Unified method for generating pseudo-CT image based on multi-sequence and multi-anatomical-site MRI image and storage medium

Abstract

The invention relates to a unified method for generating a pseudo-CT image based on multi-sequence and multi-anatomical-site MRI images and a storage medium. The method comprises the following steps: allocating prompt codes special for different sequences of MRI images of a plurality of different human body parts to the sequences of the parts, forming training data by the prompt codes, the corresponding MRI images and corresponding real CT images, training a unified model through the training data, adding a part-specific attention module RSAM in the unified model, the MRI image of the human body part is input into the trained unified model, a pseudo-CT image is generated, and the MRI image is distributed with a prompt code special for the human body part corresponding to the MRI image. According to the method, pseudo-CT images generated by MRI images of different parts of a human body are more accurate.

Description

Technical Field

[0001] The present invention relates to the field of medical image conversion, in particular to a method, storage medium and device for generating pseudo CT images based on multi-sequence and multi-anatomical part MRI images, and is applicable to a unified generation model of multi-anatomical parts and multi-MRI sequences. Background Art

[0002] Existing MRI-to-pseudo-CT generation models work well when processing specific anatomical sites or a single MRI sequence, but the generation effect is often poor when inputting MRI images from multiple anatomical sites and different sequences. This is because the characteristics of different sites and sequences vary significantly and are difficult to be fully processed in a single model. Therefore, a unified model that can simultaneously process multiple sequences and multiple anatomical sites of MRI is crucial to improve the accuracy and consistency of generated pseudo-CT images. Summary of the invention

[0003] Based on this, a unified method for generating pseudo CT images based on multi-sequence and multi-anatomical MRI images is provided. This method can simultaneously process MRI images from multiple parts and multiple sequences to ensure the accuracy and consistency of the generated pseudo CT images. By mixing the input of MRI data from different parts and sequences, the model can learn the characteristics of different anatomical parts and MRI sequences, and then achieve unified generation.

[0004] A unified method for generating pseudo CT images based on multi-sequence, multi-anatomical MRI images, comprising:

[0005] Assign prompt codes specific to the sequence of MRI images of multiple different human body parts, and form training data with the prompt codes, corresponding MRI images, and corresponding real CT images.

[0006] The unified model is trained by training data, and the part-specific attention module RSAM is added to the unified model.

[0007] An MRI image of a human body part is input into a trained unified model to generate a pseudo CT image, wherein the MRI image is assigned a prompt code dedicated to the human body part corresponding to the MRI image.

[0008] In one embodiment,

[0009] The part-specific attention module RSAM is used to receive input features from the residual block and adjust the weights of these features according to the hint code.

[0010] In one embodiment,

[0011] The total loss function of the unified model is:

[0012]

[0013] Among them, α GAN =1.0,α cycle =10.0,α identity =5.0,α perceptual =0.01, α adv =1.0.

[0014] In one embodiment,

[0015]

[0016] in represents the synthetic CT image generated by the generator, represents the input MRI image, i is used to distinguish images of different body parts,

[0017] represents the real CT image, D CT is the discriminator, the expected value and They represent the expected values ​​of the generated samples and the real samples respectively,

[0018]

[0019] in, represents the input MRI image, represents the real CT image,

[0020] G MR→CT represents the generator mapping from MRI images to synthetic CT images,

[0021] G CT→MR represents the inverse mapping from CT image to MRI image,

[0022]

[0023] in, represents the input MRI image, G CT→MR represents the inverse mapping from CT image to MRI image, represents the real CT image, G MR→CT represents the generator mapping from MRI images to synthetic CT images,

[0024]

[0025] in, represents the features extracted by the MAE model at the kth scale, K is the number of scales, G MR→CT represents the generator mapping from MRI images to synthetic CT images, represents the real CT image, and ∥·∥2 is the L2 norm, which is used to calculate the difference between feature vectors.

[0026] In one embodiment,

[0027] The invention also includes a protection method for the unified model after training, and the protection method includes:

[0028] By means of the authorized device identifier, the authorized region information and the encryption and decryption mechanism of the unified model, it is ensured that the unified model only runs on the authorized device and the authorized region.

