Brain tumor classification method and system based on circuit bootstrap and SNN

Abstract

The invention provides a brain tumor classification method and system based on circuit bootstrap and SNN, and belongs to the technical field of medical image analysis, an SNN model comprises a feature extraction layer, at least one spiking nerve unit and an output spiking nerve cell layer, and the spiking nerve unit comprises a first spiking nerve cell layer and a bootstrap circuit; the method comprises the following steps: performing pulse feature extraction on a received encrypted magnetic resonance image based on a feature extraction layer to obtain encrypted pulse features; performing pulse transmission on the encrypted pulse characteristics based on the first pulse neuron layer to obtain an encrypted pulse sequence; performing noise suppression on the encrypted pulse sequence based on a bootstrap circuit to obtain an optimized encrypted pulse sequence; and performing output pulse transmission on the optimized encrypted pulse sequence based on the output pulse neuron layer to obtain a brain tumor classification result. According to the invention, noise suppression is realized in the encryption domain by using the bootstrap circuit, the calculation complexity is reduced, the classification efficiency is improved, and the power consumption demand is reduced by using the SNN model.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a brain tumor classification method and system based on circuit bootstrapping and SNN. Background Art

[0002] Existing brain tumor classification methods primarily rely on deep learning techniques, such as convolutional neural networks (CNNs) and Transformer models, which require large amounts of labeled data for training. However, MRI data often contains sensitive patient information, making data sharing and model training difficult. Federated learning is considered a potential solution to addressing medical data privacy issues. This approach allows individual medical institutions to train models locally, sharing only model parameters, not the original data. However, existing federated learning frameworks still suffer from several key flaws. First, model parameters may contain implicit information about the original data. Through carefully designed parameter inference attacks, attackers can reconstruct key features of the training data. Second, federated learning typically relies on a centralized parameter server architecture, which not only introduces a single point of failure risk but also creates a new channel for privacy leakage. More importantly, the differences in MRI equipment parameters used by different medical institutions can lead to significant data distribution shift, making traditional federated learning frameworks often perform poorly in real-world medical scenarios. Differential privacy is another commonly used privacy-preserving approach that prevents privacy leakage by adding specific noise to the data or model parameters. However, its applicability to medical image classification tasks is severely limited. On the one hand, MRI images typically have fine anatomical features, and adding excessive noise can significantly degrade image quality, affecting doctors' diagnostic accuracy. On the other hand, deep learning models are highly sensitive to variations in input data, and even moderate addition of noise can lead to a significant drop in model performance.

[0003] Homomorphic encryption (HE), one of the core technologies in privacy-preserving computing, provides a theoretically perfect solution to the privacy protection problem in medical image analysis. This technology allows computations to be performed directly on encrypted data without prior decryption, fundamentally ensuring the confidentiality of the data during processing. However, in practical medical image analysis applications, especially for complex tasks such as MRI brain tumor classification, homomorphic encryption faces severe technical challenges. Specifically: First, while fully homomorphic encryption (FHE) theoretically supports computations of arbitrary complexity on encrypted data, it faces significant noise management challenges in practical applications. The noise in FHE schemes increases exponentially with increasing computational depth, a characteristic that severely limits its application in deep neural networks. Second, in homomorphic encryption environments, convolutional neural networks and Transformer models have high computational complexity and require significant computing resources.

[0004] Therefore, there is an urgent need to provide a brain tumor classification method and system based on circuit bootstrapping and SNN to reduce computing resources, reduce noise accumulation, and thus improve the efficiency of brain tumor classification in the encrypted domain. Summary of the Invention

[0005] In view of this, it is necessary to provide a brain tumor classification method and system based on circuit bootstrapping and SNN to solve the technical problems existing in the existing technology, such as excessive noise accumulation and high computing resources required, resulting in low efficiency of brain tumor classification in the encrypted domain.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a brain tumor classification method based on circuit bootstrapping and SNN, wherein the SNN model includes a feature extraction layer, at least one spiking neural unit, and an output spiking neuron layer, wherein the spiking neural unit includes a first spiking neuron layer and a bootstrapping circuit; The method comprises: Extracting pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features; Performing pulse transmission on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence; performing noise suppression on the encrypted pulse sequence based on the bootstrap circuit to obtain an optimized encrypted pulse sequence; The optimized encrypted pulse sequence is output pulse transmitted based on the output pulse neuron layer to obtain a brain tumor classification result.

