Craniocerebral image detection method and model construction method based on domain generalization neural network
Through the method based on domain generalization neural network, the problems of high data labeling cost and domain drift in cranial brain image detection are solved, and accurate diagnosis of different populations and patients is achieved, and the efficiency and accuracy of cranial brain health detection are improved.
Patent Information
- Application Number
- CN202410920695.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-12
AI Technical Summary
The existing cranial brain image detection method based on artificial intelligence technology is difficult to achieve accurate diagnosis among different populations and patients due to the high cost of data labeling and domain drift.
Using a domain generalization neural network method, the mixed loss function of the maximum mean difference measurement loss and classification loss is designed by collecting the cranial brain ultrasonic image data of different populations, and the cranial brain image detection model is trained to enable the model to learn data from different domains. Distribute characteristics to enhance its generalization ability.
Accurate diagnosis among different groups of people and patients is achieved, the cost of data collection and labeling is reduced, domain drift is alleviated, and the efficiency and accuracy of craniocerebral health diagnosis is improved.
Smart Images

Figure CN118941501A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of medical image detection based on computer vision, and in particular relates to a cranial brain image detection method based on a domain generalization neural network and a model construction method. Background Art
[0002] Ultrasonic testing is based on the propagation characteristics of ultrasound in human tissue. Ultrasonic waves are mechanical waves with frequencies higher than the human hearing range and good penetrating power. When ultrasound encounters tissue interfaces with different densities or acoustic impedances, reflection, refraction, scattering, and attenuation occur. These phenomena provide basic information for imaging.
[0003] Under the current hierarchical medical system, the accuracy of craniocerebral diagnosis cannot meet the needs due to the influence of multiple factors such as the professional training level of diagnostic doctors, the experience of inspectors, and instruments and equipment. Using artificial intelligence technologies such as fully connected networks, recurrent neural networks such as LSTM, and classic convolutional neural networks to accurately predict the health status of the craniocerebral has become a hot topic today. However, existing artificial intelligence-based technologies usually require a large amount of labeled data. In the real medical world, it is difficult to collect enough labeled data due to the high cost of collecting cases and asking experts to manually label them, which makes it impossible for existing neural networks to perform at their maximum performance.
[0004] In addition, existing medical image detection based on artificial intelligence technology, especially deep learning technology, follows a basic assumption that the training data and the prediction data conform to the same distribution, which is also an important reason why deep learning technology can achieve the highest accuracy in various fields. However, due to the specificity of the human body, there is a certain domain drift between the ultrasound data collected from different healthy people or patients and the data generated in the target human body to be tested, which breaks the basic assumptions followed by deep learning, seriously weakens the performance of the neural network reasoning stage, and makes the diagnosis results fail to meet the requirements. Summary of the invention
[0005] In view of the above problems, the first aspect of the present invention provides a method for constructing a cranial brain image detection model based on a domain generalization neural network, comprising the following process:
[0006] Step 1: collect brain ultrasound image datasets of different populations and annotate a portion of the image data; divide the entire image dataset into source domain 1 data and source domain 2 data;
[0007] Step 2, preprocessing the image data set, wherein the preprocessing includes smoothing filtering to remove noise;
[0008] Step 3, constructing a cranial brain image detection model, which includes a feature extraction network and a classification neural network;
[0009] Step 4: Design a hybrid loss function based on the maximum mean difference metric loss and classification loss to train the cranial brain image detection model; enable the model to learn the differences in distribution characteristics of source domain 1 data and source domain 2 data, enhance the generalization ability of the model, and obtain the final cranial brain image detection model.
[0010] Preferably, the smoothing filtering to remove noise in step 2 is specifically as follows:
[0011] Take the average value of adjacent pixels in the image, for a grayscale image A:
[0012]
[0013] Among them, A is the grayscale image pixel matrix, a ij is the pixel value at position (i, j) in the image;
[0014] Define a window of size L and calculate the mean of the elements in the window as the corresponding window Output value at position
[0015] The formula is as follows:
[0016]
[0017] Among them, Vmean is the mean of the pixels in the window, and L is the batch size.
[0018] Preferably, in step 1, a data set of cranial ultrasound images of different populations is collected, including ultrasound image data of different cranial health conditions of adolescents, middle-aged people, and the elderly; the cranial ultrasound images of adolescents and middle-aged groups are used as source domain 1 data, and the cranial ultrasound images of the elderly group are used as source domain 2 data; or the labeled image data are used as source domain 1 data, and the unlabeled data are used as source domain 2 data.