[0029] In one of the embodiments, the authorized area information includes jointly encoding geographic coordinates and text information. When defining the text information, geographic grid division is combined to enhance regional awareness. Specifically, the geographic area is first divided into multiple grids, and then the grid IDs of these grids are used as auxiliary information. Then, the specific hospital and department descriptions are combined to generate a regional awareness embedding.

[0030] In one embodiment,

[0031] The protection method specifically includes:

[0032] Match the authorized device identifier and authorized region information with the unified model,

[0033] Encrypt the parameters within the unified model,

[0034] The internal parameters of the unified model are decrypted only when the device identifier of the device actually using the unified model is consistent with the authorized device identifier and the region information of the actual use region is consistent with the authorized region information.

[0035] In one embodiment,

[0036] The authorized device identifier is generated by the following method:

[0037] The hardware information of the authorized device is processed through a hash algorithm to generate an authorized device identifier.

[0038] Authorized region information is generated by the following method:

[0039] The information of the authorized use area of ​​the unified model is processed by a hash algorithm to generate the authorized area information.

[0040] In one embodiment,

[0041] Encrypt the parameters in the unified model, including:

[0042] Based on the authorized device identifier and authorized region information, an encryption matrix with the same dimension as the internal parameters of the unified model is generated by using a pseudo-random generation algorithm.

[0043] Internal parameters of the unified model are encrypted based on the encryption matrix.

[0044] A computer storage medium stores at least one executable instruction, which enables a processor to execute the operation of the unified method for generating pseudo CT images based on multi-sequence and multi-anatomical site MRI images.

[0045] A computer device includes: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus, and the memory is used to store at least one executable instruction, wherein the executable instruction enables the processor to execute operations corresponding to the unified method for generating pseudo CT images based on multi-sequence and multi-anatomical site MRI images.

[0046] The beneficial effects of this application are:

[0047] 1. The method of the present application can simultaneously process different sequence MRI slice images from multiple different human body parts. This capability enables the model to simultaneously consider the characteristics of different human body parts, so that the corresponding pseudo CT images can be generated more comprehensively and accurately.

[0048] 2. The unified model of this application performs mixed training on data from multiple human body parts, captures and adapts to the feature differences of different anatomical parts in the same architecture, and then generates high-precision pseudo CT images. Since the MRI data of different anatomical parts of the human body are processed uniformly in the model, the model can learn the commonalities and specific features of features across parts, thereby improving the accuracy and consistency of pseudo CT image generation. The flexibility of the unified model enables it to adapt to MRI input sequences of any anatomical part, realize unified processing of multi-part and multi-sequence data, and ensure that the generated pseudo CT images have excellent quality in different parts and sequences.

[0049] 3. The method of the present application aims to overcome the limitations of the prior art in processing information of multiple different human body parts. By jointly processing MRI slice images of multiple different human body parts, the method of the present application can better capture the subtle differences between different human body parts and improve the accuracy and detail performance of the generated pseudo-CT images in various parts.

[0050] 4. The method of this application introduces an attention mechanism for multiple human body parts, which can dynamically adjust the focus of attention according to the characteristics and structures of different human body parts. This attention mechanism specific to each human body part helps to generate high-quality pseudo CT images of the corresponding parts more accurately.

[0051] 5. The method of the present application is not limited to processing MRI images of multiple human body parts, but can also process multi-modal data input. In addition to processing MRI slice images of different human body parts, the method of the present application also has the ability to process different MRI imaging modalities (such as T1 weighted, T2 weighted, etc.), which makes the model more versatile and flexible and can adapt to different types of medical image data.

[0052] 6. It has cross-domain adaptability, that is, it can process medical image data generated by different sources, different brands or different devices. This ability makes it suitable for various data distribution situations, thereby improving its versatility in actual clinical applications.

