Blood brain barrier opening hemorrhage prediction method and system based on time domain signal and medium

Abstract

The application provides a blood-brain barrier opening hemorrhage prediction method and system based on a time domain signal, a medium, all pulse acoustic signals of a single treatment are divided into multiple short time domain acoustic signal segments, a time-frequency relationship coding module is used to extract features of the acoustic signals in the frequency domain and the time domain, the time resolution and the dynamic change capturing capability are improved, the information loss in the frequency domain conversion is reduced, a magnetic resonance contrast agent does not need to be injected, the living body detection capability is maintained, and non-invasive detection is realized. Further, through a gating aggregation module and a multi-path classification optimization design of a gating attention mechanism, the distinguishing capability of the model for different state characteristics is enhanced, the classification precision and the prediction performance are improved.

Description

Technical Field

[0001] This application relates to the field of therapeutic ultrasound technology, and in particular to a method, system, and medium for predicting blood-brain barrier opening and bleeding based on time-domain signals. Background Technology

[0002] The blood-brain barrier (BBB) ​​is a physiological barrier structure composed of brain capillary endothelial cells, the basement membrane, and astrocyte terminales. Its core function is to maintain the homeostasis of the central nervous system's internal environment and prevent harmful substances from entering brain tissue. However, while preventing harmful substances from entering the brain, this barrier structure also severely hinders the cross-barrier delivery of therapeutic drugs (such as chemotherapy drugs and neuroprotective agents), resulting in central nervous system therapeutic drugs failing to reach effective therapeutic concentrations.

[0003] In recent years, the technique of focusing ultrasound combined with microbubble ultrasound contrast agents has been able to reversibly and specifically increase the permeability of the blood-brain barrier through acoustic cavitation, thereby significantly improving drug delivery efficiency. Current methods for evaluating this modulatory effect include clinically used contrast-enhanced magnetic resonance imaging (MRI) and preclinical studies using dye tracing and staining methods.

[0004] (1) Magnetic Resonance Imaging: The current gold standard for assessing the effectiveness of blood-brain barrier opening is enhanced T1-weighted magnetic resonance imaging (MRI). When gadolinium, an intravenously injected contrast agent, circulates in the blood vessels, if ultrasound-modulated target areas show increased blood-brain permeability, gadolinium can penetrate the blood vessels and enter the tissue. T1-weighted MRI can then show the high-brightness distribution of gadolinium in the tissue, thus assessing the effect of permeability modulation. The clinical method for assessing the safety of blood-brain barrier opening is T2-weighted MRI, where cerebral hemorrhage appears as a high-brightness area in the image. The disadvantages of this method are that it requires expensive MRI equipment, has a long scan time, requires the injection of gadolinium (a toxic and difficult-to-excrete contrast agent) for effectiveness assessment, and is not suitable for patients with implanted pacemakers or other metallic implants, or those with claustrophobia.

[0005] (2) Dye Tracing Method: First, a dye (such as Evans Blue) is injected into the animal's vein. When the dye passes through the target area regulated by focused ultrasound, if the permeability of the blood-brain barrier in that area increases, the dye molecules will permeate out of the blood vessel and enter the tissue. Then, the animal is euthanized, the target tissue is removed, and the dye distribution is detected using a spectrophotometer or fluorescence microscope to assess the regulatory effect. The disadvantages of this method are that it is cumbersome, time-consuming, and lacks in vivo assessment capabilities, making it difficult to meet the needs of real-time clinical assessment.

[0006] (3) Staining method: After staining the brain sections of experimental animals with hematoxylin-eosin dye, the red blood cells caused by hemorrhage will be stained orange-red. Microscopic examination of the staining results can be used to assess the safety of treatment (whether there is bleeding). This method also has the limitations of being cumbersome, time-consuming and lacking the ability to assess in vivo. Summary of the Invention

[0007] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a method, system and medium for predicting blood-brain barrier opening and bleeding based on time-domain signals, in order to solve the technical problems of existing blood-brain barrier opening assessment techniques, which require the injection of magnetic resonance contrast agents with certain toxicity and difficult to excrete from the body, are time-consuming, and lack the ability to assess in vivo.

[0008] To achieve the above and other related objectives, a first aspect of this application provides a method for predicting open hemorrhage of the blood-brain barrier based on time-domain signals, comprising: acquiring multiple historical acoustic signal samples during the modulation of the blood-brain barrier by ultrasound combined with contrast agents, and preprocessing each of the historical acoustic signal samples to obtain multiple acoustic signal training samples; wherein each acoustic signal training sample includes multiple short-time-domain acoustic signal segments; pairing each of the acoustic signal training samples with multiple acquired blood-brain barrier open hemorrhage assessment samples to construct a training dataset; and training deep learning using the training dataset. The model generates a blood-brain barrier open hemorrhage prediction result corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample; based on the blood-brain barrier open hemorrhage prediction result corresponding to the acoustic signal training sample and the blood-brain barrier open hemorrhage assessment sample, and based on a preset loss function, the parameters of the deep learning model are adjusted to construct a blood-brain barrier open hemorrhage prediction model; based on the blood-brain barrier open hemorrhage prediction model, and based on the real-time acoustic signal samples obtained from the ultrasound combined with contrast agent in the process of regulating the blood-brain barrier, the blood-brain barrier open hemorrhage result data is obtained.

[0009] In some embodiments of the first aspect of this application, each of the historical acoustic signal samples includes multiple pulsed time-domain acoustic signals; wherein, the method of performing data preprocessing on each of the historical acoustic signal samples to obtain multiple acoustic signal training samples includes: performing time-series splicing preprocessing on each pulsed time-domain acoustic signal in the historical acoustic signal samples to generate a continuous long time-domain acoustic signal; performing segmented trimming preprocessing on the continuous long time-domain acoustic signal to generate multiple short time-domain acoustic signal segments, and constructing acoustic signal training samples.