[0019] Preferably, the feature extraction network in step 3 includes a global average pooling layer, a first fully connected layer, a ReLU activation function layer, a second fully connected layer and a Sigmoid activation function layer connected in sequence; first, the global average pooling layer compresses the feature map of each channel of the patient's cranial ultrasound image into a scalar, then the first fully connected layer compresses the dimension of each scalar, then the ReLU activation function is used to perform a nonlinear change on the compressed scalar, then the second fully connected layer restores the dimension of the compressed scalar after the nonlinear transformation, and finally the Sigmoid activation function is used to obtain the weight of each channel.
[0020] Preferably, the classification neural network includes two fully connected layers and one Softmax activation function layer, wherein the first fully connected layer is responsible for mapping the features obtained by the feature extraction network into high-dimensional implicit features, the second fully connected layer converts the high-dimensional implicit notifications into brain health scores, and the Softmax activation function layer converts the brain health scores into brain health prediction values; wherein the first fully connected layer is responsible for mapping the features obtained by the feature extraction network into 512 dimensions, and the second fully connected layer maps the 512-dimensional features to specific scores of different brain health conditions.
[0021] Preferably, the step 4 trains the cranial brain image detection model, inputs a small amount of labeled data of the source domain 1 data and the source domain 2 data into the feature extraction network, and maps the extracted features to the Hilbert space through the Gaussian kernel function. In this space, the minimum feature distance difference is obtained by the maximum mean difference measurement method, and the feature distance difference is continuously optimized through network iteration, so that the feature distance becomes closer and closer; at the same time, deep correlation alignment is used to align the conditional features step by step; and a classification neural network is used to extract sufficient discriminative information from the domain invariant features, so that it has a stronger classification ability in unseen data.
[0022] Preferably, a hybrid loss function based on the maximum mean difference metric loss and classification loss is designed, and the mathematical expression is as follows:
[0023] L all =L ce +λL dg
[0024] Among them, L ce is the classification loss function, L dg is the maximum mean difference metric loss function, and λ is the penalty coefficient, which is used to determine the degree of domain generalization;
[0025] The classification loss function L ce The mathematical expression is as follows:
[0026]
[0027] Among them, n is the number of samples, y is the true label of the cranial brain image, and a is the predicted value of the cranial brain image detection model;
[0028] The mathematical expression of the maximum mean difference metric loss function is finally shown as follows:
[0029]
[0030] Where σ is the regenerated Hilbert space mapping function, Y 1 is the source domain 1 sample X 1The corresponding label, Y 2 is the source domain 2 sample X 2 The corresponding label, P(Y 1 =c) is the probability value of category c in the source domain 1, P(Y 2 =c) is the probability value of category c in source domain 2;
[0031] The process of updating the neural network parameters based on the loss function is as follows:
[0032]
[0033] Among them, ε is the learning rate, γ and λ are balance parameters, Θ fe are neural network parameters.
[0034] The second aspect of the present invention provides a method for cranial brain image detection based on a domain generalization neural network, characterized in that the cranial brain image detection model constructed by the construction method described in the first aspect is deployed in a medical system, and includes the following process:
[0035] Step 1, obtaining a cranial ultrasound image of a patient to be diagnosed;
[0036] Step 2: input the image into the deployed trained brain image detection model;
[0037] Step 3: Output the image detection results for doctors to make a diagnosis.
[0038] The third aspect of the present invention provides a cranial brain image detection device based on a domain generalization neural network, characterized in that: the device includes at least one processor and at least one memory, the processor and the memory are coupled; the memory stores a computer execution program of the cranial brain image detection model constructed by the construction method described in the first aspect; when the processor executes the computer execution program stored in the memory, the processor executes a cranial brain image detection method based on a domain generalization neural network.
[0039] The fourth aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program or instruction of a cranial brain image detection model constructed by the construction method described in the first aspect, and when the program or instruction is executed by a processor, the processor executes a cranial brain image detection method based on a domain generalization neural network.
[0040] Compared with the prior art, the present invention has the following beneficial effects:
[0041] The present invention utilizes domain generalization technology, which can not only overcome the above-mentioned data collection difficulties and data labeling difficulties, but also greatly alleviate the domain drift phenomenon and enhance the practicality of artificial intelligence technology in craniocerebral health diagnosis. Specifically, it includes:
[0042] (1) Only one network training is required to perform more accurate real-time detection of ultrasound images of all new patients, greatly improving the efficiency of examinations.