[0053] 7. The method of the present application provides stronger interpretability and visualization capabilities, that is, while generating pseudo CT images, it can also output interpretable information about the generated images. This includes the basis on which the model generates images, the location of the corresponding human body parts, the key features that the model focuses on, etc. This capability is very important for medical professionals because they need to understand how the model makes inferences and generates images, thereby enhancing trust in the model results and helping to make more accurate diagnosis and treatment decisions.

[0054] 8. The protection method of this application is conducive to the reasonable use of the unified model in authorized devices and authorized areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The present invention is a flowchart of a unified method for generating pseudo CT images based on multi-sequence and multi-anatomical site MRI images according to an embodiment of the present application.

[0056] Figure 2 This is a structural diagram of the model used in the unified method for generating pseudo CT images based on multi-sequence and multi-anatomical site MRI images according to an embodiment of the present application.

[0057] Figure 3 A comparison chart of pseudo CT images generated by the unified method of generating pseudo CT images based on multi-sequence and multi-anatomical MRI images of the present application and pseudo CT images generated by other existing methods. The second column is the real CT image, and the sixth column is the pseudo CT image generated by the method of the present application. DETAILED DESCRIPTION

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0059] like Figure 1 and Figure 2As shown, the first embodiment of the present application provides a unified method for generating pseudo CT images based on multi-sequence and multi-anatomical site MRI images, the method comprising:

[0060] Assign prompt codes dedicated to the sequence of MRI images of different parts of the human body to the parts, and form training data with prompt codes, corresponding MRI images and corresponding real CT images. For example, the first sequence of MRI images of the human brain is assigned with a dedicated prompt code 1, the second sequence of MRI images of the human brain is assigned with a dedicated prompt code 2, the first sequence of MRI images of the human abdomen is assigned with a dedicated prompt code 3, and the second sequence of MRI images of the human abdomen is assigned with a dedicated prompt code 4. The different sequences include but are not limited to the following sequences: T1-weighted Imaging (T1WI), T2-weighted Imaging (T2WI), FLAIR (Fluid-Attenuated Inversion Recovery), Diffusion-Weighted Imaging (DWI), etc.

[0061] The unified model is trained by training data, and a part-specific attention module RSAM is added to the unified model. Specifically, the unified model adopts a CycleGAN network, and the generator of the CycleGAN network adopts a ResNet architecture. An additional self-attention module is added to one of the residual blocks of ResNet to implement a self-attention mechanism to constitute the above-mentioned part-specific attention module RSAM.

[0062] An MRI image of a human body part is input into a trained model to generate a pseudo CT image, wherein the MRI image is assigned a prompt code specific to the human body part corresponding to the MRI image.

[0063] It should be noted that the MRI image may be three MRI slice images formed by continuous slicing of a three-dimensional CT image.

[0064] It should be noted that the human body parts refer to the brain, chest, abdomen, pelvic area, arms, legs, etc. of the human body.

[0065] It should be noted that the above prompt code can be composed of numbers or vectors, etc.

[0066] In one of the embodiments, the part-specific attention module RSAM is used to receive input features from the residual block and adjust the weights of these features according to the hint code.

[0067] Specifically, Figure 2As shown, the specific operation process of the site-specific attention module RSAM of the present application is as follows.

[0068] From input features to output features, the part-specific attention module RSAM completes feature processing through the following steps: The part-specific attention module RSAM is referred to as RSAM module below.

[0069] 1. Input features: The input features first enter the RSAM module. These features are usually preliminary feature maps after convolution, normalization and activation operations.

[0070] 2. Prompt component:

[0071] 1) Prompt code reception: The prompt component of the RSAM module receives the prompt code, which can use a one-dimensional one-hot encoding to represent a specific human body part. For example, if there are N different human body parts, the prompt vector is an N-dimensional one-hot vector, in which only one element is 1 and the rest are 0.

[0072] 2) Multi-layer perceptron processing:

[0073] Feature vector generation: The prompt code first performs nonlinear mapping through a multi-layer perceptron (MLP1) to generate a feature vector that matches the input feature space. This feature vector carries the contextual information of a specific area and can help the model identify which part of the human body the current feature belongs to.