[0010] In some embodiments of the first aspect of this application, the deep learning model includes a time-frequency relationship encoding module, a gating aggregation module, and a state classification module; wherein, the method for generating the blood-brain barrier open hemorrhage prediction result corresponding to the acoustic signal training sample includes: using the time-frequency relationship encoding module to perform time-frequency domain feature fusion and adjacent segment relationship modeling operations on each short time-domain acoustic signal segment in the acoustic signal training sample to obtain the acoustic coding feature vector corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample; using the gating aggregation module to perform a dynamic weighted summation operation on the acoustic coding feature vector corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample to generate gating aggregation feature vectors corresponding to different categories of blood-brain barrier open hemorrhage states; using the state classification module to perform a mapping operation on the gating aggregation feature vectors corresponding to different categories of blood-brain barrier open hemorrhage states to obtain prediction scores corresponding to different categories of blood-brain barrier open hemorrhage states, and selecting the category of blood-brain barrier open hemorrhage state corresponding to the highest prediction score as the blood-brain barrier open hemorrhage prediction result corresponding to the acoustic signal training sample.

[0011] In some embodiments of the first aspect of this application, the time-frequency relationship coding module includes a convolutional encoder, a spectrum encoder, and a pyramid-shaped adjacency coding module; wherein, the method of using the time-frequency relationship coding module to perform time-frequency domain feature fusion and adjacency segment relationship modeling operations on each short-time domain acoustic signal segment in the acoustic signal training sample to obtain the acoustic coding feature vector corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample includes: using the convolutional encoder to perform time-domain feature extraction operations on each short-time domain acoustic signal segment in the acoustic signal training sample to obtain the time-domain feature vector corresponding to each short-time domain acoustic signal segment; using the spectrum encoder to perform time-frequency domain feature extraction operations on each short-time domain acoustic signal segment in the acoustic signal training sample to obtain the time-domain feature vector corresponding to each short-time domain acoustic signal segment; and using the spectrum encoder to perform time-frequency domain feature extraction operations on each short-time domain acoustic signal segment in the acoustic signal training sample to obtain the time-domain feature vector corresponding to each short-time domain acoustic signal segment. Frequency domain feature extraction is performed on each short-time acoustic signal segment to obtain the frequency domain feature vector corresponding to each short-time acoustic signal segment. An element-wise addition algorithm is used to perform feature fusion on the time-domain feature vector and the frequency-domain feature vector corresponding to each short-time acoustic signal segment to obtain the fused feature vector corresponding to each short-time acoustic signal segment. After layer normalization, the fused feature vector corresponding to each short-time acoustic signal segment is modeled using the pyramid-shaped adjacent coding module to obtain the acoustic coding feature vector corresponding to each short-time acoustic signal segment in the acoustic signal training sample.

[0012] In some embodiments of the first aspect of this application, the gating aggregation module includes multiple gating sub-modules with identical structures; wherein, the method of using the gating aggregation module to perform dynamic weighted summation on the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample to generate gating aggregated feature vectors corresponding to different categories of blood-brain barrier open bleeding states includes: inputting the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample into each of the gating sub-modules in parallel; each of the gating sub-modules, based on a gating attention mechanism, performs dynamic weighted summation on the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample to output the gating aggregated feature vectors corresponding to the corresponding category of blood-brain barrier open bleeding states.

[0013] In some embodiments of the first aspect of this application, the state classification module includes multiple fourth fully connected layers with different parameters; wherein, the method of using the state classification module to map the gated aggregated feature vectors corresponding to different categories of open blood-brain barrier bleeding states to obtain prediction scores corresponding to different categories of open blood-brain barrier bleeding states includes: inputting the gated aggregated feature vectors corresponding to different categories of open blood-brain barrier bleeding states into the corresponding fourth fully connected layers respectively; each fourth fully connected layer performs category mapping on the gated aggregated feature vectors corresponding to the input categories of open blood-brain barrier bleeding states to generate prediction scores corresponding to different categories of open blood-brain barrier bleeding states.

[0014] In some embodiments of the first aspect of this application, the convolutional encoder includes a first normalization layer, a convolutional module, a first fully connected layer, and a first activation function layer; the spectral encoder includes an exponential transform layer, a second normalization layer, a second fully connected layer, and a second activation function layer; the pyramid-shaped adjacent coding module includes multiple one-dimensional convolutions.

[0015] In some embodiments of the first aspect of this application, the convolutional module includes a convolutional layer, a third activation function layer, and a max pooling layer.

[0016] To achieve the above and other related objectives, a second aspect of this application provides a blood-brain barrier open hemorrhage prediction system based on time-domain signals, comprising: a data preprocessing module for acquiring multiple historical acoustic signal samples during the modulation of the blood-brain barrier by ultrasound combined with contrast agents, and performing data preprocessing on each of the historical acoustic signal samples to obtain multiple acoustic signal training samples; wherein each acoustic signal training sample includes multiple short-time-domain acoustic signal segments; a data pairing module for pairing each of the acoustic signal training samples with the acquired multiple blood-brain barrier open hemorrhage assessment samples to construct a training dataset; and a model training module for using the training dataset. A deep learning model is trained, and based on each short-time-domain acoustic signal segment in the acoustic signal training samples, a prediction result for blood-brain barrier open hemorrhage corresponding to the acoustic signal training samples is generated. A model building module is used to adjust the parameters of the deep learning model based on the prediction result for blood-brain barrier open hemorrhage corresponding to the acoustic signal training samples and the blood-brain barrier open hemorrhage assessment samples, and based on a preset loss function, to construct a prediction model for blood-brain barrier open hemorrhage. A prediction module is used to obtain blood-brain barrier open hemorrhage result data based on the prediction model for blood-brain barrier open hemorrhage and based on the real-time acoustic signal samples obtained during the process of ultrasound combined with contrast agent regulating the blood-brain barrier.

[0017] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the blood-brain barrier opening hemorrhage prediction method based on time-domain signals as described above.

[0018] As described above, the method, system, and medium for predicting blood-brain barrier opening hemorrhage based on time-domain signals of this application have the following beneficial effects:

[0019] (1) No need to inject magnetic resonance contrast agent: Compared with the current clinical gold standard - enhanced T1-weighted magnetic resonance imaging and T2-weighted magnetic resonance imaging, no need to inject magnetic resonance contrast agent, avoiding a long scanning process, and realizing real-time and safe monitoring of blood-brain barrier opening and bleeding status.

[0020] (2) Possesses in vivo detection capability: Compared with dye tracing and staining methods, it does not require dye injection or tissue staining, avoiding the impact of additional operations on the experimental subjects, maintaining in vivo detection capability, and realizing non-invasive detection.