[0007] In a possible implementation, the feature extraction layer includes an input layer and a convolutional layer; The input layer is used to perform pulse encoding on the received encrypted magnetic resonance image to obtain encrypted pulses; The convolution layer is used to perform homomorphic addition and homomorphic multiplication calculations on the encrypted pulse to obtain the encrypted pulse features.

[0008] In a possible implementation, the bootstrap circuit includes a plurality of logic selection gates, and the expressions of the logic selection gates are:

[0009] Where, is the selection gate operator; is the first candidate ciphertext; is the second candidate ciphertext; is the selection bit; Select a ciphertext for the destination.

[0010] In a possible implementation, before extracting pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features, the method further includes: Performing fully homomorphic encryption on the received magnetic resonance image based on a fully homomorphic encryption algorithm to obtain the encrypted magnetic resonance image; The fully homomorphic encryption algorithm is the LWE algorithm.

[0011] In a possible implementation, the input ciphertext format of the bootstrap circuit is RGSW format, and the spiking neural unit further includes a ciphertext structure conversion layer; The ciphertext structure conversion layer is used to perform ciphertext structure conversion on the encrypted magnetic resonance image in the LWE format to obtain the encrypted magnetic resonance image in the RGSW format.

[0012] In a possible implementation, the spiking neural unit further includes an average pooling layer and a fully connected layer sequentially connected to the bootstrap circuit; The average pooling layer is used to perform feature dimensionality reduction on the optimized encrypted pulse sequence to obtain a reduced-dimensional encrypted pulse sequence; The fully connected layer is used to map the dimension-reduced encrypted pulse sequence to an output space to obtain an output pulse sequence; The output pulse neuron layer is used to transmit the output pulse sequence to obtain the brain tumor classification result.

[0013] In a possible implementation, the spiking neural unit further includes a flattening layer connected between the average pooling layer and the fully connected layer; The flattening layer is used to integrate multiple time steps of the dimensionally reduced encrypted pulse sequence to obtain a one-dimensional vector; The fully connected layer is used to map the one-dimensional vector to the output space to obtain the output pulse sequence.

[0014] In a possible implementation, before performing fully homomorphic encryption on the received magnetic resonance image based on the fully homomorphic encryption algorithm to obtain the encrypted magnetic resonance image, the method further includes: acquiring an initial magnetic resonance image acquired by a magnetic resonance device; performing image preprocessing on the initial magnetic resonance image to obtain a preprocessed image; The pixel values ​​of the preprocessed image are format converted based on the format requirements of the fully homomorphic encryption algorithm to obtain the magnetic resonance image.

[0015] In a possible implementation, the image preprocessing includes grayscale normalization, denoising, compression, and edge detection.

[0016] In a second aspect, the present invention further provides a brain tumor classification system based on circuit bootstrapping and SNN, wherein the SNN model includes a feature extraction layer, at least one spiking neural unit, and an output spiking neuron layer, wherein the spiking neural unit includes a first spiking neuron layer and a bootstrapping circuit; The system comprises: a feature extraction module, configured to extract pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features; a pulse transmission module, configured to perform pulse transmission on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence; A bootstrap processing module, configured to perform noise suppression on the encrypted pulse sequence based on the bootstrap circuit to obtain an optimized encrypted pulse sequence; A classification module is used to perform output pulse transmission on the optimized encrypted pulse sequence based on the output pulse neuron layer to obtain a brain tumor classification result.

[0017] The beneficial effects of the present invention are as follows: the brain tumor classification method based on circuit bootstrapping and SNN provided by the present invention implements brain tumor classification of encrypted magnetic resonance images through a spiking neural network (SNN) model. The SNN model triggers action potentials (pulses) by accumulating the membrane potential generated by input pulses. When the membrane potential exceeds the threshold, the neuron emits a pulse, then enters a refractory period and resets the potential. That is, the neuron only updates its state when receiving an input pulse or emitting a pulse, avoiding the layer-by-layer synchronous calculation of traditional convolutional neural networks and Transformers. Compared with traditional convolutional neural networks and Transformers, it has lower power consumption. That is, the computing requirements in a homomorphic encryption environment are significantly reduced, thereby reducing computing resources and improving the efficiency of brain tumor classification in the encrypted domain. In addition, the SNN model combines temporal coding in the brain tumor classification process, achieving efficient reasoning of encrypted magnetic resonance images, improving classification efficiency and accuracy.