[0074] Adjustment coefficient generation: The prompt code also generates an adjustment coefficient γ through another multi-layer perceptron (MLP2). This coefficient controls the intensity of the output feature map so that the features of different parts of the human body can show different significance in the output.

[0075] 3. Self-attention mechanism:

[0076] 1) Generation of query, key, and value: The input features are respectively generated into query, key, and value feature maps through three 1×1 convolutional layers. The channel dimensions of query and key features are compressed to 1 / 8 of the original to reduce computational overhead, while the value features retain the original channel dimensions to ensure that they contain rich feature information.

[0077] 2) Fusion of feature vector and query feature:

[0078] ■Feature vector expansion: The generated feature vector is expanded to match the channel dimension of the query feature.

[0079] ■ Element-wise weighting: The feature vector is multiplied element-wise by the query feature to generate a weighted query matrix. This operation is controlled by a learnable weight parameter α to dynamically adjust the influence of the feature vector in the query. The initial value of is small (such as 0.1) to ensure that regional cues do not overly dominate attention allocation during the initial training of the model.

[0080] 4. Attention map calculation:

[0081] 1) Dot product operation: The weighted query feature matrix is ​​dot-producted with the key feature matrix to generate the attention weight matrix.

[0082] 2) Normalization: The attention weight matrix is ​​normalized by the Softmax function to obtain an attention map. This attention map represents the correlation distribution of each feature position, helping the model focus on the important features in the current area.

[0083] 5. Output feature generation:

[0084] 1) Value feature weighting: The attention map is multiplied by the value feature matrix to generate a weighted output feature map. At this point, the output feature map has incorporated region-specific contextual information, making it anatomically consistent with the target region.

[0085] 2) Adjustment coefficient scaling: The output feature map is scaled by the adjustment coefficient γ to control the output intensity and ensure the significance of regional features.

[0086] 3) Residual connection: Finally, the scaled output feature map is added to the input feature to obtain the final output feature, realizing the residual connection of the feature. This operation maintains the integrity of the original feature information while enhancing the guiding effect of the regional prompt.

[0087] It should be noted that in the method of the present application, the network after adding the RSAM module allows the simultaneous input of MRI images of different parts of the human body, and is adjusted through a parameterized attention mechanism to accurately generate high-quality pseudo CT images of the corresponding parts. In addition, the RSAM module also supports the input of 3D MRI, and can arbitrarily select slice images of continuous layers as input, and output corresponding pseudo CT slice images.

[0088] The key point of the RSAM model of this application is the introduction of a self-Attention mechanism specific to each body part. First, ensure that the MRI data of each different body part is assigned a unique prompt code, which can be an integer or a vector representing the identifier of each body part. In the input stage, MRI data of multiple different body parts are provided to the model. The MRI data of each body part undergoes the same self-attention mechanism, but through the different embedding of the prompt code, the focus of attention of each body part will be dynamically adjusted. This mechanism adopts a parametric design, which allows the focus of attention to be dynamically adjusted through the target prompt code, which is also the core idea of ​​the RSAM module.

[0089] Specifically, in order to better adapt to the characteristics of different parts when processing MRI slice image inputs of multiple human parts, and to improve the quality and accuracy of generated pseudo CT images, this application adopts a method combining prompt codes and self-attention modules. First, a unique prompt code is assigned to each MRI slice of a different human body part, and then these prompt codes are embedded in the model. Next, learnable parameters are introduced, which are connected to the prompt code to dynamically adjust the focus of attention. In the self-attention module, the attention score is calculated by calculating the similarity between Query (Q) and Key (K), where the learnable parameters affect the calculation process of the similarity. Subsequently, an activation function (such as softmax) is applied to obtain a dynamically adjusted attention weight, which is used to weight Value (V) to generate the final output feature representation.

[0090] Different cue codes lead to different self-attention weights, allowing the model to pay more attention to the features of the part corresponding to the cue code. This set of methods combining cue codes and self-attention modules enables the model to achieve dynamic attention allocation to adapt to the characteristics and structures of different parts of the human body, thereby improving the quality and accuracy of generated pseudo CT images. In this process, the key is to introduce information related to the cue code into the self-attention module and learn how to effectively use this information to adjust attention during model training.

