing field of two-channel images, characterized in that it comprises the following steps:

[0020] S10, using a photoacoustic image database as an original data set, performing preprocessing operations on the arterial plaque image, i.e. removing unclear images, image scaling and contrast adjustment;

[0021] S20, randomly dividing the photoacoustic image into a training group, a verification group and a test group;

[0022] S30, migrating a convolutional neural network, i.e. CNN, pre-trained parameters in other classification tasks;

[0023] S40, fine-tuning the CNN model full connection layer parameters according to the vulnerable plaque classification task, and building a vulnerable plaque classification diagnosis network model through network optimization;

[0024] S50, testing the network model using an independent test set, and checking the classification of the model through a confusion matrix;

[0025] S60, forming a matching classification standard according to effective historical data, combined with histology and imaging characteristics;

[0026] S70, verifying the diagnostic difference part of the two-channel image according to the standard obtained in step S60, and obtaining a recommended result with the maximum expected value;

[0027] S80, displaying the analysis recommendation result on the digital terminal in the form of a pie chart combined with numbers, and waiting for the determination of the professional personnel.

[0028] In the preferred scheme, the reconstructed sound source image is matched with the electrical parameters and absorption peaks of the definitely diagnosed arterial plaque, and is used as a training sample to train a generative adversarial network;

[0029] The relationship between the pathological characteristics of the arterial plaque and the electrical parameters is clarified based on the physical model and the data-driven model;

[0030] The image information is iterated by applying the fuzzy logic algorithm and the structural similarity index;

[0031] Different noise pixels are classified into homogeneous, detailed and edge regions, and then the best filter is adaptively selected for each region;

[0032] The corresponding contact information is labeled using fine-grained label information;

[0033] The matched data is stored in the database, laying a foundation for further deep imaging and electrical parameter problems of the image analysis model;

[0034] The signal processing and image display of the imaging process of the thermoacoustic imaging technology are optimized.

[0035] In the preferred scheme, the image analysis model is established according to historical diagnosis and treatment data to predict the prognosis effect, recurrence and highest mortality risk of the patient, and the development trend of vulnerable plaques.

[0036] In the preferred scheme, the temperature-related parameters in the reconstructed image are realized by indirect measurement using the collected thermoacoustic signals, and real-time temperature is realized.

[0037] The blood volume, blood glucose and hemoglobin oxygen saturation information are provided by combining the TAI and PAI imaging modes;

[0038] The change of blood volume and the concentration information of blood glucose component are reflected by the intensity of the vascular photoacoustic signal, that is, a prediction fitting model is established by using the difference of photoacoustic signals caused by different glucose concentrations and changes of blood volume, and the correlation between them is analyzed to obtain the concentration of blood glucose component and blood volume information in the tissue;

[0039] Taking blood glucose concentration determination as an example, a harmonic signal showing oscillation and damping is used, which is suitable for frequency domain processing;

[0040] Different frequency components are separated from each other by using fast Fourier transform.

[0041] Based on the liquid photoacoustic resonance theory, the amplitude of the resonance frequency is used to represent the concentration of the solution;

[0042] Based on the spectral specificity, the oxygen saturation of hemoglobin is calculated;

[0043] At the same time, the system algorithm is integrated and loaded on the chip, which can realize the wide promotion of wearable devices and fixed devices, thereby providing relevant protection for disease prevention and healthy life of specific groups.

[0044] In the preferred scheme, the image analysis model keeps the data label category consistent with the target category during the training of the convolutional neural network, divides the image of the arterial plaque into IV categories or 10 categories, or performs detailed classification based on the classification database of the verification module; one of the categories is selected to train the convolutional neural network; the actual problem is solved by using the fine-grained label information; the inference class is completed through the classification tree structure diagram, so that each inference class can obtain the corresponding node; when the input image belongs to any one of the training categories, the inference probability of the image needs to be obtained, otherwise the iterative algorithm is used to solve.

[0045] Compared with the prior art, the microwave thermoacoustic imaging mode has the advantages of high contrast, high resolution and non-invasiveness, and through signal processing (image inversion), the structure and morphology of biological tissues and other information can be reconstructed; the electrical parameters can reflect the pathological state of the tissue in early diagnosis, recovery and rehabilitation, and the change of dielectric properties is often earlier than the change of structure, which can promote the development of preventive medicine; deep transfer learning is used to apply the information learned in other types of images to the analysis and processing field of two-channel images, which effectively solves the problem of small-scale data set of images. BRIEF DESCRIPTION OF DRAWINGS

[0046] The present application will be further described below in conjunction with the drawings and examples:

[0047] Figure 1 is an intelligent processing and analysis module flowchart;

[0048] Figure 2 is a verification module and pathological classification processing flowchart;

[0049] Figure 3 is a physiological index detection flowchart;

[0050] Figure 4 is a schematic diagram of the overall architecture of the arterial plaque detection system based on thermoacoustic imaging technology. DETAILED DESCRIPTION

[0051] As shown in Figures 1 to 4 , an arterial plaque detection method and system based on thermoacoustic imaging technology.

[0052] Example 1:

[0053] Subject A, 55 years old, went to a medical institution for a physical examination on a certain day, and a routine examination was performed on the carotid artery.