[0043] (2) The intelligent detection algorithms of traditional neural networks all follow a basic assumption, that is, the training data and the test data are independent and identically distributed. However, due to the differences in human bodies and diseases, the actual situation often violates this basic assumption, resulting in the serious weakening of the performance of the traditional neural network method in the reasoning stage. This method does not require the patient's ultrasound image to meet the basic assumption of independent and identical distribution, which greatly improves the universality and application value of the algorithm.
[0044] (3) Manual labeling of medical images often requires experienced people to spend a lot of time labeling, which is often too costly. This method does not require manual labeling of all people and all diseases, which can greatly reduce the cost of data collection. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is the overall structural block diagram of the cranial brain image detection model of the present invention.
[0046] Figure 2 This is a flow chart of the brain image detection model training process of the present invention.
[0047] Figure 3 This is an accuracy iteration diagram of the source domain 1 neural network and the source domain 2 neural network during the training process in Example 1 of the present invention.
[0048] Figure 4 This is a comparison chart of the predicted value of craniocerebral health status and the true label in Example 1 of the present invention.
[0049] Figure 5 This is a simplified structural diagram of the cranial brain image detection device in Example 2. DETAILED DESCRIPTION
[0050] The present invention is further described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] Embodiment 1:
[0052] The overall concept of the present invention is to combine feature distance measurement with classification neural network to propose an efficient and accurate domain generalization craniocerebral health diagnosis method. First of all, there are certain differences in the data characterizing the craniocerebral health status collected from different human bodies, but they all have specific features that can characterize the craniocerebral health status. This part of the features has characteristics that are irrelevant to the human body and are only related to the health status. In order to extract this part of the key features, a feature distance difference measurement method is introduced-the maximum mean difference. Using the kernel method, the data passing through the backbone network is mapped to the Hilbert space, and the differences between the features collected from different human bodies are measured. By reducing this part of the difference, the extracted features have domain-invariant characteristics within a certain dimension, that is, in the Hilbert space, the extracted features are common characteristics.
[0053] Later, in order to make the common features discriminative, a classification neural network was introduced, using its powerful feature learning ability to deeply explore the discriminativeness in the common features, thereby improving the accuracy of craniocerebral health diagnosis. By combining the feature distance difference metric with the convolutional neural network and making full use of the advantages of both, the challenges of craniocerebral health diagnosis can be better addressed. Finally, by training the optimized neural network model, the accuracy of craniocerebral health diagnosis was improved, effectively solving the challenges faced by the current intelligent craniocerebral health diagnosis methods.
[0054] The overall process of the brain image detection method based on the domain generalization neural network of the present invention includes:
[0055] Step 1: collect brain ultrasound image datasets of different populations and annotate a portion of the image data; divide the entire image dataset into source domain 1 data and source domain 2 data;
[0056] Step 2, preprocessing the image data set, wherein the preprocessing includes smoothing filtering to remove noise;
[0057] Step 3, constructing a cranial brain image detection model, which includes a feature extraction network and a classification neural network;
[0058] Step 4: Design a hybrid loss function based on the maximum mean difference metric loss and classification loss to train the cranial brain image detection model; enable the model to learn the differences in distribution characteristics of source domain 1 data and source domain 2 data, enhance the generalization ability of the model, and obtain the final cranial brain image detection model.
[0059] Step 5: deploy the final model in the medical system, obtain the cranial ultrasound image of the patient to be diagnosed, and output the image detection results for the doctor to make a diagnosis.
[0060] 1. Data Collection
[0061] Doppler ultrasound (TCCS) technology is used, and a low-frequency linear array is selected. After finely adjusting the probe working parameters, including the transmission frequency, gain, depth, and dynamic range, the probe is placed in the scanning stage, and the key images are frozen to store important images. The stored important images are handed over to professional doctors, who manually label them. Compared with conventional methods, the proposed method has a small number of acquisitions and labels, which greatly saves the cost of data acquisition.