[0091] In addition, in the joint training strategy, the RSAM module of the present application is designed to be able to learn the features of multiple human body parts at the same time. This means that the model can process MRI slice images of multiple human body parts, each of which has its own unique characteristics. In joint training, the model not only focuses on learning the features of a single human body part, but also considers the features of multiple human body parts at the same time to improve the overall performance of the model. The idea of ​​parameter sharing and specialization enhances the adaptability of the model when dealing with multiple organs. Specifically, some parameters are shared between multiple human body parts, which allows the model to learn the commonalities between different human body parts. At the same time, there are also some parameters that are specific to each human body part to ensure that the model can adapt to the unique characteristics of each human body part.

[0092] Furthermore, in this embodiment, the present application also optimizes the loss function.

[0093] To ensure high-quality synthetic CT images, a composite loss function consisting of multiple loss terms is used. The adversarial loss promotes the authenticity and consistency of the image through an adversarial process between the generator and the discriminator. Aims to generate images that can deceive the discriminator, and the adversarial loss of the discriminator We focus on accurately classifying real and synthetic images. The adversarial loss function is defined as follows:

[0094] The total loss function of the unified model is:

[0095]

[0096] Among them, α GAN =1.0,α cycle =10.0,α identity =5.0,α perceptual =0.01, α adv =1.0.

[0097]

[0098] in represents the synthetic CT image generated by the generator, represents the input MRI image,

[0099] represents the real CT image, D CT is the discriminator, the expected value and Represent the expected values ​​of the generated samples and real samples respectively. The adversarial process between the generator and the discriminator continuously enhances the authenticity and structural consistency of the synthetic CT images.

[0100] In addition to the adversarial loss, a cycle consistency loss is used to preserve the structural features of the image. This loss ensures that the image can be restored to its original form after transformation:

[0101]

[0102] in, represents the input MRI image, represents the real CT image, G MR→CT represents the generator mapping from MRI images to synthetic CT images, G CT→MR Represents the inverse mapping from CT images to MRI images.

[0103] Again, an identity loss is applied to maintain style consistency, ensuring that the input and output images remain the same when no transformation is applied:

[0104]

[0105] in, represents the input MRI image, G CT→MR represents the inverse mapping from CT image to MRI image, represents the real CT image, G MR→CT Represents the generator mapping from MRI images to synthetic CT images.

[0106] The perceptual loss function calculates the feature difference between the generated image and the real image at multiple scales. Through multi-scale perceptual loss, the model can evaluate the quality of the generated image at a more detailed level. The features provided by the supervised learning model (such as MAE) are extracted at multiple scales. and real images Feature extraction is performed at different scales, and then the perceptual loss is calculated. Specifically, the perceptual loss function in this application is:

[0107] in, G represents the features extracted by the MAE model at the kth scale. By extracting features at multiple scales, it is possible to capture information in the image from different levels (for example, low scale captures local details, high scale captures global structure). K is the number of scales, for example, features at low, medium, and high scales can be extracted. MR→CT represents the generator mapping from MRI images to synthetic CT images, represents the input MRI image, represents the real CT image, and ∥·∥2 is the L2 norm, which is used to calculate the difference between feature vectors.

[0108] The above perceptual loss function has the following advantages:

[0109] 1. Multi-scale feature extraction: Using multi-scale feature extraction can better capture the details and structures of different levels of the image. Low-scale features focus on global structure and large-scale information, while high-scale features focus on more subtle local features and details.

[0110] 2. Enhanced perception capabilities: Calculating losses at multiple scales enables the generator to optimize the generated images at each scale, improving the quality of image generation, especially in the reconstruction of details and structures.

[0111] 3. Improve the details and global consistency of generated images: Multi-scale learning can effectively avoid the limitations of a single scale and help the generator to ensure the rationality of both details and global structure when processing medical images (such as the mapping between MRI and CT images).