[0054] When the target biological tissue to be detected is relatively superficial, a microwave source excitation system is first used:

[0055] The microwave source excitation system is composed of a magnetron and corresponding modulation modules and power supply modules;

[0056] The microwave source is generated by a S-band magnetron (such as e2v, MG5240F) oscillation, connected by a flexible coaxial cable to the end of the waveguide and the open waveguide, which irradiates upward at an angle of 45° to the horizontal plane, and performs thermal acoustic excitation on the subject's neck;

[0057] The touch screen panel is used to control the microwave source, and the microwave energy is output through an N-type coaxial adapter, and a TTL trigger signal output is configured for synchronous operation during data acquisition;

[0058] The thermal acoustic signals propagating from the biological tissue are collected by an ultrasonic transducer (flexible array detector), characterized in that:

[0059] The flexible array detector is used to detect vulnerable plaques on the blood vessel wall (such as carotid plaques);

[0060] The flexible array detector has 128 crystals, which are equally spaced in a self-made ring-shaped instrument support;

[0061] When in use, the crystals in the detector start to rotate and scan as a whole;

[0062] In order to better couple the ultrasonic waves, transformer oil is filled between the detector and the skin, and a TPU film with a thickness of 0.03mm is used to separate the skin and the transformer oil; an ultrasonic coupling agent is used between the skin and the film to couple the ultrasonic waves;

[0063] The collected ultrasonic signals are first amplified by a self-made preamplifier, then converted into digital signals by a channel data acquisition card and transmitted to a computer;

[0064] The image reflecting the microwave energy absorption distribution inside the biological tissue is reconstructed by using a delay and sum algorithm (DAS):

[0065] The signal is discretely processed, and the signal propagation delay time and spatial distance are converted into grid points according to the propagation speed of sound waves in the medium for numerical operation;

[0066] The signals are evenly distributed to each grid point covered by the aperture angle according to the delay time, and a circle is drawn with the distance corresponding to the delay time as the radius and the detector as the center. The grid points on the circle are assigned the signal amplitude of the delay time point;

[0067] Each detector receives an ultrasonic time sequence, and the time sequence is multiplied by the sound speed to obtain a distance sequence;

[0068] The ultrasonic time sequence obtained at each position is back projected onto a circle with the distance sequence as the radius and the position point of the detector as the center (the detector is simplified as a point detector during reconstruction);

[0069] The reconstruction is repeated until the signal inversion of 360 positions is completed, and finally the absorber is enhanced due to coherent superposition;

[0070] In order to improve the image quality and compensate for the defocus distortion, the synthetic aperture focusing technique (SAFT) and the coherent weighting factor are combined to reconstruct the image of the data:

[0071] SAFT synthesizes a large aperture by appropriate delay and superposition of adjacent scan signals, formula:

[0072]

[0073] Where s(i, t) represents the thermoacoustic signal detected at the ith position, Δt i is the time delay applied to the thermoacoustic signal received at the ith position. N represents the total number of adjacent scanning positions included in the SAFT summation.

[0074] In order to reduce the sidelobe and further improve the signal-to-noise ratio, a signal weighting factor is added:

[0075]

[0076] The above two formulas are multiplied to obtain the final thermoacoustic signal:

[0077] s weighted (t)=s SAFT (t)CF(t)

[0078] The signal intensity of each grid point on the distance-detector cross section is delayed and superimposed:

[0079] The principle of DAS is to use the signals collected by the transducers to perform reverse calculation in space, and then superimpose the signals collected by each transducer channel, so as to calculate the signal intensity of each point in space. DAS can be expressed as:

[0080]

[0081] where a(x, y) represents the intensity of the signal at coordinate (x, y); S k (l k ) is the signal amplitude of the kth ultrasonic transducer at a distance l k from the thermoacoustic signal source; l k represents the distance between the kth transducer and the thermoacoustic signal source.

[0082] The transducer wafer position coordinates are initialized according to the wafer arrangement relationship of the annular transducer;

[0083] The distance l k between the kth transducer and the thermoacoustic signal source in the figure and the thermoacoustic signal coordinate (x, y) are related by

[0084]

[0085] Let the thermoacoustic discrete signal time series collected by the data acquisition system be s(i), and the relationship between the thermoacoustic signal sampled by the transducer is as follows:

[0086] s(l) = s(i · v s · t s ) (2-9)

[0087] where s(l) represents the signal amplitude of the ultrasonic transducer at l; i represents the order of the collected signal points; v s represents the speed of sound; and t s represents the sampling interval;

[0088] Since the signal collected by the data acquisition card is discrete, the signal of adjacent sampling points is used to approximate the signal intensity at l, and we have:

[0089]

[0090] In an ideal case, assume that the detection extension lines of each transducer channel of the annular ultrasonic transducer intersect at a focal point (x0, y0), the distance from each transducer to the focal point is the convergence radius r0, and the angle between each transducer and the focal point is θ k , then the distance r0 between the kth transducer and the focal point and the thermoacoustic signal coordinate (x, y) have the relationship:

[0091]

[0092] The expression of the distance l between the thermoacoustic signal source (x, y) and the kth transducer is as follows:

[0093]

[0094] By combining (2-6, 2-7, and 2-12), the signal intensity detected by the annular transducer is:

[0095]

[0096] where a(x, y) represents the intensity of the signal at coordinate (x, y); represents the amplitude of the signal collected by the kth transducer at a distance from the source. From (2-6, 2-7, 2-9, 2-13) we have:

[0097]

[0098] where a(x, y) represents the intensity of the signal at coordinate (x, y); S k represents the discrete time signal sequence collected by the kth transducer; l k represents the distance between the kth transducer and the thermoacoustic signal source; θ k represents the angle between each transducer and the horizontal line.

[0099] Preferably, taking the single detector as the research object, the curve problem is regarded as a straight line problem in the mathematical limit problem, the width of the ultrasonic transducer wafer is used to approximate the corresponding arc length, and then the angle of each ultrasonic transducer wafer locking angle is calculated through the corresponding arc length and angle conversion formula, so as to calculate the corresponding position coordinates (x k , y k ) of each ultrasonic transducer wafer;

[0100] Considering the packaging edge of the ultrasonic transducer, the arc length and angle conversion formula is obtained:

[0101]

[0102] where θ is the angle between the ultrasonic transducer wafer and the x axis; d b is the length of the packaging edge; n is the wafer number; and N is the number of ultrasonic transducer wafers.

