Aortic dissection intelligent diagnosis method and system based on ascension AI and physical perception deep learning

Abstract

The application discloses an aortic dissection intelligent diagnosis method and system based on Ascend AI and physical perception deep learning, and belongs to the technical field of intelligent medical instruments. The method comprises the following steps: generating a sweep excitation by a DDS module controlled by an Ascend AI processor and synchronously collecting a magnetoelastic sensor response signal; reconstructing the signal by using a VMD algorithm for adaptive denoising to obtain a high signal-to-noise ratio effective signal; performing spectrum analysis on the signal to extract physical characteristics, and performing double-parameter temperature compensation in combination with temperature and amplitude; inputting the calibrated characteristics and the signal into a pre-trained PG-ANet for double-flow fusion reasoning to output a D-dimer concentration prediction value and a risk level; evaluating the prediction confidence based on a magnetoelastic mass-frequency linear physical model, and if the confidence is insufficient, the excitation parameters are dynamically adjusted and re-measured until a high-trust diagnosis result is output. The application solves the problems of difficult signal extraction in a high-viscosity blood environment, great temperature interference, poor AI model generalization due to a small sample, and physically untrustworthy prediction results.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent medical instruments, and more particularly to an aortic dissection intelligent diagnosis method and system based on Ascending AI and physical perception deep learning. BACKGROUND

[0002] In the field of early diagnosis of aortic dissection (AD), medical imaging and biomarker detection are two core but fragmented technical paths. Traditional imaging, as the gold standard, can accurately display anatomical structures, but is expensive, complex, and not suitable for rapid bedside screening, which can easily delay rescue opportunities. Blood tests based on D-dimer and other markers are fast and convenient, but traditional laboratory methods are time-consuming, and their specificity is insufficient to distinguish AD from other thrombotic diseases. In recent years, the point-of-care testing (POCT) technology based on magnetoelasticity has attracted much attention due to its wireless, passive, and fast characteristics, but its clinical translation faces fundamental bottlenecks. First, the high viscosity of blood leads to a dramatic increase in sensor signal damping and resonance peak broadening, making it difficult for traditional spectral analysis methods to extract weak and effective features from strong noise, resulting in low detection sensitivity. Second, the core material of the sensor is extremely sensitive to temperature, and small fluctuations in the clinical environment can cause significant frequency drift, resulting in false positive errors and making it difficult to meet the reliability requirements of medical treatment. Existing technologies attempt to introduce AI models to improve intelligence, but AD clinical samples are scarce, and simple data-driven "black box" models are prone to overfitting in small samples, have poor generalization ability, and may output absurd results that violate physical laws, lacking explainability and clinical credibility. Existing improvement schemes are mostly local optimizations of algorithms or hardware, and fail to build a complete "precise perception-intelligent decision-dynamic control" closed loop. The system cannot reliably "perceive" and reconstruct high-fidelity resonance signals in a complex blood environment; lacks a "decision" model that integrates physical laws and data features to achieve high-robustness concentration and risk assessment; and cannot perform dynamic adjustment of front-end excitation parameters based on real-time confidence to achieve adaptive measurement.

[0003] Therefore, there is an urgent need in the art for an intelligent system that integrates software and hardware, mechanism and data, to systematically solve core problems such as signal extraction, environmental interference, model generalization, and closed-loop optimization, and to promote the paradigm shift of AD diagnosis from "passive offline detection" to "active real-time perception". SUMMARY

[0004] In order to overcome the deficiencies in the prior art, the present application provides an aortic dissection intelligent diagnosis method and system based on Ascending AI and physical perception deep learning to solve the above problems.

[0005] In order to achieve the above purpose, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides an aortic dissection intelligent diagnosis method based on Ascending AI and physical perception deep learning, comprising:

[0007] S1: an Ascending AI processor controls a direct digital frequency synthesis (DDS) module to send a sweep excitation signal to a magnetoelastic sensor; at the same time, an original time domain response signal of the magnetoelastic sensor under the action of a biomarker is synchronously collected through an interface;

[0008] S2: the original time domain response signal is input to the Ascending AI processor, and a variational mode decomposition algorithm (VMD) is used to denoise and reconstruct the original time domain response signal to obtain an effective signal with high signal-to-noise ratio;

[0009] Further, the original time domain response signal is denoised and reconstructed by the variational mode decomposition algorithm (VMD) to obtain an effective signal with high signal-to-noise ratio, specifically comprising:

[0010] S21: the original time domain response signal is decomposed into a plurality of intrinsic mode function components with different center frequencies by the Ascending AI processor;

[0011] Optionally, step S21 specifically comprises:

[0012] The Ascending AI processor is used to minimize the sum of the estimated bandwidths of the plurality of intrinsic mode function components with different center frequencies as an optimization target, and to equalize the sum of the superpositions of all intrinsic mode function components with different center frequencies as a constraint condition, to construct and solve a constrained variational problem, so as to adaptively decompose the original time domain response signal into the plurality of intrinsic mode function components with different center frequencies.