[0021] (3) Improved temporal resolution: All pulse acoustic signals from a single treatment are segmented into multiple short-time domain acoustic signal segments and input into the model, which improves the temporal resolution and enables the model to capture the acoustic dynamic changes during the opening of the blood-brain barrier. This not only locates the key time points of the changes in the state of the blood-brain barrier, but also provides rich temporal information.

[0022] (4) Avoid information loss: Using a time-frequency relationship coding module to extract features of acoustic signals in both the frequency and time domains helps to reduce information loss caused by Fourier transform.

[0023] (5) Adaptive feature aggregation: The gated attention mechanism is used to achieve information aggregation, which can adaptively focus on the time segment most important to the classification and improve the model's efficiency in utilizing important features.

[0024] (6) Multi-path classification optimization: The multi-path design of the gating aggregation module and the state classification module, combined with the corresponding activation function, enhances the model's ability to distinguish different state features, improves classification accuracy and prediction performance. Attached Figure Description

[0025] Figure 1 The diagram shown is a flowchart illustrating a method for predicting blood-brain barrier opening and hemorrhage based on time-domain signals in one embodiment of this application.

[0026] Figure 2 The diagram shown is a structural schematic of a deep learning model in one embodiment of this application.

[0027] Figure 3 The diagram shown is a flowchart illustrating the training process of a deep learning model in one embodiment of this application.

[0028] Figure 4 The diagram shown is a structural schematic of a convolutional encoder according to an embodiment of this application.

[0029] Figure 5 The diagram shown is a structural schematic of a convolution module in one embodiment of this application.

[0030] Figure 6 The diagram shown is a schematic representation of the structure of a spectrum encoder in one embodiment of this application.

[0031] Figure 7 The diagram shown is a schematic representation of the structure of a pyramid-shaped adjacent coding module in one embodiment of this application.

[0032] Figure 8 The diagram shown is a schematic block diagram of a blood-brain barrier opening hemorrhage prediction system based on time-domain signals in one embodiment of this application. Detailed Implementation

[0033] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0034] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0035] <1> T1: Represents the time constant for the longitudinal magnetization of an organism to return to equilibrium in a magnetic field, and describes the process of energy being released into the surrounding lattice.

[0036] <2> T2: The time constant reflecting the transverse magnetization decay of proton spins, characterizing the loss of phase coherence between spins, and sensitive to free water.

[0037] <3> Deep learning models are machine learning models based on artificial neural networks (ANNs). Their core idea is to learn the feature representations and patterns of data through multi-layered neural network structures.

[0038] <4> Transformer model: A model based on self-attention mechanism that can process all elements in the input sequence in parallel.

[0039] <5> Adam optimizer (Adaptive Moment Estimation) is a gradient descent optimization algorithm that combines momentum and adaptive learning rate. It is widely used for training deep learning models. It dynamically adjusts the learning rate of each parameter by calculating the first moment (mean) and second moment (uncentered variance) of the gradient, thereby achieving efficient convergence during training.

[0040] In recent years, the technique of focusing ultrasound combined with microbubble ultrasound contrast agents has been able to reversibly and specifically increase the permeability of the blood-brain barrier through acoustic cavitation, thereby significantly improving drug delivery efficiency. Current methods for evaluating this modulatory effect include clinically used contrast-enhanced magnetic resonance imaging (MRI) and preclinical studies using dye tracing and staining methods.

[0041] (1) Magnetic resonance imaging: It requires expensive magnetic resonance equipment, has a long scanning time, and requires the injection of gadolinium, a magnetic resonance contrast agent that is toxic and difficult to excrete, during the effectiveness assessment. It is also not suitable for patients with metal implants such as pacemakers or those with claustrophobia.

[0042] (2) Dye tracing and staining methods: The process is cumbersome, time-consuming and lacks the ability to assess in vivo, making it difficult to meet the needs of real-time clinical assessment.

[0043] To address the problems mentioned above, this application provides a method, system, and medium for predicting blood-brain barrier opening and bleeding based on time-domain signals. This addresses the technical problems of existing blood-brain barrier opening assessment techniques, which require the injection of magnetic resonance contrast agents with certain toxicity that are difficult to excrete, are time-consuming, and lack in vivo assessment capabilities.

[0044] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 This document illustrates a flowchart of a blood-brain barrier opening hemorrhage prediction method based on time-domain signals, as described in an embodiment of the present invention. The method mainly includes the following steps:

[0045] S101: Acquire multiple historical acoustic signal samples during the process of ultrasound combined with contrast agent regulating the blood-brain barrier, and perform data preprocessing on each of the historical acoustic signal samples to obtain multiple acoustic signal training samples; wherein, each of the acoustic signal training samples includes multiple short-time domain acoustic signal segments.

[0046] In this embodiment, a passive cavitation detection probe is used to collect multiple historical acoustic signal samples during the modulation of the blood-brain barrier by ultrasound combined with contrast agents. Under focused ultrasound, the ultrasound contrast agent generates a cavitation effect, which produces mechanical forces acting on the blood-brain barrier. This mechanical action, combined with mechanisms such as possible cell ion channel switching, can achieve a temporary enhancement of blood-brain barrier permeability. If the ultrasound pressure is too high, the cavitation activity of the ultrasound contrast agent may also lead to hemorrhage. During this process, the ultrasound contrast agent emits acoustic signals, and these signals are closely related to the oscillation state of the ultrasound contrast agent, which determines the modulation effect on blood-brain barrier permeability. Therefore, a passive cavitation detection probe is used to collect these acoustic signals for subsequent processing.

[0047] In this embodiment, each of the historical acoustic signal samples includes multiple pulsed time-domain acoustic signals; wherein, the method of performing data preprocessing on each of the historical acoustic signal samples to obtain multiple acoustic signal training samples includes:

[0048] (1) Perform time-series splicing preprocessing on each pulsed time-domain acoustic signal in the historical acoustic signal sample to generate a continuous long time-domain acoustic signal.

[0049] (2) The continuous long time-domain acoustic signal is segmented and preprocessed to generate multiple short time-domain acoustic signal segments and construct acoustic signal training samples.

[0050] In this embodiment, each historical acoustic signal sample consists of multiple pulsed time-domain acoustic signals synchronously collected during a single treatment session of focused ultrasound pulses. Each pulsed time-domain acoustic signal is a sound signal represented in the time domain and has pulse characteristics, meaning its duration is relatively short and concentrated, consistent with the duration of the applied ultrasound pulse, which is several milliseconds. The acoustic signals collected by the passive cavitation detection probe are sampled using a digital converter for storage and processing.