[0062] It can include the collection of ultrasonic image data of different age groups (teenagers, middle-aged, elderly) and different brain health conditions (normal, brain atrophy, brain tumor, hydrocephalus, cerebral hemorrhage), and manually labeling some of them. The brain ultrasound images of adolescents and middle-aged groups are used as source domain 1 data, and the brain ultrasound images of the elderly group are used as source domain 2 data; or the labeled image data is used as source domain 1 data, and the unlabeled data is used as source domain 2 data.
[0063] 2. Data Preprocessing
[0064] Including the process of image denoising; using smoothing filtering to remove noise, taking the average value of adjacent pixels in the image, for a grayscale image A:
[0065]
[0066] Among them, A is the grayscale image pixel matrix, a ij is the pixel value at position (i, j) in the image;
[0067] Define a window of size L and calculate the mean of the elements in the window as the corresponding window Output value at position
[0068] The formula is as follows:
[0069]
[0070] Among them, Vmean is the mean of the pixels in the window, and L is the batch size.
[0071] 3. Brain image detection model
[0072] The overall block diagram of the brain image detection model is as follows: Figure 1 As shown in the figure, it mainly includes: feature extraction network and classification neural network. The advantage of the structure of this model is that there is no need to manually extract features, and the feature fusion strategy can ensure the reliability and credibility of the input data. Secondly, it can solve the problems of data collection difficulties and data labeling difficulties currently faced by deep learning technology. Thirdly, the proposed method can alleviate the problem of domain drift of data signals caused by differences in human body structure. In summary, the structure of the proposed method is suitable for the diagnosis of the health status of the cranial brain.
[0073] The feature extraction network includes a global average pooling layer, a fully connected layer 1, a ReLU activation function layer, a fully connected layer 2, and a Sigmoid activation function layer connected in sequence; first, the global average pooling layer compresses the feature map of each channel of the patient's cranial ultrasound image into a scalar, then the fully connected layer 1 compresses the dimension of each scalar, and then the ReLU activation function is used to perform nonlinear changes on the compressed scalar, and then the fully connected layer 2 restores the dimension of the compressed scalar after the nonlinear transformation, and finally the Sigmoid activation function is used to obtain the weight of each channel.
[0074] The classification neural network consists of two fully connected layers and one Softmax activation function layer. The first fully connected layer is responsible for mapping the features obtained by the feature extraction network into high-dimensional implicit features, the second fully connected layer converts the high-dimensional implicit notifications into craniocerebral health status scores, and the Softmax activation function layer converts the craniocerebral health status scores into craniocerebral health status prediction values.
[0075] The specific parameters are shown in Table 1:
[0076] Table 1 Specific parameters of classification neural network
[0077]
[0078] 4. Hybrid loss function design
[0079] The mathematical expression of the hybrid loss function is as follows:
[0080] L all =L ce +λL dg
[0081] Among them, L ce is the classification loss function, L dg is the maximum mean difference metric loss function, and λ is the penalty coefficient, which is used to determine the degree of domain generalization.
[0082] A small amount of labeled data from source domain 1 and source domain 2 is input into the feature extraction network, and the extracted features are mapped to the Hilbert space through the Gaussian kernel function. In this space, the minimum feature distance difference is obtained through the maximum mean difference measurement method, and the feature distance difference is continuously optimized through network iteration, making the feature distance closer and closer. At the same time, deep correlation alignment is used to align the conditional features step by step. Using the classification neural network, sufficient discriminative information is extracted from the domain invariant features, so that it can have a strong classification ability in unseen data.
[0083] Classification loss function L ce The mathematical expression is as follows:
[0084]
[0085] Among them, n is the number of samples, y is the true label of the brain image, and a is the predicted value of the brain image detection model.