[0018] Furthermore, in order to address the problem of noise accumulation during the homomorphic encryption process, the present invention sets a bootstrap circuit to suppress the noise of the encrypted pulse sequence obtained by the corresponding first pulse neuron layer, which can effectively reduce noise accumulation and further reduce the computational complexity, thereby further improving the efficiency of brain tumor classification in the encryption domain.

[0019] Furthermore, the SNN model of the present invention processes encrypted magnetic resonance images, thereby preventing the privacy information in the magnetic resonance images from being leaked during the classification process, protecting the patient's data privacy, and improving the security of magnetic resonance image processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A schematic diagram of the structure of an embodiment of the SNN model provided by the present invention; Figure 2 A schematic flow chart of an embodiment of a brain tumor classification method based on circuit bootstrapping and SNN provided by the present invention; Figure 3 A schematic diagram of the structure of an embodiment of the logic selection gate provided by the present invention; Figure 4 A schematic diagram of a flow chart of an embodiment of preprocessing and quantifying an initial magnetic resonance image provided by the present invention; Figure 5This is a schematic structural diagram of an embodiment of a brain tumor classification system based on circuit bootstrapping and SNN provided by the present invention. DETAILED DESCRIPTION

[0022] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0023] It should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the present invention illustrate operations implemented according to some embodiments of the present invention. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps that have no logical contextual relationship can be reversed in order or implemented simultaneously. In addition, those skilled in the art, guided by the content of the present invention, can add one or more other operations to the flowcharts or remove one or more operations from the flowcharts. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and / or processor systems and / or microcontroller systems.

[0024] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0025] The present invention provides a brain tumor classification method and system based on circuit bootstrapping and SNN, which are described below.

[0026] The implementation of the brain tumor classification method based on circuit bootstrapping and SNN in the embodiment of the present invention is based on the SNN model. Figure 1 A schematic diagram of the structure of an embodiment of the SNN model provided by the present invention is shown as follows: Figure 1 As shown, the SNN model includes a feature extraction layer, n spiking neural units (n ≥ 1), and an output spiking neuron layer, and the spiking neural unit includes a first spiking neuron layer and a bootstrap circuit; Figure 2 A flow chart of an embodiment of the brain tumor classification method based on circuit bootstrapping and SNN provided by the present invention is shown as follows: Figure 2As shown, the brain tumor classification method based on circuit bootstrapping and SNN includes: S201 . Perform pulse feature extraction on the received encrypted magnetic resonance image based on a feature extraction layer to obtain encrypted pulse features.

[0027] Among them, encrypted magnetic resonance images refer to magnetic resonance images encrypted based on a fully homomorphic encryption algorithm to ensure that patient privacy is not leaked during the brain tumor classification process and to ensure data privacy security.

[0028] S202 : Perform pulse transmission on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence.

[0029] The first spiking neuron layer operates by accumulating the membrane potential generated by input pulses, triggering action potentials (pulses). Specifically, the neuron fires a pulse only when the membrane potential exceeds a threshold, then enters a refractory period and resets its potential. This eliminates the layer-by-layer synchronous computation required in traditional neural networks, significantly reducing power consumption.

[0030] S203 , performing noise suppression on the encrypted pulse sequence based on the bootstrap circuit to obtain an optimized encrypted pulse sequence.

[0031] Among them, the bootstrap circuit achieves noise reset through a more underlying circuit structure, so that the noise level is reset to the initial state, greatly reducing the amount of noise transmitted to the subsequent stage and achieving noise suppression.

[0032] S204 , performing output pulse transmission on the optimized encrypted pulse sequence based on the output pulse neuron layer to obtain a brain tumor classification result.

[0033] Among them, the working principle of the output pulse neuron layer is the same as that of the first pulse neuron layer, which will not be described here.

[0034] It should be noted that the triggering action of the output pulse neuron layer and the first pulse neuron layer can be expressed as:

[0035] The principle of the above formula is as follows: When the membrane potential Exceeding the issuance threshold A trigger pulse is generated when the membrane potential Not exceeding the issuance threshold , no trigger pulse is generated.