[0051] In this embodiment, the timing splicing preprocessing involves concatenating the pulsed time-domain acoustic signals from each historical acoustic signal sample in chronological order of pulse occurrence, forming a single, continuously extended acoustic signal, i.e., a continuous long-time-domain acoustic signal. This continuous long-time-domain acoustic signal retains the timing relationships and waveform information of all the original pulses.

[0052] In this embodiment, the segmented clipping preprocessing uses a fixed-length window to clip the continuous long time-domain acoustic signal, dividing it into multiple non-overlapping short time-domain acoustic signal segments. These short time-domain acoustic signal segments constitute the acoustic signal training samples used to train the deep learning model.

[0053] It is worth noting that existing artificial intelligence methods for predicting the effectiveness of focused ultrasound in assessing blood-brain barrier opening typically perform Fourier transform preprocessing on the acquired acoustic signals, fusing the acoustic signals over the entire pulse duration into a single spectral feature as model input. However, the temporal resolution is limited by the pulse duration, making it impossible to analyze the dynamics within the pulse and the instantaneous interactions between pulses. This results in the model only being able to learn the superficial statistical regularities of static spectral features, lacking sufficient sensitivity for bleeding prediction, and failing to meet the medical needs for safety monitoring.

[0054] This application uses short-time-domain acoustic signal segments as input to the model, which can identify transient features such as the steepness of the rising / falling edge of the pulse and small fluctuations within the duration, effectively improving the model's temporal resolution. This enables the model to capture acoustic dynamic changes during the opening of the blood-brain barrier, thereby identifying abnormal acoustic patterns that may lead to hemorrhage, and thus improving the model's performance in hemorrhage prediction, providing a safety warning for clinical operations.

[0055] S102: Pair each of the acoustic signal training samples with the acquired multiple blood-brain barrier open hemorrhage assessment samples to construct a training dataset.

[0056] In this embodiment, each acoustic signal training sample is paired with a corresponding blood-brain barrier open hemorrhage assessment sample to construct a training dataset. The training dataset includes multiple acoustic signal training samples and blood-brain barrier open hemorrhage assessment samples corresponding to each of the acoustic signal training samples.

[0057] S103: Train a deep learning model using the training dataset, and generate a blood-brain barrier opening and hemorrhage prediction result corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample.

[0058] In this embodiment, as Figure 2 The diagram shown illustrates the structure of a deep learning model in an embodiment of the present invention. Figure 3 The diagram illustrates the flowchart for training a deep learning model in an embodiment of the present invention. The deep learning model includes a time-frequency relationship encoding module, a gating aggregation module, and a state classification module; wherein, the method for generating the blood-brain barrier open hemorrhage prediction result corresponding to the acoustic signal training samples includes:

[0059] S1031: The time-frequency relationship coding module is used to perform time-frequency domain feature fusion and adjacent segment relationship modeling operations on each short time-domain acoustic signal segment in the acoustic signal training sample to obtain the acoustic coding feature vector corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample.

[0060] In this embodiment, the time-frequency relationship coding module includes a convolutional encoder, a spectrum encoder, and a pyramid-shaped adjacency coding module; wherein, the method of using the time-frequency relationship coding module to perform time-frequency domain feature fusion and adjacency segment relationship modeling operations on each short-time domain acoustic signal segment in the acoustic signal training sample to obtain the acoustic coding feature vector corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample includes:

[0061] (1) The convolutional encoder is used to perform temporal feature extraction on each short temporal acoustic signal segment in the acoustic signal training sample to obtain the temporal feature vector corresponding to each short temporal acoustic signal segment.

[0062] (2) The frequency domain feature extraction operation is performed on each short time domain acoustic signal segment in the acoustic signal training sample using the spectrum encoder to obtain the frequency domain feature vector corresponding to each short time domain acoustic signal segment.

[0063] (3) The time-domain feature vector and frequency-domain feature vector corresponding to each short-time-domain acoustic signal segment are fused using the element-by-element addition algorithm to obtain the fused feature vector corresponding to each short-time-domain acoustic signal segment.

[0064] (4) After the fusion feature vectors corresponding to each short-time domain acoustic signal segment are processed by layer normalization, the pyramid-type adjacent coding module is used to model the fusion feature vectors corresponding to adjacent short-time domain acoustic signal segments after layer normalization, so as to obtain the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample.

[0065] In this embodiment, as Figure 4 The diagram shows a schematic representation of the convolutional encoder in an embodiment of the present invention. Figure 5 The diagram shown illustrates the structure of the convolution module in an embodiment of the present invention. Figure 6 The diagram shows a schematic representation of the spectrum encoder in an embodiment of the present invention. Figure 7 The diagram illustrates the structure of the pyramid-shaped adjacent coding module in an embodiment of the present invention. The convolutional encoder includes a first normalization layer, a convolutional module, a first fully connected layer, and a first activation function layer; the spectral encoder includes an exponential transform layer, a second normalization layer, a second fully connected layer, and a second activation function layer; the pyramid-shaped adjacent coding module includes multiple one-dimensional convolutions. The convolutional module includes a convolutional layer, a third activation function layer, and a max-pooling layer.

[0066] In this embodiment, as Figure 4 As shown, the convolutional encoder has two first normalization layers and two convolutional modules. The first normalization layer normalizes the feature dimensions of the input short temporal acoustic signal segments, eliminating distribution differences caused by energy fluctuations between different segments, stabilizing backpropagation gradients, and accelerating convergence. The convolutional modules are used for hierarchical extraction of temporal features. Another first normalization layer performs secondary normalization to correct data distribution shifts caused by convolution operations. The first fully connected layer learns cross-channel temporal correlations, further extracts temporal features, and achieves feature dimensionality reduction. The first activation function layer introduces nonlinearity, which helps in modeling complex patterns. Thus, the convolutional encoder, through the first normalization layer, convolutional modules, first fully connected layer, and first activation function layer, transforms short temporal acoustic signal segments into high-information-density temporal feature vectors.

[0067] In this embodiment, as Figure 5As shown, the convolutional layer in the convolution module is a one-dimensional convolutional layer, used to extract local temporal features of short-time acoustic signal segments. The third activation function layer is used to inject nonlinear transformation capability into the linear convolution operation, enabling the model to fit the nonlinear relationship of the interaction between sound wave and ultrasound contrast agent. The max pooling layer is used to preserve pulse peak features, suppress random noise, and achieve feature dimensionality reduction.