[0086] Domain generalization includes deep correlation comparison and maximum mean difference measurement; the collected data is divided into source domain 1 data and source domain 2 data, and based on empirical risk minimization, its feature distance measurement can be expressed as
[0087]
[0088] in is a set of functions on the sample space, and p and q are the characteristic distributions of the collected craniocerebral health status data and the craniocerebral data to be tested, respectively. represents the expected value of function f on random variable x under distribution p, It represents the function f that maximizes the above difference among all possible function sets. Furthermore, for the convenience of calculation, the formula is rewritten as:
[0089]
[0090] Among them, n 1 and n 2 Represent the batch sizes of source domain 1 data and source domain 2 data respectively. H is the regenerated Hilbert space, Φ represents the mapping function that maps source domain 1 data samples and source domain 2 data samples to the regenerated Hilbert space, and X 1 is the data sample of source domain 1, X 2 is the data sample of source domain 2, is the source domain 1 data sample dataset, is a sample data set of source domain 2 data. After calculation by the formula, the feature distance between source domain 1 data and source domain 2 data can be measured. When the feature distance is small enough, it means that the extracted feature is a domain-invariant feature. However, since the above method can only ensure the alignment of edge feature distribution, the conditional feature distribution cannot be aligned, so the effectiveness of the domain-invariant feature cannot be guaranteed. Therefore, deep correlation comparison is added to align the conditional feature distribution of source domain 1 data and source domain 2 data to provide higher quality domain-invariant features. The calculation formula is as follows:
[0091]
[0092] Among them, Cov 1 With Cov 2 are the covariances of source domain 1 and source domain 2, is a constant coefficient used to scale the final loss to ensure that its value is moderate and does not depend on the feature dimension. Its calculation formula is as follows:
[0093]
[0094] Where I is a row vector whose elements are all equal to 1.
[0095] The mathematical expression of the maximum mean difference metric loss function is finally shown as follows:
[0096]
[0097] Where σ is the regenerated Hilbert space mapping function. 1 is the source domain 1 data sample X 1 The corresponding label, Y 2 is the source domain 2 data sample X 2 The corresponding label, P(Y 1 =c) is the probability value of category c in the source domain 1 data, P(Y 2 =c) is the probability value of category c in the source domain 2 data.
[0098] The process of updating neural network parameters based on the mixed loss function is shown in the formula:
[0099]
[0100] Among them, ε is the learning rate, γ and λ are balance parameters, Θ fe are neural network parameters.
[0101] The network hyperparameters use a decreasing learning rate, and the initial learning rate is 0.0005. The number of iterations of the fusion neural network is 500, the batch size is 64, and the cross entropy is used as the loss function. The activation function of all hidden layers uses the ReLU function. The optimizer uses Adam optimization, and the dropout is set to 0.5 in the training phase and 1.0 in the test phase. The overall process is as follows Figure 2 shown.
[0102] 5. Experimental results verification
[0103] Figure 3 The curves of the accuracy rate changing with the number of iterations during the training of the neural network in source domain 1 and the neural network in source domain 2. It can be seen that with the increase of the number of iterations, the accuracy rate of the neural network in source domain 1 and the neural network in source domain 2 increases rapidly, proving that the model constructed by the present invention has a good convergence speed.
[0104] Figure 4 The comparison between the prediction results of the test set samples by the model of the present invention and the true labels. It can be seen that the model of the present invention can correctly predict the labels of most samples, proving that the technology proposed by the present invention has a high prediction accuracy.
[0105] Embodiment 2:
[0106] like Figure 5 As shown, the present invention also provides a brain image detection device based on a domain generalized neural network, the device includes at least one processor and at least one memory, and also includes a communication interface and an internal bus; the memory stores a computer execution program of a brain image detection model constructed by the construction method described in Example 1; when the processor executes the computer execution program stored in the memory, the processor can execute a brain image detection method based on a domain generalized neural network. Wherein the internal bus can be an industry standard architecture (Industry Standard Architecture, ISA) bus, a peripheral component interconnect (Peripheral Component, PCI) bus or an extended industry standard architecture (Xtended Industry Standard Architecture, EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of the present application is not limited to only one bus or one type of bus. Wherein the memory may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a disk or an optical disk, etc.
[0107] The device may be provided as a terminal, a server or other forms of devices.
[0108] Figure 5 is a block diagram of a device shown as an example. The device may include one or more of the following components: a processing component, a memory, a power component, a multimedia component, an audio component, an input / output (I / O) interface, a sensor component, and a communication component. The processing component generally controls the overall operation of the electronic device, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component may include one or more processors to execute instructions to complete all or part of the steps of the above method. In addition, the processing component may include one or more modules to facilitate interaction between the processing component and other components. For example, the processing component may include a multimedia module to facilitate interaction between the multimedia component and the processing component.
[0109] The memory is configured to store various types of data to support operations on the electronic device. Examples of such data include instructions for any application or method operating on the electronic device, contact data, phone book data, messages, pictures, videos, etc. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.