[0013] S22: the permutation entropy value of each intrinsic mode function component is calculated;

[0014] S23: the intrinsic mode function component with a permutation entropy value higher than an entropy value threshold is determined as a noise dominant component and is removed;

[0015] S24: the remaining intrinsic mode function components are determined as signal dominant components and are superimposed to obtain an effective signal with high signal-to-noise ratio.

[0016] S3: the effective signal is subjected to spectrum analysis, and a physical feature vector is extracted, the physical feature vector at least including a measured resonance frequency, a quality factor and a maximum amplitude, and temperature drift compensation is performed in combination with a measured ambient temperature;

[0017] Further, the effective signal is subjected to spectrum analysis, and a physical feature vector is extracted, specifically comprising:

[0018] performing fast Fourier transform (FFT) on the effective signal with high signal-to-noise ratio to obtain a complex impedance spectrum;

[0019] extracting, from the complex impedance spectrum, a physical feature vector including at least a measured resonance frequency, a quality factor and a maximum amplitude.

[0020] Optionally, temperature drift compensation is performed in combination with the measured ambient temperature, including:

[0021] S31: obtaining the measured ambient temperature and extracting the maximum amplitude and the measured resonance frequency from the physical feature vector;

[0022] S32: calculating a temperature and amplitude joint correction amount based on a difference between the measured ambient temperature and a preset reference temperature, the maximum amplitude, a preset reference amplitude, and in combination with a preset thermal drift coefficient;

[0023] S33: subtracting the temperature and amplitude joint correction amount from the measured resonance frequency to obtain a calibrated resonance frequency, and updating the measured resonance frequency in the physical feature vector based on the calibrated resonance frequency.

[0024] S4: inputting the physical feature vector after temperature drift compensation and the effective signal into a pre-trained physical perception attention neural network (PG-ANet) for fusion inference, and outputting a biomarker concentration prediction value and an aortic dissection risk level corresponding to the concentration prediction value;

[0025] Further, the physical perception attention neural network (PG-ANet) specifically includes:

[0026] a time series feature flow branch that adopts a one-dimensional residual network in combination with an attention module to extract local morphological features from the effective signal and adaptively weight the resonance peak region;

[0027] a physical mechanism flow branch that adopts a multi-layer fully connected network to process the physical feature vector;

[0028] a feature splicing layer that fuses the output results of the time series feature flow branch and the output results of the physical mechanism flow branch to obtain fused features;

[0029] a regression analysis layer that analyzes the fused features to obtain a final prediction result.

[0030] Further, the pre-trained physical perception attention neural network (PG-ANet) specifically includes:

[0031] obtaining a plurality of training data sets, the training data sets including at least sample concentration values and corresponding magneto-elastic sensor frequency shift measured values;

[0032] The parameters of the Physically Aware Attention Neural Network (PG-ANet) are optimized based on the total loss function, thereby guiding the prediction results of the Physically Aware Attention Neural Network (PG-ANet) to converge in a direction that conforms to physical laws; the total loss function includes a physical constraint loss function and a mean squared error loss function.

[0033] The physical constraint loss function is constructed based on the mass-frequency linear theoretical model of the magnetoelastic sensor, and specifically includes:

[0034] For each training data sample, the predicted sensitivity is calculated based on the concentration prediction value output by the Physical Awareness Attention Neural Network (PG-ANet) and the measured frequency shift value of the magnetoelastic sensor.

[0035] The physical constraint penalty is calculated based on the predicted sensitivity, the theoretical sensitivity of the magnetoelastic sensor, and the tolerance threshold.

[0036] Optionally, step S4 specifically includes:

[0037] S41: Input the temperature drift-compensated physical feature vector into the physical mechanism flow branch to extract the nonlinear correlation features between physical parameters; at the same time, input the effective signal into the time-series feature flow branch for feature extraction, and adaptively enhance the feature weights of the resonance peak region of the effective signal to obtain deep time-series features;

[0038] S42: Input the nonlinear correlation features between the physical parameters and the deep temporal features into the feature splicing layer for fusion to form a fused feature;

[0039] S43: Input the fusion features into the regression analysis layer and output the predicted concentration values ​​of the biomarkers and the corresponding aortic dissection risk levels.