[0103] The thermoacoustic image can be calculated through formula (2-13) and formula (3-3), and the tomography of microwave absorption at different layers is obtained;

[0104] Similarly, the delay and stack algorithm is used to reconstruct the data collected by the photoacoustic imaging:

[0105] The goal of image reconstruction is to find the initial sound pressure distribution, which is actually proportional to the product of the light energy distribution and the light absorption coefficient in space;

[0106] The filtered back projection method is derived from the time domain photoacoustic equation:

[0107]

[0108] where H(r, t) is the light source term:

[0109]

[0110] The energy density distribution of the light absorption of the imaged tissue at position r is denoted by P(r, t), and I(t) describes the pulse shape of the laser, where the pulse function of the laser is set as a delta function:

[0111] I(t) = δ(t)

[0112] By using an array to obtain the sound pressure signals received at different positions, imaging can be performed, and we take a circular scan as an example. According to the wave equation, we can obtain:

[0113]

[0114] The original equation is further simplified by using the delay-and-sum algorithm to achieve faster imaging:

[0115]

[0116] where s f (r, t) represents the sound signal intensity of a point in the tissue at a certain time (t time); s i (t) represents the ultrasound signal received from the ith detector at t time; is a weighting coefficient corresponding to the ith detector, and the weighting coefficient is related to the signal amplitude; a delay in time is defined as:

[0117]

[0118] where R i represents the position of the ith detector, r f is the actual position of the point on the image, and v is the propagation speed of ultrasound.

[0119] Enter the data intelligent processing and analysis module:

[0120] The Gaussian mixture model algorithm is combined with prior knowledge to automatically extract the region of interest (ROI):

[0121] The carotid reconstruction image is cropped to remove the region of useless information and only keep the region related to the carotid vascular information;

[0122] The image is converted into one-dimensional data points as the elements of clustering, and the GMM algorithm is used to divide all points in the image into three categories. Finally, the data points are converted into an image and displayed;

[0123] GMM is a commonly used variable distribution model that performs clustering based on a probability model, and its distribution is:

[0124]

[0125] where k is the number of classes, x i represents the i-th object, q k is the probability that the pixel belongs to the k-th class, μ k is the mean vector of the multivariate Gaussian distribution corresponding to the k-th class, Φ k is the covariance matrix of the multivariate Gaussian distribution corresponding to the k-th class.

[0126] Θ = { {q1, q2, …, q k}, { μ1, μ2, …, μ k}, { Φ1, Φ2, …, Φ k}} is estimated by the Expectation Maximization (EM) algorithm.

[0127] The EA algorithm is used to calculate the GMM parameters, and the specific steps are as follows:

[0128] Initialize the model parameters q k (0), μ k (0), Φ k (0).

[0129] E step:

[0130]

[0131] M step:

[0132]

[0133]

[0134]

[0135] where,

[0136]

[0137] Iterate the above steps until the parameters converge.

[0138] The gray value of 0-5 is the carotid artery lumen, and the gray value of 180-190 is the carotid adventitia. According to the gray value characteristics of the blood vessels, the clustered carotid artery images are gray graded, the average value of the pixel values of the three images is calculated, the gray value of the image with the lowest value is mapped to 0, and the gray values of the other two images are mapped to 255, and the three images are fused;

[0139] The image is closed, which can fill the image and keep the position and shape of the image content unchanged;

[0140] Two large connected domains near and far from the blood vessel wall are reserved, and small connected domains in the image are deleted. According to the large centroid of the far blood vessel wall, the near connected domain is deleted, and the final ROI is obtained.

[0141] The convolutional neural network (CNN) is used to automatically learn features from a large data set, and finally all the features are connected through a fully connected layer to generate an output value, and the output value is input into a classifier to obtain a final classification result:

[0142] The image is calibrated, and the ROI with plaque is calibrated as "1", and the rest is calibrated as "0". After calibration, the ROI is extracted and the resolution is adjusted to a uniform size, and the calibration image sequence is randomly mixed to prevent the model from memorizing the data;

[0143] The calibrated image is trained and features are extracted;

[0144] The five-fold cross-validation method is used for training, and the final carotid plaque detection model is obtained after 50 iterations;

[0145] The image is input into the carotid plaque detection model to obtain a preliminary discrimination result;

[0146] The image is used to assist verification, and the thermal acoustic imaging image and the photoacoustic imaging image processing result are verified and self-checked;

[0147] The two groups of images obtained by deep learning are compared, and if the results are different, the thermal acoustic imaging is used for self-checking, and the diagnosis opinion is given or actively marked for the next verification.

[0148] The suspected vulnerable plaque site is marked and the position is reversed, and the data processing module matches and marks the inspection tissue position, ultrasonic probe position and data sequence position, providing necessary basic information for vulnerable plaque classification and treatment;

[0149] Embodiment 2:

[0150] Subject B, 56 years old, went to a medical institution for physical examination, and performed a routine examination on the coronary artery.

[0151] When the target biological tissue to be detected is relatively deep, the microwave source excitation system imaging cannot meet the diagnostic requirements, and an injected current excitation source system (pulse excitation source system) is used:

[0152] The pulse excitation source system includes a signal generator system, a pulse source and a pair of excitation copper electrodes;

[0153] The signal generator inputs a square wave pulse signal to the pulse source to control the working state of the pulse source, and the pulse source uses a power switch tube to control the rapid discharge of the energy storage capacitor;

[0154] Preferably, the applied excitation current is a safe current for human body after safety evaluation, and the current effect has no adverse effect on human body;

[0155] The ultrasonic signal detection system mainly comprises an ultrasonic transducer, a low-noise preamplifier, a filter and a data acquisition system;

[0156] After the ultrasonic transducer detects the ultrasonic signal, the signal is amplified by the low-noise preamplifier and subjected to corresponding band-pass filtering, and the processed sound pressure signal is collected and stored as experimental data for sound source image reconstruction;

[0157] According to the injection current type thermoacoustic imaging theory, the image reconstruction unit reconstructs the image of the thermoacoustic source by using the collected sound pressure data:

[0158] The temperature field heat balance equation is represented as

[0159]

[0160] In the formula, c p is the specific heat capacity, k is the thermal conductivity, and when the heat conduction is ignored, there is

[0161]

[0162] The Joule heat generated in the target body causes thermal expansion to excite ultrasonic waves, and the ultrasonic signals are detected by the ultrasonic probe, and the thermoacoustic wave equation is:

[0163]

[0164] In the injection current type thermoacoustic imaging, the sound pressure wave equation satisfied by the sound wave is combined with the temperature field heat balance equation and the sound pressure wave equation:

[0165]

[0166] In the formula, I(t) is the injected current, p(r, t) is the sound pressure at the position r of the ultrasonic probe at time t, r' is the position of the sound source of the target body, c s represents the sound velocity, and β is the thermal expansion coefficient.