[0110] The power supply assembly provides power to various components of the electronic device. The power supply assembly may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device. The multimedia assembly includes a screen that provides an output interface between the electronic device and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touch, slide, and gestures on the touch panel. The touch sensor may not only sense the boundaries of the touch or slide action, but also detect the duration and pressure associated with the touch or slide operation. In some embodiments, the multimedia assembly includes a front camera and / or a rear camera. When the electronic device is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera may receive external multimedia data. Each front camera and rear camera may be a fixed optical lens system or have a focal length and optical zoom capability.
[0111] The audio component is configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC), and when the electronic device is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode, the microphone is configured to receive an external audio signal. The received audio signal can be further stored in a memory or sent via a communication component. In some embodiments, the audio component also includes a speaker for outputting an audio signal. The I / O interface provides an interface between the processing component and the peripheral interface module, and the above-mentioned peripheral interface module can be a keyboard, a click wheel, a button, etc. These buttons may include, but are not limited to: a home button, a volume button, a start button, and a lock button.
[0112] The sensor assembly includes one or more sensors for providing various aspects of status assessment for the electronic device. For example, the sensor assembly can detect the on / off state of the electronic device, the relative positioning of components, such as the display and keypad of the electronic device, and the sensor assembly can also detect the position change of the electronic device or a component of the electronic device, the presence or absence of user contact with the electronic device, the orientation or acceleration / deceleration of the electronic device and the temperature change of the electronic device. The sensor assembly may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly may also include an accelerometer, a gyroscope sensor, a magnetic sensor, a pressure sensor or a temperature sensor.
[0113] The communication component is configured to facilitate wired or wireless communication between the electronic device and other devices. The electronic device can access a wireless network based on a communication standard, such as WiFi, 2G or 3G, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
[0114] In an exemplary embodiment, the electronic device may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above methods.
[0115] Embodiment 3:
[0116] The present invention also provides a computer-readable storage medium, which stores a computer program or instruction of a cranial brain image detection model constructed by the construction method described in Example 1. When the program or instruction is executed by a processor, the processor can execute a cranial brain image detection method based on a domain generalization neural network.
[0117] Specifically, a system, device or equipment equipped with a readable storage medium may be provided, on which a software program code for implementing the functions of any of the above-mentioned embodiments is stored, and a computer or processor of the system, device or equipment reads and executes the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the above-mentioned embodiments, so the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.
[0118] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk (such as CD-ROM, CD-R, CD-RW, DVD-20ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tape, etc. The storage medium can be any available medium that can be accessed by a general or special-purpose computer.
[0119] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the invention may be directly implemented by a hardware processor or implemented by a combination of hardware and software modules in the processor.
[0120] It should be understood that the storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in a terminal or server.
[0121] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.
[0122] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.
[0123] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
[0124] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. A method for constructing a cranial brain image detection model based on a domain generalization neural network, characterized in that: The process includes: Step 1: collect brain ultrasound image datasets of different populations and annotate a portion of the image data; divide the entire image dataset into source domain 1 data and source domain 2 data; Step 2, preprocessing the image data set, wherein the preprocessing includes smoothing filtering to remove noise; Step 3, constructing a brain image detection model, which includes a feature extraction network and a classification neural network; Step 4: Design a hybrid loss function based on the maximum mean difference metric loss and classification loss to train the cranial brain image detection model; enable the model to learn the differences in distribution characteristics of source domain 1 data and source domain 2 data, enhance the generalization ability of the model, and obtain the final cranial brain image detection model.
2. A method for constructing a cranial brain image detection model based on a domain generalization neural network as claimed in claim 1, characterized in that: The smoothing filtering and noise removal in step 2 are specifically as follows: Take the average value of adjacent pixels in the image, for a grayscale image A: Among them, A is the grayscale image pixel matrix, a ij is the pixel value at position (i, j) in the image; Define a window of size L and calculate the mean of the elements in the window as the corresponding window Output value at position The formula is as follows: Among them, Vmean is the mean of the pixels in the window, and L is the batch size.
3. The method for constructing a cranial brain image detection model based on a domain generalization neural network as claimed in claim 1, characterized in that: In the step 1, a cranial ultrasound image dataset of different populations is collected, including ultrasound image data of different cranial and brain health conditions of adolescents, middle-aged people, and the elderly; the cranial ultrasound images of adolescents and middle-aged groups are used as source domain 1 data, and the cranial ultrasound images of the elderly group are used as source domain 2 data; or the labeled image data is used as source domain 1 data, and the unlabeled data is used as source domain 2 data.