[0040] S5: Evaluate the confidence score of the concentration prediction value in step S4; if the confidence score is lower than the preset threshold, the Ascend AI processor dynamically adjusts the parameters of the direct digital frequency synthesis (DDS) module and returns to step S1 to re-execute; otherwise, output the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value.

[0041] Further, step S5 specifically includes:

[0042] The concentration prediction value is physically consistent verified based on a mass-frequency linear physical model of magneto-elastic sensing, to generate the confidence score, if the confidence score is lower than the preset threshold, the center frequency in a direct digital frequency synthesis (DDS) module is dynamically adjusted by the Ascending AI processor and the first peak is locked, and step S1 is re-executed, otherwise the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value are output.

[0043] It can be understood that the execution subject of the present application can be an aortic dissection intelligent diagnosis system based on Ascending AI and physical perception deep learning, and can also be a terminal or a server, which is not limited here. The system is taken as an example for description in the embodiments of the present application.

[0044] In a second aspect, the present application provides an aortic dissection intelligent diagnosis system based on Ascending AI and physical perception deep learning, comprising:

[0045] A signal excitation and collection module is configured to control a direct digital frequency synthesis (DDS) module to send a sweep excitation signal to a magneto-elastic sensor by an Ascending AI processor; at the same time, the original time domain response signal of the magneto-elastic sensor under the action of a biomarker is synchronously collected through an interface;

[0046] A VMD denoising and reconstruction module is configured to input the original time domain response signal into the Ascending AI processor, and utilize a variational mode decomposition algorithm (VMD) to denoise and reconstruct the original time domain response signal to obtain an effective signal with high signal-to-noise ratio;

[0047] A physical feature extraction and calibration module is configured to perform frequency spectrum analysis on the effective signal, extract a physical feature vector, and the physical feature vector at least includes a measured resonance frequency, a quality factor and a maximum amplitude, and combine a measured ambient temperature to perform temperature drift compensation;

[0048] A PG-ANet fusion inference module is configured to input the temperature drift compensated physical feature vector and the effective signal into a pre-trained physical perception attention neural network (PG-ANet) for fusion inference, and output a concentration prediction value of the biomarker and an aortic dissection risk level corresponding to the concentration prediction value;

[0049] A closed-loop control module is configured to evaluate the confidence score of the concentration prediction value in step S4; if the confidence score is lower than the preset threshold, the Ascending AI processor dynamically adjusts the parameters of a direct digital frequency synthesis (DDS) module, and returns to step S1 for re-execution, otherwise the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value are output.

[0050] In summary, the present application creatively constructs a new paradigm of intelligent diagnosis of aortic dissection by deeply integrating the physical mechanism of magnetoelastic sensing, signal processing and Ascending AI computing power, which is an integration of "high-fidelity signal perception-physically interpretable intelligent decision-making-confidence-driven closed-loop control". Under the premise of ensuring rapid detection and bedside applicability, the paradigm systematically overcomes the core technical bottlenecks such as signal extraction in a high-viscosity blood environment, large interference of clinical temperature fluctuations, poor AI model generalization under small sample conditions, and physically untrusted prediction results, and realizes the fundamental change from traditional "passive, offline single measurement" to "active, real-time perception-optimization". The present application not only significantly improves the accuracy, reliability and robustness of early aortic dissection diagnosis, but also provides an innovative software and hardware collaborative framework and solution for other POCT detection of trace markers in complex biological environments. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only exemplary, and those skilled in the art can also obtain other implementation drawings from the provided drawings without creative labor.

[0052] The structures, proportions, sizes, etc. shown in the present specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and do not define the limiting conditions for the implementation of the present application, so they do not have technical significance. Any modification of the structure, change of the proportion relationship or adjustment of the size, without affecting the effect and purpose that the present application can produce, should still fall within the scope of the technical content disclosed by the present application.

[0053] Figure 1 Flowchart of the intelligent diagnosis method of aortic dissection based on Ascending AI and physical perception deep learning;

[0054] Figure 2 Schematic diagram of performing VMD on the original time domain response signal for denoising and reconstruction;

[0055] Figure 3 Architecture diagram of a physical perception attention neural network (PG-ANet);

[0056] Figure 4 Schematic diagram of the overall hardware architecture and data link of the intelligent diagnosis system of aortic dissection;

[0057] Figure 5 System module diagram of an intelligent diagnosis of aortic dissection based on Ascending AI and physical perception deep learning. DETAILED DESCRIPTION

[0058] The specific embodiments of the present application will be described in the following detailed description, which should be considered in conjunction with the accompanying drawings. It will be apparent to those skilled in the art from this disclosure that various advantages can be achieved in the preferred embodiments and as such, the present application should not be deemed to be limited only to such preferred embodiments. Rather, other embodiments within the scope of the present application will be recognized by those skilled in the art.