[0068] In this embodiment, each short-time-domain acoustic signal segment in the acoustic signal training samples undergoes Fourier transform processing before being fed into a spectrum encoder, enabling the spectrum encoder to extract frequency domain features from the short-time-domain acoustic signal segments. For example... Figure 6 As shown, the exponential transform layer in the spectrum encoder is used to nonlinearly enhance the frequency domain amplitude, extending high-amplitude signals to a wider range, which is beneficial for detecting increases in frequency components. The second normalization layer is used to standardize the distribution of different frequency domain features, stabilizing the backpropagation gradient and accelerating convergence. The second fully connected layer is used to extract spectral features. The second activation function layer is used to introduce nonlinear responses of spectral features to obtain the frequency domain feature vectors corresponding to each short-time-domain acoustic signal segment.

[0069] In this embodiment, each short-time-domain acoustic signal segment in the acoustic signal training sample is processed by a convolutional encoder and a spectral encoder with shared weights to extract the time-domain feature vector and frequency-domain feature vector corresponding to the short-time-domain acoustic signal segment. Then, the time-domain features and frequency-domain features are fused by an element-wise addition algorithm. After layer normalization, the relationship between adjacent short-time-domain acoustic signal segments is modeled by a pyramid-type adjacent coding module to obtain the acoustic coding feature vector corresponding to each short-time-domain acoustic signal segment in the acoustic signal training sample.

[0070] In this embodiment, as Figure 7 As shown, the one-dimensional convolutions in the pyramid-shaped adjacency coding module have different sizes, which are used to extract features and model cross-scale relationships of short time-domain acoustic signal segments after layer normalization. The pyramid-shaped adjacency coding module uses three different sizes of one-dimensional convolutions to form a pyramid-shaped hierarchical structure, focusing on acoustic features at different time scales. Through this multi-scale parallel processing, features at different time resolutions are concatenated through skip connections to form an acoustic coding feature vector containing complete time-frequency information. This preserves the transient details of the signal, integrates global contextual information, and effectively captures the dynamic interactions between adjacent signal segments, improving the model's ability to understand acoustic scenes.

[0071] S1032: The gating aggregation module is used to perform dynamic weighted summation on the acoustic coding feature vectors corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample to generate gating aggregation feature vectors corresponding to different categories of blood-brain barrier open hemorrhage states.

[0072] In this embodiment, the gating aggregation module includes multiple gating sub-modules with identical structures; wherein, the method of using the gating aggregation module to perform dynamic weighted summation on the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training samples to generate gating aggregation feature vectors corresponding to different categories of blood-brain barrier open hemorrhage states includes:

[0073] (1) Input the acoustic coding feature vectors corresponding to each short time domain acoustic signal segment in the acoustic signal training sample into each of the gated sub-modules in parallel.

[0074] (2) Each of the gated submodules performs a dynamic weighted summation operation on the acoustic coding feature vectors corresponding to each short time domain acoustic signal segment in the acoustic signal training sample based on the gated attention mechanism, so as to output the gated aggregated feature vector corresponding to the blood-brain barrier open bleeding state of the corresponding category.

[0075] In this embodiment, for example, there are three gating sub-modules, corresponding to the gating aggregated feature vectors of the blood-brain barrier opening and bleeding states in three categories: "blood-brain barrier not open," "blood-brain barrier successfully open and no bleeding," and "blood-brain barrier open and bleeding." Each gating sub-module has the same structure, including a third fully connected layer and a corresponding activation function. Figure 2 In this context, Tanh represents the hyperbolic tangent function, which performs a non-linear mapping on the features, compressing the feature values ​​to a symmetric interval to generate eigenvectors representing the correlation between the fragment and the target state. Sigmoid represents the probability function, which calculates an initial weight in the interval [0, 1] for each fragment, representing the fragment's contribution to the current target state (e.g., "open bleeding"). Softmax represents the global weight normalization function, which transforms the initial weights of all fragments into a probability distribution.

[0076] In this embodiment, each gating submodule simultaneously receives the acoustic coding feature vectors corresponding to each short-time-domain acoustic signal segment. A gating attention mechanism is then used to dynamically weight the features, outputting a corresponding gating aggregated feature vector. This vector represents the global, optimized feature representation of the entire acoustic signal training sample within its assigned blood-brain barrier open / bleeding state category. The three gating aggregated feature vectors output by the gating aggregation module represent the aggregated features for the "blood-brain barrier not open" state, the "blood-brain barrier successfully open and no bleeding" state, and the "blood-brain barrier open and bleeding" state, respectively.

[0077] In this embodiment, the gated attention mechanism is a deep learning model design that combines gating units and attention mechanisms. It aims to enhance the model's ability to focus on key features by dynamically controlling the information flow. The core idea of ​​the gated attention mechanism is to introduce learnable gating signals to dynamically adjust the distribution of attention weights, determining which parts need to be strengthened or suppressed. The gating mechanism acts as a "filter," preventing irrelevant information from interfering and improving the model's robustness.

[0078] S1033: The state classification module is used to map the gated aggregated feature vectors corresponding to different categories of open blood-brain barrier bleeding states to obtain the prediction scores corresponding to different categories of open blood-brain barrier bleeding states. The category of open blood-brain barrier bleeding state corresponding to the highest prediction score is selected as the prediction result of open blood-brain barrier bleeding corresponding to the acoustic signal training sample.

[0079] In this embodiment, the state classification module includes multiple fourth fully connected layers with different parameters; wherein, the method of using the state classification module to map the gated aggregated feature vectors corresponding to different categories of blood-brain barrier open hemorrhage states to obtain the prediction scores corresponding to different categories of blood-brain barrier open hemorrhage states includes:

[0080] (1) Input the gating aggregated feature vectors corresponding to different categories of blood-brain barrier open bleeding states into the corresponding fourth fully connected layer.

[0081] (2) Each of the fourth fully connected layers performs class mapping on the gated aggregated feature vectors corresponding to the blood-brain barrier open bleeding state of the input category, so as to generate prediction scores corresponding to different categories of blood-brain barrier open bleeding state.