4. The method for constructing a cranial brain image detection model based on a domain generalization neural network as claimed in claim 1, characterized in that: The feature extraction network in step 3 includes a global average pooling layer, a first fully connected layer, a ReLU activation function layer, a second fully connected layer and a Sigmoid activation function layer connected in sequence; first, the global average pooling layer compresses the feature map of each channel of the patient's cranial ultrasound image into a scalar, then the first fully connected layer compresses the dimension of each scalar, then the ReLU activation function is used to perform a nonlinear change on the compressed scalar, then the second fully connected layer restores the dimension of the compressed scalar after the nonlinear transformation, and finally the Sigmoid activation function is used to obtain the weight of each channel.
5. The method for constructing a cranial brain image detection model based on a domain generalization neural network as claimed in claim 1, characterized in that: The classification neural network includes two fully connected layers and a Softmax activation function layer, wherein the first fully connected layer is responsible for mapping the features obtained by the feature extraction network into high-dimensional implicit features, the second fully connected layer converts the high-dimensional implicit notifications into brain health status scores, and the Softmax activation function layer converts the brain health status scores into brain health status prediction values; wherein the first fully connected layer is responsible for mapping the features obtained by the feature extraction network into 512 dimensions, and the second fully connected layer maps the 512-dimensional features to specific scores of different brain health statuses.
6. The method for constructing a cranial brain image detection model based on a domain generalization neural network according to claim 1, characterized in that: The step 4 trains the cranial brain image detection model, inputs a small amount of labeled data of source domain 1 and source domain 2 into the feature extraction network, and maps the extracted features to the Hilbert space through the Gaussian kernel function. In this space, the minimum feature distance difference is obtained by the maximum mean difference measurement method, and the feature distance difference is continuously optimized through network iteration, so that the feature distance is closer and closer; at the same time, deep correlation alignment is used to align the conditional features step by step; and a classification neural network is used to extract sufficient discriminative information from the domain invariant features, so that it has a stronger classification ability in unseen data.
7. The method for constructing a cranial brain image detection model based on a domain generalization neural network according to claim 6, characterized in that: Design a hybrid loss function based on the maximum mean difference metric loss and classification loss. The mathematical expression is as follows: Lallll=Lce+λldg Among them, L ce is the classification loss function, L dg is the maximum mean difference metric loss function, and λ is the penalty coefficient, which is used to determine the degree of domain generalization; The classification loss function L ce The mathematical expression is as follows: Among them, n is the number of samples, y is the true label of the cranial brain image, and a is the predicted value of the cranial brain image detection model; The mathematical expression of the maximum mean difference metric loss function is finally shown as follows: Where σ is the regenerated Hilbert space mapping function, Y1 is the label corresponding to the sample X1 in source domain 1, Y2 is the label corresponding to the sample X2 in source domain 2, P(Y1=c) is the probability value of category c in source domain 1, and P(Y2=c) is the probability value of category c in source domain 2; The process of updating the neural network parameters based on the loss function is as follows: Among them, ε is the learning rate, γ and λ are balance parameters, Θ fe are neural network parameters.
8. A method for cranial brain image detection based on a domain generalization neural network, characterized in that: The cranial brain image detection model constructed by the construction method according to any one of claims 1 to 7 is deployed in a medical system, and includes the following process: Step 1, obtaining a cranial ultrasound image of a patient to be diagnosed; Step 2: Input the image into the deployed trained brain image detection model; Step 3: Output the image detection results for doctors to make a diagnosis.
9. A brain image detection device based on a domain generalization neural network, characterized in that: The device includes at least one processor and at least one memory, the processor and the memory are coupled; the memory stores a computer execution program of a cranial brain image detection model constructed by the construction method according to any one of claims 1 to 7; when the processor executes the computer execution program stored in the memory, the processor executes a cranial brain image detection method based on a domain generalization neural network.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program or instruction of a cranial brain image detection model constructed by the construction method according to any one of claims 1 to 7. When the program or instruction is executed by a processor, the processor executes a cranial brain image detection method based on a domain generalization neural network.
Citation Information
Patent Citations
Unsupervised domain-adaptive brain tumor semantic segmentation method based on deep adversarial learning
CN108062753A
Domain generalization method and device of image processing model, and image processing method and device
CN115880538A
Fault diagnosis method based on semi-supervised confrontation domain generalization intelligent model
CN116089812A
Heterogeneous domain-based medical image classification method and system
CN116128876A