[0036] It should be understood that the brain tumor classification method based on circuit bootstrapping and SNN in the embodiments of the present invention can be implemented in any device based on this method, such as a medical image processing device. Specifically, the brain tumor classification method based on circuit bootstrapping and SNN is stored in the aforementioned device as a pre-programmed program. When the device is powered on, the program is invoked and the brain tumor classification method based on circuit bootstrapping and SNN is implemented.

[0037] Compared to existing technologies, the brain tumor classification method based on circuit bootstrapping and SNN provided in embodiments of the present invention implements brain tumor classification in encrypted MRI images through a spiking neural network (SNN) model. The SNN model triggers action potentials (pulses) by accumulating the membrane potential generated by input pulses. When the membrane potential exceeds a threshold, the neuron fires a pulse, then enters a refractory period and resets its potential. Specifically, the neuron updates its state only when receiving an input pulse or firing a pulse, avoiding the layer-by-layer synchronous computation required by traditional convolutional neural networks and transformers. Compared to traditional convolutional neural networks and transformers, this method offers lower power consumption, significantly reducing computational requirements in a homomorphic encryption environment. This, in turn, reduces computing resources and improves brain tumor classification efficiency within the encrypted domain. Furthermore, the SNN model incorporates temporal coding during brain tumor classification, enabling efficient inference of encrypted MRI images, improving both classification efficiency and accuracy.

[0038] Furthermore, in order to address the problem of noise accumulation during the homomorphic encryption process, the present invention sets a bootstrap circuit to suppress the noise of the encrypted pulse sequence obtained by the corresponding first pulse neuron layer, which can effectively reduce noise accumulation and further reduce the computational complexity, thereby further improving the efficiency of brain tumor classification in the encryption domain.

[0039] Furthermore, the SNN model of the embodiment of the present invention processes encrypted magnetic resonance images, thereby preventing the privacy information in the magnetic resonance images from being leaked during the classification process, protecting the patient's data privacy, and improving the security of magnetic resonance image processing.

[0040] Since the SNN model processes pulse signals, and the encrypted magnetic resonance image is image data, in some embodiments of the present invention, such as Figure 1 As shown, the feature extraction layer includes the input layer and the convolution layer; The input layer is used to perform pulse encoding on the received encrypted magnetic resonance image to obtain encrypted pulses; The convolutional layer is used to perform homomorphic addition and homomorphic multiplication on the encrypted pulse to obtain the encrypted pulse features.

[0041] In a specific embodiment of the present invention, in order to preserve the temporal representation of the encrypted magnetic resonance image, the pulse encoding method is time encoding.

[0042] In a specific embodiment of the present invention, the bootstrap circuit includes a plurality of logic selection gates, and the structure of the logic selection gates is as follows: Figure 3 As shown by Figure 3 It can be seen that the expression of the logic selection gate is:

[0043] Where, is the selection gate operator; is the first candidate ciphertext; is the second candidate ciphertext; is the selection bit; Select a ciphertext for the destination.

[0044] The embodiment of the present invention achieves noise reset of ciphertext in the encrypted magnetic resonance image recognition process at the circuit level based on a bootstrap circuit, thereby reducing the amount of noise and ensuring the classification efficiency of brain tumor classification.

[0045] In some embodiments of the present invention, before step S201, the brain tumor classification method based on circuit bootstrapping and SNN further includes: Performing fully homomorphic encryption on the received magnetic resonance image based on a fully homomorphic encryption algorithm to obtain an encrypted magnetic resonance image; The fully homomorphic encryption algorithm is the LWE (Learning with Errors) algorithm.

[0046] The format of the encrypted magnetic resonance image after encryption by the fully homomorphic encryption algorithm is LWE format, while the input ciphertext format required by the bootstrap circuit is RSGW (Gentry Sahai Water's ring version) format. Therefore, in order to achieve successful bootstrapping of the bootstrap circuit, in some embodiments of the present invention, such as Figure 1 As shown, the spiking neural unit also includes a ciphertext structure conversion layer; The ciphertext structure conversion layer is used to convert the encrypted magnetic resonance image in the LWE format into an encrypted magnetic resonance image in the RGSW format.