[0082] In this embodiment, for example, the state classification module includes three fourth fully connected layers with different parameters. Each fourth fully connected layer optimizes its parameters for its corresponding state category to parse the acoustic feature patterns of different categories of open blood-brain barrier bleeding states. The three gated aggregation feature vectors generated by the gated aggregation module are input into the corresponding fourth fully connected layers, and each fourth fully connected layer independently maps features to classification scores. By comparing the predicted scores corresponding to different categories of open blood-brain barrier bleeding states, the state category with the highest score is taken as the final prediction result of the model. This ensures that each state category has its own dedicated feature transformation path, maintaining the independence of the discriminative features for each category while automatically learning the optimal decision boundary through end-to-end training.

[0083] It is worth noting that existing artificial intelligence methods for predicting the effectiveness of focused ultrasound in assessing the opening of the blood-brain barrier often employ a model structure that includes an acoustic coding module, a feature aggregation module, and a classification module. The acoustic coding module is used to encode the acoustic signal data, the feature aggregation module is used to perform feature fusion processing on the encoded acoustic signal, and the classification module is used to output the result data of the opening of the blood-brain barrier based on the acoustic signal after feature fusion. The acoustic signal throughout the entire pulse duration is fused into a single spectral feature as the model input. The temporal resolution is limited by the pulse duration, making it difficult to capture the subtle features in the dynamic changes of the blood-brain barrier.

[0084] This application segments the acoustic signals collected during all pulses of a single treatment session into several short-time-domain acoustic signal segments as input to the model. This improves temporal resolution, enabling the localization of key time points in the blood-brain barrier state changes. Furthermore, it increases the length of the input sequence, providing the model with richer temporal information. The use of a time-frequency relationship encoding module to extract features from the input data in both the frequency and time domains helps reduce information loss caused by Fourier transform.

[0085] Furthermore, compared to the feature aggregation modules in existing technologies, this application employs a gated aggregation module based on a gated attention mechanism to achieve information aggregation. This attention mechanism adaptively focuses on the time segments most important for classification, improving feature utilization efficiency. Finally, both the gated aggregation module and the state classification module used in this application are multi-path designs, enabling the model to simultaneously learn specific feature representations of different states. When combined with the designed activation function, this helps enhance the model's classification performance.

[0086] S104: Based on the blood-brain barrier open hemorrhage prediction results and blood-brain barrier open hemorrhage assessment samples corresponding to the acoustic signal training samples, and based on a preset loss function, adjust the parameters of the deep learning model to construct a blood-brain barrier open hemorrhage prediction model.

[0087] In this embodiment, the preset loss function includes:

[0088]

[0089] in, Indicates the preset loss function; Represents the cross-entropy loss function; α represents the similarity loss function; α represents the weight coefficient.

[0090] In this embodiment, cross-validation is used to test the deep learning model. The model uses the Adam optimizer with a learning rate of 0.0003. The preset loss function is a hybrid loss function of cross-entropy loss and similarity loss. The training epochs do not exceed 150, and an early stopping strategy is applied to save the optimal model parameters. The weight coefficient α is used to adjust the model's sensitivity to classification loss and similarity loss. The larger α is, the more it encourages different gating aggregation modules in the model to distinguish different feature vectors. An α of 0.1 can be used for model training.

[0091] In this embodiment, cross-validation is an effective method for evaluating the performance of machine learning models. It involves dividing the training dataset into multiple subsets, such as a training subset and a test subset, and repeatedly training and testing the model. In each iteration of cross-validation, the deep learning model is trained using the training subset. The test subset is used to evaluate model performance, constructing a confusion matrix and calculating evaluation metrics such as accuracy, recall, precision, and F1 score to more stably and reliably assess the model's generalization ability. The confusion matrix is ​​a commonly used tool for evaluating the performance of classification models; it calculates a series of performance metrics by statistically analyzing the differences between the model's predictions and the actual results.

[0092] In this embodiment, the model's accuracy refers to the proportion of correctly predicted samples out of the total number of samples, and its calculation formula is as follows:

[0093]

[0094] Where Accuracy represents the model's accuracy; TP represents the number of samples that the model predicted as positive but were actually positive; TN represents the number of samples that the model predicted as negative but were actually negative; FP represents the number of samples that the model predicted as positive but were actually negative; and FN represents the number of samples that the model predicted as negative but were actually positive.

[0095] In this embodiment, the model's recall rate refers to the proportion of samples correctly predicted as positive out of the actual positive samples, and its calculation formula is as follows:

[0096]

[0097] Where Recall represents the model's recall rate; TP represents the number of samples that the model predicted as positive but were actually positive; and FN represents the number of samples that the model predicted as negative but were actually positive.

[0098] In this embodiment, the model's precision refers to the proportion of samples that the model predicts to be positive but are actually positive. The calculation formula is as follows:

[0099]

[0100] Where Precision represents the model's accuracy; TP represents the number of samples that the model predicted as positive but were actually positive; and FP represents the number of samples that the model predicted as positive but were actually negative.

[0101] In this embodiment, the F1 score of the model refers to the harmonic mean of precision and recall, and its calculation formula is as follows:

[0102]

[0103] Where Precision represents the model's precision, and Recall represents the model's recall.

[0104] S105: Based on the blood-brain barrier opening hemorrhage prediction model, and according to the real-time acoustic signal samples obtained by ultrasound combined with contrast agent in the process of regulating the blood-brain barrier, the blood-brain barrier opening hemorrhage result data are obtained.

[0105] In this embodiment, real-time acoustic signal samples from the actual treatment process are input into the blood-brain barrier opening and bleeding prediction model to quickly assess the effect of ultrasound-mediated blood-brain barrier permeability control. The assessed control effect refers to whether the blood-brain barrier is opened and whether there is bleeding. Category 1 represents "blood-brain barrier not open", Category 2 represents "blood-brain barrier successfully opened and without bleeding", and Category 3 represents "blood-brain barrier open and bleeding". This model can assess not only the opening effect of the blood-brain barrier but also whether there is bleeding.

[0106] In summary, the blood-brain barrier opening hemorrhage prediction method based on time-domain signals proposed in this application has the following advantages:

[0107] (1) No need to inject magnetic resonance contrast agent: Compared with the current clinical gold standard - enhanced T1-weighted magnetic resonance imaging and T2-weighted magnetic resonance imaging, no need to inject magnetic resonance contrast agent, avoiding a long scanning process, and realizing real-time and safe monitoring of blood-brain barrier opening and bleeding status.