[0167] The sound field at the position of the ultrasonic transducer is obtained by solving the sound pressure wave equation by using the Green function integral method:

[0168]

[0169] Considering the response characteristics h(t) of the ultrasonic transducer, the sound signal received by the probe can be represented as:

[0170]

[0171] where p(r,t) is the final signal measured by the ultrasonic transducer, represents a convolution operation.

[0172] According to the convolution theorem, the deconvolution problem about signal recovery and system identification is solved by means of Fourier transform, and the sound signal received by the ultrasonic probe can be expressed as:

[0173] W(r,w) = P(r,w)H(w)

[0174] where W(r,w), P(r,w), H(w) are the frequency domain representations of w(r,t), p(r,t), h(t) respectively

[0175] Based on the least mean square error estimation, the sound pressure p(r,t) at the transducer position is recovered by using the Wiener filter;

[0176]

[0177] where FFT -1 is the Fourier inverse transform, * represents the conjugate complex number, C = 1 / (a|H(w)|) is a regularization factor, and a is a maturity factor used to match the ratio of the input signal and the noise signal power spectrum in the entire |H(w)| period.

[0178] The sound signal collected by the ultrasonic probe is converted into the original sound signal by using the above formula;

[0179] The collected sound wave is processed by using the filtered back-projection method to reconstruct the sound source distribution of the target body;

[0180] The entire data collection process is controlled by a Labview program, and the image reconstruction is realized by a Matlab program;

[0181] Similarly, the delay-and-sum algorithm is used to reconstruct the photoacoustic imaging collected data:

[0182] The goal of image reconstruction is to find the initial sound pressure distribution, which is actually proportional to the product of the light energy distribution and the light absorption coefficient in space;

[0183] The filtered back-projection method is derived from the time-domain photoacoustic equation:

[0184]

[0185] where H(r,t) is the light source term:

[0186]

[0187] The energy density distribution of light absorption in the imaging tissue at position r, I(t) is used to describe the pulse shape of the laser, where the pulse function of the laser is set as a delta function:

[0188] I(t) = δ(t)

[0189] By using an array to obtain the sound pressure signals received at different positions, imaging can be performed, where a circular scan is taken as an example. According to the wave equation, the following can be obtained:

[0190]

[0191] The original equation is further simplified by using the delay-and-sum algorithm to achieve faster imaging:

[0192]

[0193] Where s f (r, t) represents the sound signal intensity of a point in the tissue at a certain time (t time); s i (t) represents the ultrasound signal received at t time from the i-th detector; is a weighting coefficient corresponding to the i-th detector, which is related to the signal amplitude; a delay in time is defined:

[0194]

[0195] Where R i represents the position of the i-th detector, r f is the actual position of the point on the image, and v is the propagation speed of ultrasound.

[0196] Similarly, the delay-and-sum algorithm is used to reconstruct the data collected by photoacoustic imaging:

[0197] The goal of image reconstruction is to find the initial sound pressure distribution, which is actually proportional to the product of the light energy distribution and the light absorption coefficient in space;

[0198] The filtered back-projection method is derived from the time-domain photoacoustic equation:

[0199]

[0200] Where H(r, t) is the light source term:

[0201]

[0202] The energy density distribution of light absorption in the imaging tissue at position r, I(t) is used to describe the pulse shape of the laser, where the pulse function of the laser is set as a delta function:

[0203] I(t) = δ(t)

[0204] By using the array to obtain the sound pressure signal received at different positions, imaging can be performed, and here we take a circular scan as an example. According to the wave equation, we can obtain:

[0205]

[0206] By using the delay and superposition algorithm, the original equation is further simplified to obtain faster imaging:

[0207]

[0208] where s f (r, t) represents the sound signal intensity of a point in the tissue at a certain time (t time); s i (t) represents the ultrasonic signal received from the ith detector at t time; is a weighting coefficient corresponding to the ith detector, and the weighting coefficient is related to the signal amplitude; a delay amount in time is defined:

[0209]

[0210] where R i represents the position of the ith detector, r f is the actual position of the point on the image, and v is the propagation speed of ultrasound.

[0211] enter the data intelligent processing and analysis module as Figure 2 shown:

[0212] The Gaussian mixture model algorithm is combined with prior knowledge to automatically extract the region of interest (ROI):

[0213] The carotid reconstruction image is cropped to remove the region of useless information and only the region related to the carotid vascular information is retained;

[0214] The image is converted into one-dimensional data points as the elements of clustering, and the GMM algorithm is used to divide all points in the image into three categories. Finally, the data points are converted into an image and displayed;

[0215] GMM is a commonly used variable distribution model based on a probability model for clustering, and its distribution is:

[0216]

[0217] where k is the number of categories, x i represents the ith object, q k is the probability that the pixel belongs to the k category, μ k is the mean vector of the multivariate Gaussian distribution corresponding to the k category, and Φ kCovariance matrix of the multivariate Gaussian distribution corresponding to the kth class.

[0218] Θ={{q1,q2,…,q k},{μ1,μ2,…,μ k},{Φ1,Φ2,…,Φ k}} are estimated by the Expectation Maximization (EM) algorithm.

[0219] The GMM parameters are calculated by using the EA algorithm, and the specific steps are as follows:

[0220] Initialize the model parameters q k (0)、μ k (0)、Φ k (0).