[0112] like Figure 3 As shown, Figure 3 The second column is a real CT image, and the sixth column is a pseudo CT image generated by the method of the present application. Columns 3 to 5 are pseudo CT images generated by other existing technologies. The area framed in the figure is the main comparison area. Through the comparison, it can be seen that the pseudo CT image generated by the method of the present application is more accurate and clear. And the accuracy of the pseudo CT images generated by the method of the present application for different parts of the human body is relatively high.

[0113] Furthermore, a protection method for the trained unified model is also included, and the protection method includes:

[0114] Through the authorized device identifier, authorized region information and the encryption and decryption mechanism of the unified model, it is ensured that the unified model is only run on the authorized device and the authorized region. The protection method specifically includes: matching the authorized device identifier and the authorized region information with the unified model, encrypting the parameters in the unified model, and only when the device identifier of the actual use device of the unified model is consistent with the authorized device identifier, and the region information of the actual use region is consistent with the authorized region information, the internal parameters of the unified model are decrypted.

[0115] Furthermore, the authorized area information includes joint encoding of geographic coordinates and text information.

[0116] The present application combines geographic location information for encoding, and introduces geographic coordinates (such as longitude and latitude) to enhance the diversity of regional information vectors.

[0117] In the above embodiment of the present application, geographic coordinates are mixed with text information and embedded, that is, geographic coordinates (such as the city center location) and text information (city + hospital + department) are jointly encoded. For example, GPS embedding + Transformer model is used to map geographic information to semantic space.

[0118] In the above-mentioned embodiments of the present application, location embedding and regional division are applied, that is, the region is divided into multiple grids (such as based on geographic boundaries), the grid ID is used as auxiliary information, and the description of the hospital and department is combined to generate a regional awareness embedding. That is, when defining text information (such as cities, hospitals, and departments), the geographic grid division is combined to enhance regional awareness. Specifically, the geographic area is first divided into multiple grids (for example, based on administrative regions or geographic boundaries), and then the grid IDs of these grids are used as auxiliary information. Then, a regional awareness embedding is generated in combination with the specific hospital and department descriptions, which can more accurately express the geographical and medical characteristics of different regions (cities, hospitals, departments). This method can combine geographic information with text information, so that the model can better understand and process these regional characteristics. After fusing the geographic location information, the regional vector of this application has the intuitiveness of physical space and can capture distance and proximity relationships.

[0119] The specific implementation of the above protection method is described in detail below.

[0120] 1. Extract device identifier

[0121] Extract a single hardware identifier from the authorized device or concatenate multiple hardware hardware identifiers together. These identifiers are used to generate the authorized device identifier of the device to ensure that the model file can only be decrypted and used on that specific device. Common hardware identifiers include:

[0122] CPU serial number: A unique identifier for each processor, used to distinguish different processors.

[0123] Motherboard Serial Number: A unique number on a computer's motherboard that ensures physical identification of the device.

[0124] MAC Address: The hardware address of a network adapter, used to identify the device's network interface.

[0125] Hard Drive Serial Number: A unique identifier for a storage device that helps further uniquely identify the device.

[0126] The above hardware identifier will be processed through a hash algorithm (such as SHA-256) to generate an authorized device identifier of the authorized device.

[0127] The authorized device identifier serves as the basis for device authentication throughout the encryption and decryption process, ensuring that the unified model can only be decrypted and used on a device with a matching authorized device identifier.

[0128] 2. Generate authorized area information and encryption matrix

[0129] After the authorized device identifier is generated, the authorized region information needs to be generated.

[0130] 2.1 Obtaining authorized area information: Based on the information of the authorized use area of ​​the unified model, the authorized area information is generated using hash algorithms such as SHA-256.

[0131] 2.2 The information about the authorized use area is obtained by the following method: obtaining the information about the authorized use area through GPS or network IP. The information about the authorized use area can be composed of GPS location coordinates or IP address. It can also be a joint encoding of geographic coordinates and text information.

[0132] In hospitals, the use of models is often restricted by physical location. For example, some models may only be allowed to be decrypted and used in specific departments (such as radiology and imaging) to ensure data security. For this reason, geographic location information is used as another constraint during the encryption seed generation process.