[0059] It should be noted that the terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. The terms "include" and "have" and any variations thereof used in the present application are intended to cover non-exclusive inclusion. The term "a plurality of" used in the present application refers to two or more. The term "and / or" used in the present application refers to one of the options or any combination of the options.

[0060] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings.

[0061] Embodiment 1, refer to Figure 1 The aortic dissection intelligent diagnosis method based on Ascending AI and physical perception deep learning in the embodiments of the present application includes:

[0062] S1: The Ascending AI processor controls the direct digital frequency synthesis (DDS) module to send a sweep excitation signal to the magnetoelastic sensor; at the same time, the original time domain response signal of the magnetoelastic sensor under the action of the biomarker is synchronously collected through the interface;

[0063] In the present embodiment, the Ascending AI processor controls the DDS module to emit a coarse sweep signal (for example, center frequency 50 kHz, span ±50 kHz) to the magnetoelastic sensor through the SPI bus, and synchronously collects the original time domain response signal of the magnetoelastic sensor under the action of the biomarker through the I2C interface;

[0064] The biomarker is D-dimer in blood, and the original time domain response signal is the induced voltage of the magnetoelastic sensor vibration, which is physical observation data containing amplitude, phase, and frequency complete information.

[0065] S2: input the original time-domain response signal into the Ascending AI processor, denoise and reconstruct the original time-domain response signal by using a variational mode decomposition algorithm (VMD) to obtain an effective signal with high signal-to-noise ratio;

[0066] In the embodiment, in step S2, referring to Figure 2 , the Ascending AI processor receives the original time-domain response signal, and schedules a neural network processing unit (NPU) in the Ascending AI processor to perform a variational mode decomposition algorithm (VMD) for denoising and reconstruction.

[0067] The denoising and reconstruction of the original time-domain response signal by using the variational mode decomposition algorithm (VMD) to obtain an effective signal with high signal-to-noise ratio specifically includes:

[0068] S21: decompose the original time-domain response signal into a plurality of intrinsic mode function components with different center frequencies by using the Ascending AI processor;

[0069] Further, step S21 specifically includes:

[0070] By using the Ascending AI processor, an optimization objective of minimizing the sum of estimated bandwidths of the plurality of intrinsic mode function components with different center frequencies is constructed, and a constraint condition that the sum of superpositions of all intrinsic mode function components with different center frequencies is equal to the original time-domain response signal is constructed, a constrained variational problem is solved, and thus the original time-domain response signal is adaptively decomposed into the plurality (K) of intrinsic mode function components (IMF1-IMFK) with different center frequencies;

[0071] The specific calculation process is as follows:

[0072] ;

[0073] ;

[0074] wherein, is a mode component obtained by decomposition, is a center frequency of each mode, is an original input signal, is a Dirac function, represents a convolution operation, s.t. is a constraint condition, and j is an imaginary unit.

[0075] S22: calculate an permutation entropy value (PE) of each intrinsic mode function component (IMF);

[0076] S23: determine an intrinsic mode function component with a permutation entropy value higher than an entropy value threshold as a noise-dominant component and remove the noise-dominant component;

[0077] In step S23, if the PE value of an IMF is greater than the entropy threshold value, it is determined as a noise dominant component and removed; if the PE value is less than or equal to the entropy threshold value; wherein the entropy threshold value is 0.6.

[0078] S24: The remaining intrinsic mode function components are determined as signal dominant components IMF and superimposed to obtain an effective signal with high signal-to-noise ratio .

[0079] S3: The effective signal is subjected to spectral analysis to extract a physical feature vector, and the physical feature vector at least includes a measured resonance frequency, a quality factor and a maximum amplitude, and temperature drift compensation is performed in combination with a measured ambient temperature;

[0080] In the embodiment, the effective signal is subjected to spectral analysis to extract a physical feature vector, specifically including:

[0081] The effective signal with high signal-to-noise ratio is subjected to fast Fourier transform (FFT) to obtain a complex impedance spectrum;

[0082] From the complex impedance spectrum, at least a measured resonance frequency , a quality factor and a maximum amplitude are extracted.