[0221] E step:

[0222]

[0223] M step:

[0224]

[0225]

[0226]

[0227] wherein,

[0228]

[0229] Iterate the above steps until the parameters converge.

[0230] The carotid lumen is taken as the gray value 0-5, and the carotid adventitia is taken as the gray value 180-190. The clustered carotid images are gray-graded according to the gray value characteristics of the blood vessels, the average values of the pixel values of the three images are calculated, the gray value of the image with the lowest value is mapped to 0, and the gray values of the other two images are mapped to 255, and the three images are fused;

[0231] The image is closed, which can fill the image and keep the position and shape of the image content unchanged;

[0232] The two large connected domains near the proximal and distal ends of the blood vessel wall are retained, and the connected domains on the image that are too small are deleted. According to the large centroid of the distal blood vessel wall, the proximal connected domain is deleted, and the final ROI is obtained.

[0233] The convolutional neural network (CNN) is used to automatically learn features from a large data set, and finally all the features are connected through a fully connected layer to generate an output value, and the output value is input into a classifier to obtain a final classification result:

[0234] The image is labeled, and the ROI with plaque is labeled as "1", and the rest is labeled as "0". After labeling, the ROI is extracted and the resolution is adjusted to a uniform size. The labeled image sequence is randomly mixed to prevent the model from memorizing the data;

[0235] The labeled image is trained and features are extracted;

[0236] The five-fold cross-validation method is used for training, and the final carotid plaque detection model is obtained after 50 iterations;

[0237] The image is input into the carotid plaque detection model to obtain the preliminary discrimination result;

[0238] The image is input into the carotid plaque detection model to obtain the preliminary discrimination result;

[0239] The two groups of images obtained by deep learning are compared. If the results are different, use thermal acoustic imaging to check yourself. After judgment, give a diagnosis or actively mark for the next verification.

[0240] The suspected vulnerable plaque site is marked and the position is reversed. The data processing module matches and marks the inspection tissue position, ultrasonic probe position and data sequence position, providing necessary basic information for vulnerable plaque classification and treatment;

[0241] Example 3: Single elderly person C

[0242] During the program running process, the thermal acoustic channel and the wide acoustic channel imaging results are compared and found to be different. Then it enters the verification module as shown in Figure 3

[0243] Deep transfer learning is used to apply the knowledge learned in histology, MRI and other types of images to the analysis and processing field of two-channel images. The photoacoustic image and thermal acoustic image database are used as the original data set. The arterial plaque image is preprocessed by removing unclear images, image scaling and contrast adjustment;

[0244] The reconstructed sound source image is randomly divided into training group, verification group and test group;

[0245] Transfer the convolutional neural network (CNN) parameters pre-trained in other classification tasks;

[0246] According to the vulnerable plaque classification task, fine-tune the CNN model full connection layer parameters, build a vulnerable plaque classification diagnosis network model after network optimization, and form a matching classification standard;

[0247] Test the network model using an independent test set, and view the model classification through the confusion matrix; ​

[0248] Based on the relationship between the pathological features of arterial plaques and features such as electrical parameters and absorption peaks obtained from the prediction module, the diagnostic differences between the two channels of images are self-validated to obtain the recommendation result with the highest expected value.

[0249] The analysis and recommendation results will be displayed on a digital terminal using a combination of pie charts and numbers, awaiting confirmation from professionals.

[0250] Deep transfer learning:

[0251] Based on the known data domain, construct a labeled source domain dataset {x}. sj y sj};

[0252] Based on the target domain data to be diagnosed, construct the corresponding unlabeled target domain dataset {x t} and further divide them into training data and test data;

[0253] The labeled source domain data is input into the constructed corresponding feature encoding network G. j Extracting high-dimensional feature representation {G j (x sj The result is then fed into a classifier to obtain the predicted output {C}. j (G j (x sj ))};

[0254] Input unlabeled target domain training data x t To the corresponding feature encoding network G j Extract the corresponding high-dimensional features {G} j (x sj To obtain the corresponding predicted output {C} j (G j (x sj ))};

[0255] Calculate each coding network (G) using the following formula j (x sj ) and (G j (x tj Maximum mean difference (MMD) loss and corresponding classification loss:

[0256]

[0257] Calculate {C} using the following formula j (G j (x sj )) and C i (G i (x ti )) i≠j Ldisc loss.

[0258]

[0259] According to the overall loss function, the encoding network G j and the classifier C j are trained and the parameters are updated;

[0260] In the test phase, the target domain test data is input into the encoding network to extract features, and the corresponding classifier is used to obtain the probability output of multiple classifiers, and the outputs are summed and averaged to output the final plaque recognition result.

[0261] Embodiment 4: Disease diagnosis process enters the prediction module.

[0262] Prediction module:

[0263] The two-channel reconstructed sound source image is used as a training sample;

[0264] Through unsupervised training of a generative adversarial network (GAN), new sample data is generated to achieve data enhancement;

[0265] The GAN network loss function used here is consistent with the traditional GAN network, and is represented as

[0266]

[0267] In the formula, G is the generation network; D is the discrimination network; x is the real data; z is a random vector; p data is the probability distribution of the real data; p(z) is the probability distribution of the random vector; E is the mathematical expectation; G(z) is the data generated by the generator; D(G(z)) is the output of the discriminator for discriminating the true or false of the generated data;

[0268] The optimization goal of the generator is to minimize the loss function log[1-D(G(z))];

[0269] The optimization goal of the discriminator is to maximize the loss function logD(x)+log[1-D(G(z));

[0270] Both the generator and the discriminator are MLP network structures, and the binary classification cross-entropy function is used as the loss function, and the LeakyReLU function is used as the activation function;

[0271] The Adam optimizer is used to update and optimize the network parameters (i.e. the parameters of the generator and the discriminator) in the constructed GAN model;

[0272] The discriminator with optimized network parameters is used to perform multiple unsupervised training on the generator to optimize its network parameters;

[0273] The ability of long short-term memory neural network (LSTM) to maintain the time sequence relationship of data is utilized to establish a recurrent convolutional generative adversarial network (RCGAN);

[0274] The RCGAN selects an RNN with good prediction effect as a generator, selects a CNN capable of quickly and accurately extracting data features as a discriminator, and adopts a long short-term memory neural network (LSTM);

[0275] The specific working principle of the LSTM is as follows: at the current time t, x t is an input vector, s t-1 is an output at the previous time, c t-1 is a hidden state at the previous time, the value of the current time LSTM memory unit is ct, and the output value is s t .