[0047] The embodiment of the present invention converts the format of the encrypted magnetic resonance image into a format compatible with the bootstrap circuit through the ciphertext structure conversion layer, thereby ensuring the operation success rate of the bootstrap circuit and further ensuring the bootstrap success rate, that is, improving the success rate of noise suppression.

[0048] It should be noted that the conversion process from LWE ciphertext to RGSW ciphertext is a conventional conversion process. Specifically, the LWE ciphertext is first converted to RLWE ciphertext, and then the RLEW ciphertext is converted to RGSW ciphertext.

[0049] The hierarchical homomorphic computation mode for LWE ciphertext is level 2, while that for RGSW is level 1. This achieves a level reduction during the ciphertext format conversion process. This level reduction translates to a smaller key size and lower computational overhead, further improving the computational efficiency of the SNN model and, consequently, the efficiency of brain tumor classification. In other words, circuit bootstrapping is achieved during the conversion from LWE ciphertext to RGSW ciphertext, further suppressing noise accumulation and improving brain tumor classification efficiency. Furthermore, during the conversion from LWE ciphertext to RGSW ciphertext, the blind rotation process in the function bootstrapping can be replaced by the outer product operation between the RLWE ciphertext and the RGSW ciphertext, reducing the computational effort and further improving the efficiency of the bootstrapping.

[0050] In summary, the embodiment of the present invention can achieve a faster and smaller bootstrapping process based on the more compact ciphertext structure of RGSW by setting a conversion process of converting LWE ciphertext into RGSW ciphertext.

[0051] To further improve the classification efficiency of brain tumors, in some embodiments of the present invention, Figure 1 As shown, the spiking neural unit also includes an average pooling layer and a fully connected layer sequentially connected to the bootstrap circuit; The average pooling layer is used to perform feature dimensionality reduction on the optimized encrypted pulse sequence to obtain a reduced-dimensional encrypted pulse sequence.

[0052] Among them, the average pooling layer further reduces the computational complexity by performing feature dimensionality reduction on the optimized encrypted pulse sequence, that is, further improves the computational efficiency.

[0053] The fully connected layer is used to map the dimension-reduced encrypted pulse sequence to the output space to obtain the output pulse sequence.

[0054] Each neuron in the fully connected layer is connected to all neurons in the previous layer, forming a dense connection. Its function is to integrate local features into global information and generate a global representation related to classification.

[0055] The output pulse neuron layer is used to transmit the output pulse sequence to obtain the brain tumor classification result.

[0056] Since the dimension-reduced encrypted pulse sequence output by the average pooling layer is multi-dimensional data, in order to facilitate the subsequent processing of the fully connected layer, in some embodiments of the present invention, such as Figure 1 As shown, the spiking neural unit also includes a flattening layer connected between the average pooling layer and the fully connected layer; The flattening layer is used to integrate multiple time steps of the dimensionally reduced encrypted pulse sequence to obtain a one-dimensional vector; The fully connected layer is used to map the one-dimensional vector to the output space to obtain the output pulse sequence.

[0057] The embodiment of the present invention realizes the conversion of multi-dimensional data into one-dimensional vectors by setting a flattening layer, which facilitates data processing of the fully connected layer.

[0058] In some embodiments of the present invention, Figure 4 As shown, before performing fully homomorphic encryption on the received magnetic resonance image based on the fully homomorphic encryption algorithm to obtain the encrypted magnetic resonance image, the brain tumor classification method based on circuit bootstrapping and SNN further includes: S401: Acquire an initial magnetic resonance image acquired by a magnetic resonance device.

[0059] The acquisition method of step S401 can be: acquiring an initial magnetic resonance image acquired in real time by a magnetic resonance device, or: acquiring an initial magnetic resonance image acquired historically by a magnetic resonance device, that is, the embodiment of the present invention can be a real-time classification or a delayed classification of brain tumors.

[0060] S402, performing image preprocessing on the initial magnetic resonance image to obtain a preprocessed image; S403. Perform format conversion on the pixel values ​​of the preprocessed image based on the format requirements of the fully homomorphic encryption algorithm to obtain a magnetic resonance image.

[0061] Specifically, the format requirement of the fully homomorphic encryption algorithm is: the pixel value is in 8-bit integer format. Based on this format requirement, the pixel value of the preprocessed image is quantized to obtain a magnetic resonance image in 8-bit integer format.