[0083] Temperature drift compensation is performed in combination with a measured ambient temperature, including:

[0084] S31: The measured ambient temperature is obtained, and the maximum amplitude and the measured resonance frequency are extracted from the physical feature vector;

[0085] S32: Based on the difference between the measured ambient temperature and a preset reference temperature, the maximum amplitude, a preset reference amplitude, and in combination with a preset thermal drift coefficient, a temperature and amplitude joint correction amount is calculated;

[0086] S33: The temperature and amplitude joint correction amount is subtracted from the measured resonance frequency to obtain a calibrated resonance frequency, and the measured resonance frequency in the physical feature vector is updated based on the calibrated resonance frequency.

[0087] The calculation process is as follows:

[0088] ;

[0089] wherein, is a measured ambient temperature, is a maximum amplitude, and is a preset thermal drift coefficient.

[0090] S4: inputting the physical feature vector after temperature drift compensation and the effective signal into a pre-trained physical perception attention neural network (PG-ANet) for fusion reasoning, and outputting a concentration prediction value of the biomarker and an aortic dissection risk level corresponding to the concentration prediction value;

[0091] In the embodiment, referring to Figure 3 , the physical perception attention neural network (PG-ANet) specifically includes:

[0092] The time sequence feature flow branch adopts a one-dimensional residual network (1D-ResNet) combined with an attention SE module (Squeeze-and-Excitation). The attention SE module learns the weight of each channel to adaptively enhance the feature response of the formant region and suppress the background noise band, so as to extract local morphological features from the effective signal.

[0093] The physical mechanism flow branch adopts a multi-layer fully connected network to process the physical feature vector.

[0094] More specifically, the physical feature vector after temperature drift compensation is input into a multi-layer fully connected layer (DenseLayers) and a ReLU activation function to extract the nonlinear correlation features between the physical feature vectors after temperature drift compensation.

[0095] The feature splicing layer fuses the output results of the time sequence feature flow branch and the output results of the physical mechanism flow branch to obtain fusion features.

[0096] The regression analysis layer analyzes the fusion features to obtain the final prediction result.

[0097] Further, the pre-trained physical perception attention neural network (PG-ANet) specifically includes:

[0098] A plurality of training data sets are obtained, and the training data set at least includes a sample concentration value and a corresponding magneto-elastic sensor frequency shift measured value;

[0099] The parameters of the physical perception attention neural network (PG-ANet) are optimized based on a total loss function, so as to guide the prediction result of the physical perception attention neural network (PG-ANet) to converge in a direction consistent with the physical law; the total loss function includes a physical constraint loss function and a mean square error loss function.

[0100] The calculation process is as follows:

[0101] ;

[0102] wherein, is the mean square error loss, is a physical constraint term, is a weight coefficient.

[0103] The physical constraint loss function is constructed based on a mass-frequency linear theory model of the magnetoelastic sensor, and specifically includes:

[0104] For each training data sample, a predicted sensitivity is calculated according to a concentration prediction value output by the physical perception attention neural network (PG-ANet) and a frequency shift measured value of the magnetoelastic sensor.

[0105] A physical constraint penalty quantity is calculated based on the predicted sensitivity, a theoretical sensitivity of the magnetoelastic sensor, and a tolerance threshold.

[0106] The calculation process is as follows:

[0107] ;

[0108] wherein, is a frequency shift measured value, is a concentration prediction value, is a theoretical sensitivity of the magnetoelastic sensor, is a tolerance threshold.

[0109] In the embodiment, step S4 specifically includes:

[0110] S41: inputting the physical feature vector after temperature drift compensation to the physical mechanism flow branch to extract nonlinear correlation features between physical parameters; at the same time, inputting the effective signal to the time sequence feature flow branch to extract features, and adaptively enhancing the feature weight of the resonance peak region of the effective signal to obtain deep time sequence features;

[0111] S42: inputting the nonlinear correlation features between the physical parameters and the deep time sequence features to the feature concatenation layer to fuse and form fusion features;

[0112] S43: inputting the fusion features to the regression analysis layer to output the concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value.

[0113] S5: evaluating the confidence score of the concentration prediction value in step S4; if the confidence score is lower than a preset threshold, the Ascending AI processor dynamically adjusts the direct digital frequency synthesis (DDS) module parameters, and returns to step S1 to re-execute, otherwise, outputs the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value.

[0114] In the embodiment, step S5 specifically includes:

[0115] The confidence score is calculated based on a mass-frequency linear physical model of the magnetoelastic sensor, and the concentration prediction value is physically consistent.