[0276] i t = σ (w t s t-1 + w ri x t + w ci c t-1 + b i )

[0277] f t = σ (w rf s t-1 + w xf x t + w cf c t-1 + b f )

[0278] c t = f t × c t-1 + i t × σ (w rc s t-1 + w xc x t + b c )

[0279] o t = σ (w ro s t-1 + w xo x t + w co c t-1 + b o )

[0280] s t = o t × tanh (c t )

[0281] The convolutional layer output is:

[0282] m P =f(w′·x p：p+g-1 +b′)

[0283] where m P is the pth feature obtained by convolution calculation; f is the selected activation function; w' and b' are the weight and bias corresponding to each convolution kernel, respectively; the window size of the convolution kernel is g; and x p is the pth input of the input feature x.

[0284] Each of the above features is connected as an output feature, i.e., the output feature M of the convolutional layer is

[0285] M=[m1,m2,...,m n-g+1 ]

[0286] Using max pooling, the most significant feature mq in M is extracted instead of all features in M:

[0287] m q =max[m1,m2,...,m n-g+1 ]

[0288] All the results of max pooling are connected to form the output of the pooling layer, and finally the extracted features are classified by the activation function LeakyReLU;

[0289] The Adam optimizer is used to update and optimize the network parameters (i.e., the parameters of the generator and the discriminator) in the constructed GAN model;

[0290] Deep learning is performed on the historical data and the enhanced data to explore the temporal and spatial trends of the arterial plaque;

[0291] Based on the physical model and the data-driven model, the relationship between the pathological features of the arterial plaque and the disease-related features such as the electrical parameters of the thermoacoustic image and the absorption peaks of the photoacoustic image is elucidated;

[0292] The data such as the electrical parameters of the confirmed thermoacoustic image and the absorption peaks of the photoacoustic image are matched with the disease features and stored in the database, laying a foundation for further deep imaging and image analysis model electrical parameter problems;

[0293] According to the historical diagnosis and treatment data, an image analysis model is established to predict the prognosis effect, recurrence, and highest mortality risk of the patient and the development trend of the vulnerable plaque;

[0294] Image analysis model:

[0295] When training a convolutional neural network, the training data label category is kept consistent with the target category. The images of arterial plaques are classified into IV or 10 categories, or a detailed classification is performed based on the classification database of the validation module to match them.

[0296] Choose one of the categories to train the convolutional neural network;

[0297] Solve practical problems using fine-grained label information;

[0298] By using a classification tree structure diagram, inference classes are completed, and each inference class can obtain the corresponding node; if the input image belongs to any class in the training class, the inference probability of the image is obtained.

[0299] If the subclass also belongs to the reasoning class, then an iterative algorithm is used to solve the problem.

[0300] By combining image information, electrical parameter data, absorption peak wavelengths and their characteristic correlations, a composite criterion is formed to predict patient prognosis, recurrence and maximum mortality risk, and the development trend of vulnerable plaques.

[0301] Example 5:

[0302] like Figure 4 As shown, the physiological indicator detection module:

[0303] Using the collected thermoacoustic signals, the real-time temperature detection information is reconstructed as follows:

[0304] The pressure generated by microwave excitation with a pulse width τ that satisfies stress constraints is

[0305]

[0306]

[0307] Where β is the coefficient of thermal expansion, v s It's the speed of sound, C p τ is the specific heat capacity, and I(r) is the total energy absorbed by a microwave pulse of width τ at position r.

[0308] Q(r,t) is a heating function representing the heat energy absorbed per unit time and per unit mass:

[0309] Q(r,t)≈(σ+w∈0∈″)|E(r,t)| 2 =σ eff |E(r,t)| 2

[0310] σ eff For effective conductivity, since magnetic losses are negligible in the tissue, the initial pressure is:

[0311]

[0312] The main temperature dependence is due to the thermal expansion coefficient and the effective electrical conductivity, which is calculated as a function of temperature using the following formula:

[0313]

[0314] Using both TAI and PAI imaging modes in combination provides additional information such as blood volume (heart failure and shock), blood glucose, hemoglobin oxygenation level (hypoxia), etc.

[0315] Using the intensity of the vascular photoacoustic signal to reflect the changes in blood volume, the concentration information of the blood glucose component, that is, using the difference in photoacoustic signals caused by different glucose concentrations and changes in blood volume to establish a prediction fitting model, analyzing the correlation between the two to obtain the concentration of the blood glucose component and the blood volume information inside the tissue;

[0316] Taking blood glucose concentration determination as an example:

[0317] The processing of the photoacoustic signal is frequency domain processing;

[0318] Using a resonant signal that shows oscillation and damping, which is suitable for frequency domain processing;

[0319] Fast Fourier Transform (FFT) is used to separate different frequency components from each other;

[0320] According to the following theory, the amplitude of the resonant frequency can represent the concentration of the solution:

[0321] Ignoring the thermal diffusion and viscosity of the liquid, when the laser irradiates a weakly absorbing solution, the medium absorbs the laser energy and converts it into heat. The temperature gradient causes expansion and compression in the solution, generating thermoelastic waves. The photoacoustic wave is described as:

[0322]

[0323] p(r, t) represents the acoustic pressure of the photoacoustic signal, r is the spatial distribution of the acoustic pressure p(r, t); t is the time; β is the volumetric thermal expansion coefficient, c p is the specific heat at constant pressure, c0 is the sound speed, and H(r, t) is a function defined as the heat deposited in the medium per unit volume and time.