[0062] By preprocessing the initial magnetic resonance images, the embodiments of the present invention can improve the quality of the resulting preprocessed images, thereby increasing the reliability and accuracy of brain tumor classification results. Furthermore, by converting the pixel values ​​of the preprocessed images based on the format requirements of the fully homomorphic encryption algorithm, the obtained magnetic resonance images can be directly encrypted using the fully homomorphic encryption algorithm, ensuring encryption accuracy and further ensuring the accuracy of brain tumor classification results.

[0063] In a specific embodiment of the present invention, image preprocessing includes grayscale normalization processing, denoising processing, compression processing and edge detection.

[0064] The grayscale normalization process specifically involves scaling the pixel values ​​of the initial magnetic resonance image to [0, 1] to reduce the influence of the value range.

[0065] The denoising process specifically includes: using a filtering algorithm such as Gaussian filtering or median filtering to filter the initial magnetic resonance image after grayscale normalization processing to remove image noise.

[0066] The compression processing specifically involves adjusting the image size to reduce the amount of calculation, thereby further improving the efficiency of brain tumor classification.

[0067] In a specific embodiment of the present invention, the size of the compressed initial magnetic resonance image is 128×128.

[0068] The edge detection is specifically: extracting the contour information of the compressed initial magnetic resonance image based on methods such as the Sobel operator.

[0069] In summary, the brain tumor classification method based on circuit bootstrapping and SNN proposed in the embodiments of the present invention utilizes a spiking neural network (SNN) model to reduce the power consumption required for neural network computation, significantly lowering the computational requirements compared to traditional CNNs in a homomorphic encryption environment. Furthermore, the bootstrapping circuit suppresses noise accumulation between spiking neuron layers, and the ciphertext structure conversion based on LWE and RGSW enables bootstrapping of the spiking neuron layer, effectively suppressing noise accumulation within the spiking neuron layer. The combination of these two noise suppression methods significantly reduces noise accumulation, lowers computational complexity, and improves the efficiency of brain tumor classification. Furthermore, the present invention combines SNNs with circuit bootstrapping to achieve privacy-preserving classification of sensitive magnetic resonance images, enabling intelligent diagnosis without decryption and ensuring privacy security during the classification process. In summary, the embodiments of the present invention achieve efficient and secure computation in a homomorphic encryption environment, enabling brain tumor classification without data decryption. It also optimizes encryption computation efficiency, reduces noise accumulation, and improves SNN computational performance within the encrypted domain. This invention overcomes the technical bottlenecks of low computational efficiency and high power consumption of existing fully homomorphic encryption methods, providing a more secure and efficient solution for privacy-preserving computation.

[0070] In order to better implement the brain tumor classification method based on circuit bootstrapping and SNN in the embodiment of the present invention, based on the brain tumor classification method based on circuit bootstrapping and SNN, the embodiment of the present invention also provides a brain tumor classification system based on circuit bootstrapping and SNN, wherein the SNN model includes a feature extraction layer, at least one pulse neural unit and an output pulse neuron layer, wherein the pulse neural unit includes a first pulse neuron layer and a bootstrap circuit; Figure 5 As shown, the brain tumor classification system 500 based on circuit bootstrapping and SNN includes: A feature extraction module 501 is configured to extract pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features; a pulse transfer module 502 for performing pulse transfer on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence; A bootstrap processing module 503 is used to perform noise suppression on the encrypted pulse sequence based on a bootstrap circuit to obtain an optimized encrypted pulse sequence; The classification module 504 is configured to perform output pulse transmission on the optimized encrypted pulse sequence based on the output pulse neuron layer to obtain a brain tumor classification result.

[0071] The brain tumor classification system 500 based on circuit bootstrapping and SNN provided in the above embodiment can implement the technical solution described in the above embodiment of the brain tumor classification method based on circuit bootstrapping and SNN. The specific implementation principles of the above modules or units can be found in the corresponding contents in the above embodiment of the brain tumor classification method based on circuit bootstrapping and SNN, which will not be repeated here.