[0116] The calculated theoretical frequency shift is:

[0117] ;

[0118] The measured frequency shift is calculated:

[0119] ;

[0120] Wherein, the intrinsic resonance frequency of the magnetoelastic sensor in the unloaded (no biomarker binding) state;

[0121] The confidence score is calculated:

[0122] ;

[0123] If the confidence score is lower than the preset threshold, the Ascend AI processor dynamically adjusts the center frequency in the direct digital frequency synthesis (DDS) module and locks the first peak, and returns to step S1 to re-execute, otherwise outputs the final biomarker concentration prediction value and the corresponding aortic dissection risk level.

[0124] Further, if the confidence score is lower than the preset threshold, the Ascend AI processor triggers an interrupt, instructing the DDS module to adjust the center frequency of the sweep excitation signal to the first peak frequency and narrow the scan bandwidth (for example, to ±5 kHz); wherein the preset threshold is 0.85.

[0125] Embodiment 2, see Figure 4 The overall hardware architecture and data link of the aortic dissection intelligent diagnosis system proposed by the present application include:

[0126] The Ascend AI intelligent computing layer: the development board integrated with the Ascend processor is used as the core control unit. The Ascend processor is integrated with a high-performance CPU and an NPU (neural network processing unit) inside. The CPU is responsible for logical scheduling and peripheral control of the whole system, and the NPU is specially used to accelerate the inference operation of the VMD variational modal decomposition algorithm and the physical perception attention neural network PG-ANet. The on-board high-speed memory is connected to the acquisition interface through a direct memory access (DMA) controller, ensuring zero-copy transmission of large data volume;

[0127] Signal excitation link: the Ascending AI processor is connected with a direct digital frequency synthesizer (DDS) chip (AD9833 is selected in this embodiment) through an SPI bus. The Ascending AI processor writes a frequency control word to the DDS chip, and the DDS chip generates a high-precision sinusoidal sweep signal according to the frequency control word, the precision of which can reach 0.1 Hz. After being amplified by a power amplifier, the sinusoidal sweep signal drives a solenoid to generate an alternating magnetic field;

[0128] Magnetic elastomer sensor: including a magnetic elastomer sensor chip placed in a blood sample to be measured and an external solenoid. The magnetic elastomer sensor chip is made of Metglas 2826MB amorphous alloy strip, with a size of , and is modified on the surface with specific antibodies for D-dimer. Under the excitation of an alternating magnetic field, the magnetic elastomer sensor generates magnetostrictive vibration. The magnetic flux change caused by the vibration of the magnetic elastomer sensor is reversely coupled to the solenoid, forming an induced electromotive force signal;

[0129] High-precision acquisition module: the induced signal is first subjected to signal conditioning by a low-noise preamplifier (LNA), and then enters a 24-bit high-precision analog-to-digital converter (ADC) (AD7768 is selected in this embodiment). The ADC is connected with the Ascending development board through an I2C audio interface. By using the high-bandwidth characteristic of the I2C protocol, the time-domain response signal collected is transmitted in real time to the memory of the Ascending AI processor;

[0130] Environmental monitoring and interaction: equipped with an NTC thermistor for real-time acquisition of environmental temperature (T) ) and transmission to the CPU for temperature compensation; the interaction terminal is connected with the development board through a UART or HDMI interface, for displaying the diagnostic results and system alarms.

[0131] Embodiment 3, see Figure 5 , an embodiment of the present application, provides an aortic dissection intelligent diagnosis system based on Ascending AI and physical perception deep learning, comprising:

[0132] Signal excitation and acquisition module 310, for the Ascending AI processor to control a direct digital frequency synthesis (DDS) module to send a sweep excitation signal to a magnetic elastomer sensor; at the same time, the original time-domain response signal of the magnetic elastomer sensor under the action of a biomarker is synchronously acquired through an interface;

[0133] VMD denoising and reconstruction module 320, for inputting the original time-domain response signal to the Ascending AI processor, and utilizing a variational mode decomposition algorithm (VMD) to denoise and reconstruct the original time-domain response signal, to obtain an effective signal with high signal-to-noise ratio;

[0134] The physical feature extraction and calibration module 330 is configured to perform spectrum analysis on the effective signal, extract a physical feature vector, and perform temperature drift compensation in combination with an actually measured ambient temperature, wherein the physical feature vector at least includes a measured resonance frequency, a quality factor, and a maximum amplitude;

[0135] The PG-ANet fusion inference module 340 is configured to input the physical feature vector after temperature drift compensation and the effective signal into a pre-trained physical perception attention neural network (PG-ANet) for fusion inference, and output a biomarker concentration prediction value and an aortic dissection risk level corresponding to the concentration prediction value.