[0324] To solve equation (1), Fourier transform is performed

[0325]

[0326] p(r, t) = ∫p(r, w)e -j dw

[0327] H(r,t) = ∫H(r,w)e -jwt dw

[0328] In equation (2), p(r,w) can be expressed as the sum of all positive modes:

[0329] p(r,w) = ∑ J A j (w)p j (r) (4)

[0330] In equation (4), j is the normal vibration of the mode, p j (r) is the spatial distribution, A j (w) is the amplitude of the different modes varying with time, p j (r) is expressed in cylindrical coordinates as r is the radial, is the azimuthal axis, and z is the longitudinal axis.

[0331] Since the walls of the photoacoustic cell are rigid, the vertical component of the velocity to the wall is zero, so:

[0332]

[0333] In equation (5), j m (k r r) is the mth order Bessel function, α mn is the nth root in the mth order Bessel function, and l, a are the length and radius of the photoacoustic cell.

[0334] Substituting equation (4) into equation (2) gives the amplitude as follows:

[0335]

[0336]

[0337] In equation (6), V c is the volume of the photoacoustic cell, and a mass factor Q is introduced to account for the loss of viscosity and damping;

[0338] When the laser passes through a weakly absorbing solution, the energy absorbed by the solution can be expressed as H(r,w) = aI(r,w), where I(r,w) is the light intensity and a represents the light absorption coefficient. Substituting H(r,w) into equation (6) gives:

[0339]

[0340] Setting that is, A0 represents the concentration of the weakly absorbing solution, and the amplitude A j (r,w) can be simplified as

[0341]

[0342] The resonant frequency of the determination is defined as w i The photoacoustic pressure can be described as

[0343] p = A0A i (w i )p i (r) (10)

[0344] From equation (10), the amplitude of a certain frequency can be used to determine the concentration of the solution, such as the resonant frequency;

[0345] Hemoglobin oxygen saturation index generation process:

[0346] When calculating the content of oxygenated hemoglobin, deoxygenated hemoglobin and blood oxygen saturation parameters, the wavelength is λ i The absorption coefficient of blood at this time can be expressed as

[0347]

[0348] Where ε HbR (λ i ) and are the molar extinction coefficients of deoxygenated hemoglobin and oxygenated hemoglobin at a specific wavelength, [HbR] and [HbO2] represent the molar concentrations of deoxygenated hemoglobin and oxygenated hemoglobin in blood respectively, then convert the above equation into a matrix equation form:

[0349]

[0350] The calculation method of blood oxygen saturation is:

[0351]

[0352] The ratio of HbR and HbO2 is obtained by least squares method:

[0353]

[0354] Where,

[0355] Since the photoacoustic signal of the same absorbing substance is proportional to the light absorption coefficient, the absorption coefficient in equation (2-37) can be replaced by the photoacoustic signal, then equation (2-37) can be transformed into:

[0356]

[0357] Where,

[0358] At the same time, the system algorithm is integrated on the chip, which can realize the wide promotion of wearable devices and fixed devices, thereby providing relevant protection for disease prevention and healthy life of specific groups.

[0359] Embodiment 6:

[0360] After processing the collected signals, the color change corresponds to the risk level change. After intelligent identification, the ROI (region of interest) with the highest risk of arterial plaque is enlarged and displayed on the terminal (department computer). The on-duty doctor combines more evidence to make further judgments, that is, the combination of intelligence and artificiality.

[0361] The twin network matches the ultrasound pictures obtained by the current mainstream detection method of arterial plaque with the thermoacoustic imaging pictures, further makes full use of the historical data of thermoacoustic imaging, and improves the accuracy of the detection model.

[0362] The process of "laser irradiation generates a small heat source -> generates a thermoacoustic wave and is received by the receiving device" can be integrated into an integrated automatic device. The signal collection front end is replaced according to the characteristics of carotid plaque and coronary plaque, and then it is connected with the device gear mode.

[0363] In addition, the comsol simulation software improves the construction process of the model and the identification of medical knowledge compared with the prior art (such as the process of visual recognition, the integration of infrared cameras and thermoacoustic imaging technology signal collection end, etc.).

[0364] Embodiment 7:

[0365] As shown in Figure 4 , an arterial plaque detection system based on thermoacoustic imaging technology is composed of an excitation source system, a data acquisition system and an image reconstruction module. The excitation source system is divided into two modes according to the depth of the actual arterial plaque to be detected: microwave thermoacoustic imaging and injected current thermoacoustic imaging. The excitation source system also includes a laser, an optical parametric oscillator, a lens group, an amplification circuit and a multiplexer, and a computer. The excitation source system and the data acquisition system are integrated in a ring-shaped hardware shell. The image reconstruction module is arranged in the computer, and mobile networks and cloud services can be applied as transmission media according to different application scenarios.

[0366] The system is composed of an excitation source system, a data acquisition system and an image reconstruction module. After processing by the excitation source system and the data acquisition system, the collected ultrasound signals are processed, and the results are reflected on two different channels in the software.

[0367] The excitation source system and the data acquisition system can be integrated into a handheld, portable, wearable and other integrated devices. The data transmitted back through the network is processed in the image reconstruction module, which meets the application of home self-testing, health prevention and disease diagnosis.

[0368] The injection current type thermoacoustic imaging has the advantages of increased penetration depth, improved conversion efficiency, simple, convenient and effective image reconstruction with lower energy, larger detection depth than magnetic thermoacoustic imaging, and wide application prospect.

[0369] The above-mentioned embodiments are only preferred technical solutions of the present application, and should not be regarded as limitations of the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, and the equivalent replacement solutions of the technical features recited in the claims are within the protection scope. That is, equivalent replacement improvements within this range are also within the protection scope of the present application.