[0072] Those skilled in the art will appreciate that all or part of the process flow of the above-described method embodiment can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program, and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0073] The above is a detailed introduction to the brain tumor classification method and system based on circuit bootstrapping and SNN provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A brain tumor classification method based on circuit bootstrapping and SNN, characterized in that: The SNN model includes a feature extraction layer, at least one spiking neural unit, and an output spiking neural layer, wherein the spiking neural unit includes a first spiking neural layer and a bootstrap circuit; The method comprises: Extracting pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features; Performing pulse transmission on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence; performing noise suppression on the encrypted pulse sequence based on the bootstrap circuit to obtain an optimized encrypted pulse sequence; The optimized encrypted pulse sequence is output pulse transmitted based on the output pulse neuron layer to obtain a brain tumor classification result.

2. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 1, characterized in that The feature extraction layer includes an input layer and a convolutional layer; The input layer is used to perform pulse encoding on the received encrypted magnetic resonance image to obtain encrypted pulses; The convolution layer is used to perform homomorphic addition and homomorphic multiplication calculations on the encrypted pulse to obtain the encrypted pulse features.

3. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 1, characterized in that The bootstrap circuit includes a plurality of logic selection gates, and the expression of the logic selection gate is: Where, is the selection gate operator; is the first candidate ciphertext; is the second candidate ciphertext; is the selection bit; Select a ciphertext for the destination.

4. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 1, characterized in that Before extracting pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features, the method further includes: Performing fully homomorphic encryption on the received magnetic resonance image based on a fully homomorphic encryption algorithm to obtain the encrypted magnetic resonance image; The fully homomorphic encryption algorithm is the LWE algorithm.

5. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 4, characterized in that: The input ciphertext format of the bootstrap circuit is RGSW format, and the pulse neural unit further includes a ciphertext structure conversion layer; The ciphertext structure conversion layer is used to perform ciphertext structure conversion on the encrypted magnetic resonance image in the LWE format to obtain the encrypted magnetic resonance image in the RGSW format.

6. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 1, characterized in that The spiking neural unit further includes an average pooling layer and a fully connected layer sequentially connected to the bootstrap circuit; The average pooling layer is used to perform feature dimensionality reduction on the optimized encrypted pulse sequence to obtain a reduced-dimensional encrypted pulse sequence; The fully connected layer is used to map the dimension-reduced encrypted pulse sequence to an output space to obtain an output pulse sequence; The output pulse neuron layer is used to transmit the output pulse sequence to obtain the brain tumor classification result.

7. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 6, characterized in that: The spiking neural unit further includes a flattening layer connected between the average pooling layer and the fully connected layer; The flattening layer is used to integrate multiple time steps of the dimensionally reduced encrypted pulse sequence to obtain a one-dimensional vector; The fully connected layer is used to map the one-dimensional vector to the output space to obtain the output pulse sequence.

8. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 4, characterized in that: Before performing fully homomorphic encryption on the received magnetic resonance image based on the fully homomorphic encryption algorithm to obtain the encrypted magnetic resonance image, the method further includes: acquiring an initial magnetic resonance image acquired by a magnetic resonance device; performing image preprocessing on the initial magnetic resonance image to obtain a preprocessed image; The pixel values ​​of the preprocessed image are format converted based on the format requirements of the fully homomorphic encryption algorithm to obtain the magnetic resonance image.

9. The brain tumor classification method based on circuit bootstrapping and SNN according to claim 8, characterized in that: The image preprocessing includes grayscale normalization processing, denoising processing, compression processing and edge detection.

10. A brain tumor classification system based on circuit bootstrapping and SNN, characterized in that: The SNN model includes a feature extraction layer, at least one spiking neural unit, and an output spiking neural layer, wherein the spiking neural unit includes a first spiking neural layer and a bootstrap circuit; The system comprises: a feature extraction module, configured to extract pulse features from the received encrypted magnetic resonance image based on the feature extraction layer to obtain encrypted pulse features; a pulse transmission module, configured to perform pulse transmission on the encrypted pulse feature based on the first pulse neuron layer to obtain an encrypted pulse sequence; A bootstrap processing module, configured to perform noise suppression on the encrypted pulse sequence based on the bootstrap circuit to obtain an optimized encrypted pulse sequence; A classification module is used to perform output pulse transmission on the optimized encrypted pulse sequence based on the output pulse neuron layer to obtain a brain tumor classification result.

Images