[0136] The closed-loop control module 350 is configured to evaluate a confidence score of the concentration prediction value in step S4; if the confidence score is lower than the preset threshold, the Ascending AI processor dynamically adjusts a direct digital frequency synthesis (DDS) module parameter, and returns to step S1 for re-execution, otherwise, a final biomarker concentration prediction value and an aortic dissection risk level corresponding to the concentration prediction value are output.

[0137] In summary, the present application creatively constructs a new paradigm of intelligent diagnosis of aortic dissection by deeply fusing the physical mechanism of magnetoelastic sensing, signal processing, and Ascending AI computing power, that is, an integrated new paradigm of "high-fidelity signal perception-physically interpretable intelligent decision-making-confidence-driven closed-loop control". Under the premise of ensuring rapidity of detection and bedside applicability, the paradigm systematically overcomes core technical bottlenecks such as difficulty in signal extraction in a high-viscosity blood environment, large interference of clinical temperature fluctuations, poor AI model generalization under small sample conditions, and physically untrustworthy prediction results, and realizes a fundamental change from traditional "passive, offline single measurement" to "active, real-time perception-optimization". The present application not only significantly improves the accuracy, reliability, and robustness of early aortic dissection diagnosis, but also provides an innovative software and hardware collaborative framework and solution for POCT detection of other trace markers in complex biological environments. Although the present application has been described in detail in the foregoing general description and specific embodiments, some modifications or improvements can be made on the basis of the present application, which is obvious to those skilled in the art. Therefore, these modifications or improvements made on the basis of not deviating from the spirit of the present application are within the scope of the present application.

Claims

1. An intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning, characterized in that, Includes the following steps: S1: The Ascend AI processor controls the direct digital frequency synthesis module to send a sweep frequency excitation signal to the magnetoelastic sensor; at the same time, it synchronously acquires the original time-domain response signal of the magnetoelastic sensor under the action of biomarkers through the interface; S2: Input the original time-domain response signal into the Ascend AI processor, and use the variational mode decomposition algorithm to denoise and reconstruct the original time-domain response signal to obtain an effective signal with a high signal-to-noise ratio; S3: Perform spectrum analysis on the effective signal to extract physical feature vectors. The physical feature vectors include at least the measured resonant frequency, quality factor, and maximum amplitude, and perform temperature drift compensation in combination with the measured ambient temperature. S4: Input the temperature drift-compensated physical feature vector and the effective signal into a pre-trained physical perception attention neural network for fusion reasoning, and output the concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value. Specifically, the attention neural network for physical perception includes: a temporal feature flow branch: employing a one-dimensional residual network combined with an attention module to extract local morphological features from the effective signal and adaptively weighted formant regions; a physical mechanism flow branch: using a multi-layer fully connected network to process the physical feature vector; a feature concatenation layer: fusing the outputs of the temporal feature flow branch and the physical mechanism flow branch to obtain fused features; and a regression analysis layer: analyzing the fused features to obtain the final prediction result. S5: Evaluate the confidence score of the concentration prediction value in step S4; if the confidence score is lower than the preset threshold, the Ascend AI processor dynamically adjusts the parameters of the direct digital frequency synthesis module and returns to step S1 to re-execute; otherwise, output the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value.

2. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, In step S2, the step of using variational mode decomposition algorithm to denoise and reconstruct the original time-domain response signal to obtain an effective signal with a high signal-to-noise ratio specifically includes: S21: The Ascend AI processor decomposes the original time-domain response signal into multiple intrinsic mode function components with different center frequencies; S22: Calculate the permutation entropy value of each of the intrinsic mode function components; S23: Determine the intrinsic mode function components whose entropy values ​​are higher than the entropy threshold as noise-dominant components and remove them; S24: The remaining intrinsic mode function components are determined as the dominant signal components and superimposed to obtain an effective signal with a high signal-to-noise ratio.

3. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 2, characterized in that, Step S21 specifically includes: The Ascend AI processor aims to minimize the sum of estimated bandwidths of multiple intrinsic mode function components with different center frequencies, and uses the constraint that the sum of all intrinsic mode function components with different center frequencies equals the original time-domain response signal as a constraint condition. By constructing and solving a constrained variational problem, the original time-domain response signal is adaptively decomposed into the multiple intrinsic mode function components with different center frequencies.

4. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, Perform spectral analysis on the effective signal to extract physical feature vectors, specifically including: The high signal-to-noise ratio effective signal is subjected to a fast Fourier transform to obtain the complex impedance spectrum; From the complex impedance spectrum, extract physical feature vectors that include at least the measured resonant frequency, quality factor, and maximum amplitude.

5. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, Temperature drift compensation is performed based on the measured ambient temperature, including: S31: Obtain the measured ambient temperature, and extract the maximum amplitude and the measured resonant frequency from the physical feature vector; S32: Based on the difference between the measured ambient temperature and the preset reference temperature, the maximum amplitude, the preset reference amplitude, and in combination with the preset thermal drift coefficient, calculate the joint correction amount of temperature and amplitude; S33: Subtract the temperature and amplitude joint correction from the measured resonance frequency to obtain the calibrated resonance frequency, and update the measured resonance frequency in the physical feature vector based on the calibrated resonance frequency.

6. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, The pre-trained physical perception attention neural network specifically includes: Acquire multiple sets of training datasets, wherein the training datasets include at least sample concentration values ​​and corresponding measured magnetoelastic sensor frequency shift values; The parameters of the physical perception attention neural network are optimized based on the total loss function, thereby guiding the prediction results of the physical perception attention neural network to converge in a direction that conforms to physical laws; the total loss function includes a physical constraint loss function and a mean squared error loss function. The physical constraint loss function is constructed based on the mass-frequency linear theoretical model of the magnetoelastic sensor, and specifically includes: For each training data sample, the prediction sensitivity is calculated based on the concentration prediction value output by the physical perception attention neural network and the measured frequency shift value of the magnetoelastic sensor. The physical constraint penalty is calculated based on the predicted sensitivity, the theoretical sensitivity of the magnetoelastic sensor, and the tolerance threshold.

7. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, Step S4 specifically includes: S41: Input the temperature drift-compensated physical feature vector into the physical mechanism flow branch to extract the nonlinear correlation features between physical parameters; at the same time, input the effective signal into the time-series feature flow branch for feature extraction, and adaptively enhance the feature weights of the resonance peak region of the effective signal to obtain deep time-series features; S42: Input the nonlinear correlation features between the physical parameters and the deep temporal features into the feature splicing layer for fusion to form a fused feature; S43: Input the fusion features into the regression analysis layer and output the predicted concentration values ​​of the biomarkers and the corresponding aortic dissection risk levels.

8. The intelligent diagnostic method for aortic dissection based on Ascend AI and physical perception deep learning according to claim 1, characterized in that, Step S5 specifically includes: Based on the mass-frequency linear physical model of magnetoelastic sensing, the concentration prediction value is physically consistent and a confidence score is generated. If the confidence score is lower than the preset threshold, the Ascend AI processor dynamically adjusts the center frequency in the direct digital frequency synthesis module and locks the first peak value, and returns to step S1 to re-execute. Otherwise, the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value are output.

9. A system for intelligent diagnosis of aortic dissection based on Ascend AI and physical perception deep learning, used to implement the intelligent diagnosis method for aortic dissection based on Ascend AI and physical perception deep learning as described in any one of claims 1-8, characterized in that, include: The signal excitation and acquisition module is used by the Ascend AI processor to control the direct digital frequency synthesis module to send a sweep frequency excitation signal to the magnetoelastic sensor; at the same time, it synchronously acquires the original time-domain response signal of the magnetoelastic sensor under the action of biomarkers through the interface. The VMD denoising and reconstruction module is used to input the original time-domain response signal into the Ascend AI processor, and use the variational mode decomposition algorithm to denoise and reconstruct the original time-domain response signal to obtain an effective signal with a high signal-to-noise ratio. The physical feature extraction and calibration module is used to perform spectrum analysis on the effective signal and extract physical feature vectors. The physical feature vectors include at least the measured resonant frequency, quality factor and maximum amplitude, and temperature drift compensation is performed in combination with the measured ambient temperature. The PG-ANet fusion inference module is used to input the temperature drift-compensated physical feature vector and the effective signal into a pre-trained physical perception attention neural network for fusion inference, and output the concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value. The closed-loop control module is used to evaluate the confidence score of the concentration prediction value in step S4; if the confidence score is lower than the preset threshold, the Ascend AI processor dynamically adjusts the parameters of the direct digital frequency synthesis module and returns to step S1 to re-execute; otherwise, it outputs the final concentration prediction value of the biomarker and the aortic dissection risk level corresponding to the concentration prediction value.

Images