[0133] Example of geographic location verification: requiring the device to be within the authorized area of ​​the hospital (such as the designated area of ​​the imaging department), otherwise the model will refuse to decrypt. This ensures that model decryption does not only rely on the hardware identifier of the device, but also requires decryption and use within the preset geographic area.

[0134] 2.3 Generate encryption matrix:

[0135] The above authorized device identifier and authorized region information can be concatenated together to form an encryption seed. Based on the encryption seed, an encryption matrix with the same dimension as the parameters of the model is generated using a pseudo-random generation algorithm (such as Mersenne Twister).

[0136] For example, in an encryption matrix, each matrix element is a floating point number between 0 and 1, representing an encryption factor.

[0137] Hospital regional encryption matrix: To further improve security, different "regional encryption matrices" can be generated for different areas within the hospital (such as different departments), thereby achieving fine-grained regional authority control and ensuring that the model fails to decrypt in unauthorized areas.

[0138] 3. Model encryption operation

[0139] The encryption operation of the unified model occurs at the model parameter level, that is, each weight and bias parameter of the model is encrypted. The specific steps are as follows:

[0140] Encrypt model parameters element by element: For each parameter wi in the model and the encryption factor ei in the encryption matrix, perform element-by-element multiplication to generate the encrypted parameter wi′=wi×ei.

[0141] Furthermore, an encrypted model file can be generated: the encrypted model file (such as .pt or .h5 format) is stored in a form that cannot be used directly. It can only be decrypted and restored when the device identifier of the actual use device of the unified model is consistent with the authorized device identifier, and the region information of the actual use area is consistent with the authorized region information.

[0142] 4. Model decryption operation

[0143] The decryption process of the unified model requires dual verification of the device identifier and geographic location information to ensure that the model can only be decrypted and used on legitimate devices and in designated areas. The specific steps are as follows:

[0144] 4.1 Device Verification and Location Verification: When the system loads the encryption model, it first extracts the hardware identifier of the currently used device, and generates the device identifier of the actual used device in accordance with the method of generating the authorized device identifier. At the same time, the system obtains the geographic location information (such as GPS coordinates or IP address) of the currently used device in real time, and generates the region information of the actual used region in accordance with the method of generating the authorized region information based on the geographic location information of the actual used device. The system will only continue the decryption operation when the device identifier of the actual used device in the unified model is consistent with the authorized device identifier, and the region information of the actual used region is consistent with the authorized region information.

[0145] 4.2 Generate decryption matrix: The decryption matrix is ​​the same as the encryption matrix. That is, the decryption factor ei′ of the decryption matrix is ​​the same as the corresponding encryption factor ei of the encryption matrix.

[0146] 4.3 Decryption model parameters: Perform element-by-element division on each encrypted parameter wi′ in the encryption model and the decryption factor ei′ of the corresponding decryption matrix to restore the original model parameter wi=wi′

[0147] / ei′.

[0148] 5. Security Analysis

[0149] The above two-factor authentication encryption mechanism has the following key security features:

[0150] 5.1 Uniqueness: Model encryption is based on device identifiers and geographic location information, so each model encrypted file cannot be decrypted on different devices or regions, preventing cross-device and cross-region use. If the device identifier or geographic location information does not match, the decryption matrix cannot be generated, resulting in the inability to recover the model parameters. The output of the model will become 0 or random noise, ensuring that unauthorized devices and regions cannot decrypt and use the model file.

[0151] 5.2 Geographically restricted decryption control: If the current device is not in the authorized area, even if the device identifier matches, the system will refuse to generate the decryption matrix, resulting in decryption failure. This double restriction ensures that the use of the model file is limited to the specific geographic location of the hospital.

[0152] 5.3 Anti-reversal: Encrypted seeds and matrices are generated using device identifiers and geographic location constraints to ensure that even if an attacker obtains the encrypted model file, it is impossible to decrypt or restore the original model parameters by simple modification.

[0153] 5.4 Prevent model theft and illegal use: The dual authentication mechanism of geolocation and device identifier ensures that the model is only used on authorized devices and designated areas, preventing the model from being decrypted outside the hospital or on unauthorized devices.