Claims

1. An arterial plaque detection method based on thermoacoustic imaging technology, characterized in that: The heat acoustic signal and the photoacoustic signal are collected and stored by the arranged excitation source system and the data acquisition system, and an image is further reconstructed according to an image reconstruction module; arterial plaque information of a tissue part is fed back after the two-channel reconstructed images are intelligently processed and analyzed by using a neural convolution network, and a self-checking and troubleshooting mechanism of the two channels is established based on historical data, including transfer learning and artificial judgment; the development trend of the arterial plaque is predicted based on a detailed classification formed by a classification database of the verification module; physiological index parameters are reconstructed by using the collected heat acoustic signal and photoacoustic signal, and real-time monitoring in a starting stage is realized; The intelligent processing and analysis operation is specifically as follows: after being processed by the image reconstruction module, heat acoustic imaging images and photoacoustic imaging images reconstructed are taken as inputs of the module, and then, based on further algorithm optimization, vulnerable plaque parts are automatically identified; A Gaussian mixture model algorithm is combined with deep transfer learning data to automatically extract a region of interest, i.e., ROI; Known images are converted into one-dimensional data points as clustering elements, and all points in the images are classified by using a GMM algorithm; The known images are calibrated, wherein the ROI with a plaque is calibrated as "1", and the rest is calibrated as "0"; the calibrated images are trained and features are extracted; A five-fold cross-validation method is used for training, a GMM parameter is estimated by using an expectation maximization algorithm, and the GMM parameter is calculated by using an EA algorithm; After iteration stabilization, the clustered arterial images are graded in gray scale according to arteries and related characteristics thereof; Image features are extracted by using a plaque detection algorithm based on a neural convolution network; All features are connected through a fully connected layer to generate an output value, and the output value is input into a classifier to obtain a final classification result; Unknown images are input into an arterial plaque detection model to obtain a preliminary discrimination result; The heat acoustic imaging images are used for auxiliary verification, and the heat acoustic imaging images and the photoacoustic imaging images are verified with each other and self-checked; When the results of the two groups of images processed by deep learning are different, the heat acoustic imaging is used for self-checking, and a diagnosis opinion is given after judgment or active marking is performed, and the like; A suspected vulnerable plaque part is marked and position inversion is performed, and the data processing module matches and marks an examination tissue position, an ultrasonic probe position and a data sequence position, thereby providing basic information for vulnerable plaque classification and treatment.

2. The method for detecting arterial plaque based on thermoacoustic imaging technology according to claim 1, characterized in that: Deep transfer learning is used to apply information learned in histology and MRI images to the analysis and processing field of two-channel images, and the method is characterized in that the following steps are included: S10, taking a photoacoustic image database as an original data set, performing a preprocessing operation on arterial plaque images, i.e., removing unclear images, image scaling and contrast adjustment; S20, randomly dividing the photoacoustic images into a training group, a verification group and a test group; S30, migrating a convolutional neural network, i.e., CNN, in parameters pre-trained in other classification tasks; S40, fine-tuning CNN model full connection layer parameters according to a vulnerable plaque classification task, and building a vulnerable plaque classification diagnosis network model after network optimization; S50, testing the network model by using an independent test set, and checking a classification condition of the model by using a confusion matrix; S60, according to the effective historical data, combined with histology and imaging features, form a matching classification standard; S70, according to the standard obtained in step S60, verify the diagnostic difference part of the two channel images, and obtain the recommended result with the maximum expected value; S80, the analysis recommendation result is displayed on the digital terminal in the form of pie chart combined with numbers, waiting for the determination of professional personnel.

3. The method for detecting arterial plaque based on thermoacoustic imaging technology according to claim 1, wherein the historical data features are: The reconstructed sound source image is matched with the electrical parameters and absorption peaks of the accurately diagnosed arterial plaque, and is used as a training sample to train a generative adversarial network; The relationship between the pathological features of the arterial plaque and the electrical parameters is clarified based on the physical model and the data-driven model; Fuzzy logic algorithm and structural similarity index are applied to iterate the image information; Different noise pixels are classified into homogeneity, detail and edge, and then the best filter is adaptively selected for each region; The corresponding contact information is labeled using fine-grained label information; The matched data is stored in the database, laying a foundation for further deep imaging and electrical parameter problems of the image analysis model.

4. The method of claim 1, wherein the method is based on thermoacoustic imaging technology. Further comprising the following steps: An image analysis model is established according to historical diagnosis and treatment data to predict the prognosis effect, recurrence and highest mortality risk of the patient, and the development trend of vulnerable plaques.

5. The method for detecting arterial plaque based on thermoacoustic imaging technology according to claim 1, wherein: Real-time temperature is realized by indirectly measuring temperature-related parameters in the reconstructed image using the collected thermoacoustic signals; Blood volume, blood glucose and hemoglobin oxygen saturation information are provided by combining the TAI and PAI imaging modes; The intensity of the vascular photoacoustic signal is used to reflect the change of blood volume and the concentration information of blood glucose components, that is, a prediction fitting model is established using the difference in photoacoustic signals caused by different glucose concentrations and changes in blood volume, and the correlation between the two is analyzed to obtain the concentration and blood volume information of the blood glucose components in the tissue; Taking blood glucose concentration determination as an example, a harmonic signal showing oscillation and damping is used, which is suitable for frequency domain processing; Fast Fourier transform is used to separate different frequency components from each other; Based on the liquid photoacoustic resonance theory, the amplitude of the resonance frequency is used to represent the concentration of the solution; Based on the spectral specificity, the hemoglobin oxygen saturation is calculated; The above algorithms are integrated and loaded on a chip.

6. The method for detecting arterial plaque based on thermoacoustic imaging technology according to claim 4, characterized in that: The characteristics of the image analysis model are: When training the convolutional neural network, the data label category of the training is consistent with the target category, the image of the arterial plaque is divided into four categories or ten categories, or a detailed classification database based on the verification module is matched; One of the categories is selected to train the convolutional neural network; Fine-grained label information is used to solve actual problems; Through the classification tree structure diagram, reasoning is completed, and each reasoning class can obtain the corresponding node; when the input image belongs to any one of the training categories, the reasoning probability of the image needs to be obtained, otherwise the iterative algorithm is used to solve it.

Images