[0108] (2) Possesses in vivo detection capability: Compared with dye tracing and staining methods, it does not require dye injection or tissue staining, avoiding the impact of additional operations on the experimental subjects, maintaining in vivo detection capability, and realizing non-invasive detection.

[0109] (3) Improved temporal resolution: All pulse acoustic signals from a single treatment are segmented into multiple short-time domain acoustic signal segments and input into the model, which improves the temporal resolution and enables the model to capture the acoustic dynamic changes during the opening of the blood-brain barrier. This not only locates the key time points of the changes in the state of the blood-brain barrier, but also provides rich temporal information.

[0110] (4) Avoid information loss: Using a time-frequency relationship coding module to extract features of acoustic signals in both the frequency and time domains helps to reduce information loss caused by Fourier transform.

[0111] (5) Adaptive feature aggregation: The gated attention mechanism is used to achieve information aggregation, which can adaptively focus on the time segment most important to the classification and improve the model's efficiency in utilizing important features.

[0112] (6) Multi-path classification optimization: The multi-path design of the gating aggregation module and the state classification module, combined with the corresponding activation function, enhances the model's ability to distinguish different state features, improves classification accuracy and prediction performance.

[0113] In the embodiments of this application, terms such as "first" and "second" are used to distinguish identical or similar items with essentially the same function and purpose. For example, "first fully connected layer" and "second fully connected layer" are used only to distinguish different fully connected layers and do not limit their order. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different.

[0114] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0115] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0116] Figure 8 This is a schematic block diagram of a blood-brain barrier opening hemorrhage prediction system based on time-domain signals provided in an embodiment of this application.

[0117] like Figure 8As shown, the blood-brain barrier opening hemorrhage prediction system 800 based on time-domain signals includes:

[0118] The data preprocessing module 801 is used to acquire multiple historical acoustic signal samples during the process of ultrasound combined with contrast agent regulating the blood-brain barrier, and to perform data preprocessing on each of the historical acoustic signal samples to obtain multiple acoustic signal training samples; wherein, each of the acoustic signal training samples includes multiple short time-domain acoustic signal segments.

[0119] The data pairing module 802 is used to pair each of the acoustic signal training samples with the acquired multiple blood-brain barrier open bleeding assessment samples to construct a training dataset.

[0120] The model training module 803 is used to train a deep learning model using the training dataset and generate a blood-brain barrier opening and hemorrhage prediction result corresponding to each short time-domain acoustic signal segment in the acoustic signal training sample.

[0121] The model building module 804 is used to build a blood-brain barrier open hemorrhage prediction model by adjusting the parameters of the deep learning model based on the blood-brain barrier open hemorrhage prediction results and blood-brain barrier open hemorrhage assessment samples corresponding to the acoustic signal training samples and based on a preset loss function.

[0122] The prediction module 805 is used to obtain blood-brain barrier opening hemorrhage result data based on the blood-brain barrier opening hemorrhage prediction model and the real-time acoustic signal samples obtained by ultrasound combined with contrast agent in the process of regulating the blood-brain barrier.

[0123] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0124] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0125] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the blood-brain barrier opening hemorrhage prediction method based on time-domain signals as described above.

[0126] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to execute the blood-brain barrier opening hemorrhage prediction method based on time-domain signals as described above.

[0127] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0128] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0129] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0132] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0133] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs, etc.).

[0134] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0136] In summary, this application provides a method, system, and medium for predicting blood-brain barrier open bleeding based on time-domain signals. It segments all pulsed acoustic signals from a single treatment into multiple short-time-domain acoustic signal segments and inputs them into the model. A time-frequency relationship coding module simultaneously extracts features from the acoustic signals in both the frequency and time domains, improving temporal resolution and dynamic change capture capabilities while reducing information loss during frequency domain conversion. Furthermore, it eliminates the need for MRI contrast agents, maintaining in vivo detection capabilities and achieving non-invasive detection. Further, through a gated aggregation module with a gated attention mechanism and a multi-path classification optimization design, the model's ability to distinguish features from different states is enhanced, improving classification accuracy and predictive performance. Therefore, this application effectively overcomes the various shortcomings of existing technologies and possesses high industrial applicability.

[0137] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A blood brain barrier opening hemorrhage prediction method based on a time domain signal, characterized in that, The application relates to a blood-brain barrier opening hemorrhage prediction method based on deep learning. The method comprises the following steps: acquiring a plurality of historical acoustic signal samples of ultrasound combined with a contrast agent in the process of regulating a blood-brain barrier, and respectively performing data preprocessing on each of the historical acoustic signal samples to obtain a plurality of acoustic signal training samples; wherein each of the acoustic signal training samples comprises a plurality of short-time domain acoustic signal segments; pairing each of the acoustic signal training samples with a plurality of blood-brain barrier opening hemorrhage evaluation samples obtained to construct a training data set; training a deep learning model by using the training data set, and generating a blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample according to each short-time domain acoustic signal segment in the acoustic signal training sample; the deep learning model comprises a time-frequency relationship coding module, a gating aggregation module and a state classification module; wherein the method for generating the blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample comprises: performing time-frequency domain feature fusion and adjacent segment relationship modeling operations on each short-time domain acoustic signal segment in the acoustic signal training sample by using the time-frequency relationship coding module to obtain an acoustic coding feature vector corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample; performing dynamic weighted summation operations on the acoustic coding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample by using the gating aggregation module to generate a gating aggregation feature vector corresponding to a blood-brain barrier opening hemorrhage state of different categories; performing mapping operations on the gating aggregation feature vectors corresponding to the blood-brain barrier opening hemorrhage states of different categories by using the state classification module to obtain prediction scores corresponding to the blood-brain barrier opening hemorrhage states of different categories, and selecting a blood-brain barrier opening hemorrhage state corresponding to the highest prediction score as the blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample; adjusting parameters of the deep learning model according to the blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample and the blood-brain barrier opening hemorrhage evaluation sample and based on a preset loss function to construct a blood-brain barrier opening hemorrhage prediction model; 2. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 1, wherein, based on the blood-brain barrier opening hemorrhage prediction model and according to a real-time acoustic signal sample obtained in the process of regulating the blood-brain barrier by using the ultrasound combined with the contrast agent, obtaining blood-brain barrier opening hemorrhage result data. Each of the historical acoustic signal samples comprises a plurality of pulse time domain acoustic signals; wherein the method for performing data preprocessing on each of the historical acoustic signal samples to obtain a plurality of acoustic signal training samples comprises: performing time sequence splicing preprocessing on each pulse time domain acoustic signal in the historical acoustic signal sample to generate a continuous long-time domain acoustic signal; performing segmentation and cutting preprocessing on the continuous long-time domain acoustic signal to generate a plurality of short-time domain acoustic signal segments and construct an acoustic signal training sample.

3. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 1, wherein, The time-frequency relationship encoding module comprises a convolutional encoder, a spectral encoder and a pyramid adjacent encoding module; wherein the time-frequency relationship encoding module is used to perform time-frequency domain feature fusion and adjacent segment relationship modeling operation on each short-time domain acoustic signal segment in the acoustic signal training sample, to obtain the acoustic encoding feature vector corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample in the following manner: The convolutional encoder is used to perform time domain feature extraction operation on each short-time domain acoustic signal segment in the acoustic signal training sample, to obtain the time domain feature vector corresponding to each short-time domain acoustic signal segment; The spectral encoder is used to perform frequency domain feature extraction operation on each short-time domain acoustic signal segment in the acoustic signal training sample, to obtain the frequency domain feature vector corresponding to each short-time domain acoustic signal segment; The element-by-element addition algorithm is used to perform feature fusion operation on the time domain feature vector and the frequency domain feature vector corresponding to each short-time domain acoustic signal segment, to obtain the fusion feature vector corresponding to each short-time domain acoustic signal segment; After layer normalization processing of the fusion feature vector corresponding to each short-time domain acoustic signal segment, the pyramid adjacent encoding module is used to perform modeling operation on the fusion feature vectors corresponding to adjacent short-time domain acoustic signal segments after layer normalization processing, to obtain the acoustic encoding feature vector corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample.

4. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 1, wherein, The gating aggregation module comprises a plurality of structurally identical gating sub-modules; wherein the gating aggregation module is used to perform dynamic weighted summation operation on the acoustic encoding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample, to generate the gating aggregation feature vector corresponding to the blood-brain barrier opening hemorrhage state of different categories in the following manner: The acoustic encoding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample are input in parallel to each gating sub-module; Each gating sub-module performs dynamic weighted summation operation on the acoustic encoding feature vectors corresponding to each short-time domain acoustic signal segment in the acoustic signal training sample based on the gating attention mechanism, to output the gating aggregation feature vector corresponding to the blood-brain barrier opening hemorrhage state of the corresponding category.

5. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 1, wherein, The state classification module comprises a plurality of fourth fully connected layers with different parameters; wherein the state classification module is used to perform mapping operation on the gating aggregation feature vectors corresponding to the blood-brain barrier opening hemorrhage state of different categories, to obtain the prediction score corresponding to the blood-brain barrier opening hemorrhage state of different categories in the following manner: The gating aggregation feature vectors corresponding to the blood-brain barrier opening hemorrhage state of different categories are input into the corresponding fourth fully connected layers respectively; Each fourth fully connected layer performs category mapping on the gating aggregation feature vector corresponding to the input blood-brain barrier opening hemorrhage state, to generate the prediction score corresponding to the blood-brain barrier opening hemorrhage state of different categories.

6. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 3, wherein, The convolutional encoder comprises a first normalization layer, a convolutional module, a first full connection layer, and a first activation function layer; the spectral encoder comprises an exponential transformation layer, a second normalization layer, a second full connection layer, and a second activation function layer; and the pyramid adjacent coding module comprises a plurality of one-dimensional convolutions.

7. The time-domain signal based blood brain barrier opening hemorrhage prediction method of claim 6, wherein, The convolutional module comprises a convolutional layer, a third activation function layer, and a maximum pooling layer.

8. A blood brain barrier opening hemorrhage prediction system based on time domain signals, characterized by, Comprise: A data preprocessing module is configured to acquire a plurality of historical acoustic signal samples of an ultrasound contrast agent in the process of regulating the blood-brain barrier, and perform data preprocessing on each of the historical acoustic signal samples to obtain a plurality of acoustic signal training samples; each of the acoustic signal training samples comprises a plurality of short-time acoustic signal segments. A data pairing module is configured to pair each of the acoustic signal training samples with a plurality of blood-brain barrier opening hemorrhage evaluation samples acquired to construct a training data set. A model training module is configured to train a deep learning model using the training data set, and generate a blood-brain barrier opening hemorrhage prediction result corresponding to each of the acoustic signal training samples according to each of the short-time acoustic signal segments in the acoustic signal training sample; the deep learning model comprises a time-frequency relationship coding module, a gating aggregation module, and a state classification module; wherein the blood-brain barrier opening hemorrhage prediction result corresponding to each of the acoustic signal training samples is generated in the following manner: The time-frequency relationship coding module is used to perform time-frequency domain feature fusion and adjacent segment relationship modeling operations on each of the short-time acoustic signal segments in the acoustic signal training sample to obtain an acoustic coding feature vector corresponding to each of the short-time acoustic signal segments in the acoustic signal training sample; The gating aggregation module is used to perform dynamic weighted summation operations on the acoustic coding feature vectors corresponding to each of the short-time acoustic signal segments in the acoustic signal training sample to generate a gating aggregation feature vector corresponding to different categories of blood-brain barrier opening hemorrhage states; The state classification module is used to perform mapping operations on the gating aggregation feature vectors corresponding to different categories of blood-brain barrier opening hemorrhage states to obtain prediction scores corresponding to different categories of blood-brain barrier opening hemorrhage states, and select a blood-brain barrier opening hemorrhage state corresponding to the highest prediction score as the blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample; A model construction module is configured to adjust parameters of the deep learning model based on a preset loss function according to the blood-brain barrier opening hemorrhage prediction result corresponding to the acoustic signal training sample and the blood-brain barrier opening hemorrhage evaluation sample to construct a blood-brain barrier opening hemorrhage prediction model; A prediction module is configured to obtain blood-brain barrier opening hemorrhage result data based on the blood-brain barrier opening hemorrhage prediction model and according to a real-time acoustic signal sample of an ultrasound contrast agent in the process of regulating the blood-brain barrier.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the blood-brain barrier opening hemorrhage prediction method based on a time domain signal according to any one of claims 1-7.

Images