[0154] 5.5 Operation traceability: Record operation logs during decryption and use, including device identifiers, decryption time, and geographic location information, to ensure that decryption operations meet the hospital's security and compliance requirements. This operation log is stored through secondary encryption to ensure the security of log data and prevent unauthorized access.

[0155] The second embodiment of the present application provides a computer storage medium, in which at least one executable instruction is stored, and the executable instruction enables a processor to execute the operation of the unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images.

[0156] The third embodiment of the present application provides a computer device, including: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images.

[0157] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A unified method for generating pseudo CT images based on multi-sequence, multi-anatomical MRI images, characterized in that: include: Assign prompt codes specific to the sequence of MRI images of multiple different human body parts, and form training data with the prompt codes, corresponding MRI images, and corresponding real CT images. The unified model is trained by training data, and the part-specific attention module RSAM is added to the unified model. The MRI image of the human body part is input into the trained unified model to generate a pseudo CT image, wherein the MRI image is assigned a prompt code dedicated to the human body part corresponding to the MRI image.

2. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 1, characterized in that: The part-specific attention module RSAM is used to receive input features from the residual block and adjust the weights of these features according to the hint code.

3. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 2, characterized in that: The total loss function of the unified model is: Among them, among them GAN =1.0、a cycle =10.0、a identity =5.0、a perceptual =0.01、a adv = 1.

0.

4. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 3, characterized in that: in represents the synthetic CT image generated by the generator, represents the input MRI image, represents the real CT image, D CT is the discriminator, the expected value and They represent the expected values ​​of the generated samples and the real samples respectively, in, represents the input MRI image, represents the real CT image, G MR→CT represents the generator mapping from MRI images to synthetic CT images, G CT→MR represents the inverse mapping from CT image to MRI image, in, represents the input MRI image, G CT→MR represents the inverse mapping from CT image to MRI image, represents the real CT image, G MR→CT represents the generator mapping from MRI images to synthetic CT images, in, represents the features extracted by the MAE model at the kth scale, K is the number of scales, G MR→CT represents the generator mapping from MRI images to synthetic CT images, represents the real CT image, / / · / / 2 is the L2 norm, which is used to calculate the difference between feature vectors.

5. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 1, characterized in that: The invention also includes a protection method for the unified model after training, and the protection method includes: By means of the authorized device identifier, the authorized region information and the encryption and decryption mechanism of the unified model, it is ensured that the unified model only runs on the authorized device and the authorized region.

6. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 5, characterized in that: The authorized area information includes the joint encoding of geographic coordinates and text information. When defining the text information, geographic grid division is combined to enhance regional awareness. Specifically, the geographic area is first divided into multiple grids, and then the grid IDs of these grids are used as auxiliary information. Then, a region-aware embedding is generated in combination with specific hospital and department descriptions.

7. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 5, characterized in that: The protection method specifically includes: Match the authorized device identifier and authorized region information with the unified model, Encrypt the parameters within the unified model, The internal parameters of the unified model are decrypted only when the device identifier of the device actually using the unified model is consistent with the authorized device identifier and the region information of the actual use region is consistent with the authorized region information.

8. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 7, characterized in that: The authorized device identifier is generated by the following method: The hardware information of the authorized device is processed through a hash algorithm to generate an authorized device identifier. Authorized region information is generated by the following method: The information of the authorized use area of ​​the unified model is processed by a hash algorithm to generate the authorized area information.

9. The unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images according to claim 8, characterized in that: Encrypt the parameters in the unified model, including: Based on the authorized device identifier and authorized region information, an encryption matrix with the same dimension as the internal parameters of the unified model is generated by using a pseudo-random generation algorithm. Internal parameters of the unified model are encrypted based on the encryption matrix.

10. A computer storage medium, characterized in that: The computer storage medium stores at least one executable instruction, which enables the processor to perform the operation of the unified method for generating pseudo CT images based on multi-sequence, multi-anatomical site MRI images as described in any one of claims 1 to 9.

Images