A medical image driven craniocerebral ultrasound simulation parameter modeling method

Abstract

This invention discloses a medical image-driven method for modeling parameters in cranial ultrasound simulation, comprising the following steps: acquiring multimodal medical images of the cranium, wherein the multimodal medical images include at least a first image providing bone density information and a second image providing soft tissue structure information; spatially aligning the first and second images to obtain aligned images; extracting skull tissue regions and various soft tissue regions from the aligned images to obtain tissue mask maps; mapping corresponding acoustic parameters to different tissue regions based on the tissue mask maps to obtain reference acoustic parameter distribution maps; training a generative model using the first image, the second image, and the reference acoustic parameter distribution map as training data; inputting the target medical image to be processed, and using the trained generative model to generate a target acoustic parameter distribution map, which is then used for cranial ultrasound simulation. This invention provides technical support for real-time simulation planning of the cranium while ensuring targeting accuracy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical ultrasound technology and computer vision, and in particular relates to a medical image-driven method for modeling cranial ultrasound simulation parameters. Background Technology

[0002] With the development of non-invasive ultrasound modulation technology, transcranial focused ultrasound has become an important technical means in the fields of brain science research and neurofunctional modulation. Due to the high density, non-uniformity, and complex geometry of the skull, ultrasound waves will produce significant phase distortion and energy attenuation when penetrating the skull. In order to achieve high-precision focusing control, acoustic simulation based on individual imaging data is required before ultrasound transmission to calculate the phase and amplitude compensation parameters of each transducer element.

[0003] Currently, the mainstream approach to establishing individualized intracranial acoustic models utilizes computed tomography (CT) images. This approach calculates the density and sound velocity distribution of the skull through a linear mapping of CT values, simplifying intracranial soft tissues (such as gray matter, white matter, and cerebrospinal fluid) as isotropic homogeneous media. However, this simplification ignores the non-uniform characteristics of soft tissues, leading to a shift in the calculated focal region and reducing the accuracy of the sound field simulation. While fusing CT and magnetic resonance imaging (MRI) can supplement the structural information of soft tissues, acquiring two sets of high-quality images simultaneously increases data acquisition costs and radiation exposure risks in practical applications. Furthermore, traditional modeling processes based on manual segmentation or thresholding are cumbersome and time-consuming, making it difficult to meet the efficiency requirements of rapid intraoperative planning.

[0004] Therefore, how to achieve a high-precision, high-efficiency method for modeling the acoustic parameters of the entire brain tissue that can reduce dependence on multimodal imaging is a technical problem that urgently needs to be solved in the field of transcranial focused ultrasound simulation. Summary of the Invention

[0005] To overcome the shortcomings of existing transcranial focused ultrasound simulation modeling, such as single tissue modeling, low computational efficiency, and excessive reliance on multimodal images, this invention provides a medical image-driven method for parametric modeling of cranial ultrasound simulation, comprising the following steps: Acquire multimodal medical images of the brain, wherein the multimodal medical images include at least a first image providing bone density information and a second image providing soft tissue structure information; Spatially align the first image and the second image to obtain the aligned image; Based on the aligned image, the skull tissue region and various soft tissue regions are extracted to obtain a tissue mask map; Based on the tissue mask map, acoustic parameters are mapped to different tissue regions to obtain a reference acoustic parameter distribution map; Construct a generative model that realizes end-to-end mapping from medical images to acoustic parameter distribution maps; The generative model is trained using the first image, the second image, and the reference acoustic parameter distribution map as training data. Input the target medical image to be processed, and use the trained generative model to generate the target acoustic parameter distribution map, which is used for cranial ultrasound simulation.

[0006] Optionally, spatial alignment of the first image and the second image includes: The second image is transformed to the spatial coordinate system of the first image using a rigid registration algorithm, so that the first image and the second image are spatially consistent in terms of anatomical structure.

[0007] Optionally, based on the aligned image, the skull tissue region and various soft tissue regions are extracted to obtain a tissue mask map, including: The skull tissue region is extracted from the first image based on a preset pixel threshold. Based on a preset automatic segmentation algorithm, multiple soft tissue regions are extracted from the aligned second image. These multiple soft tissue regions include gray matter regions, white matter regions, and cerebrospinal fluid regions.

[0008] Optionally, based on the tissue mask map, acoustic parameters are mapped to different tissue regions to obtain a reference acoustic parameter distribution map, including: For the skull tissue region, the sound velocity and density values ​​of each voxel are determined according to the pixel values ​​of each voxel in the first image and based on a preset physical mapping relationship. For the various soft tissue regions, according to the type of each soft tissue region, a preset standard sound velocity constant and standard density constant are assigned to the voxels in the corresponding region; The sound velocity and density values ​​calculated from the skull tissue region are spatially merged with the standard sound velocity constants and standard density constants assigned to each soft tissue region to obtain a reference acoustic parameter distribution map.

[0009] Optionally, the generative model is trained using the first image, the second image, and the reference acoustic parameter distribution map as training data, including: The generative model includes a generator and a discriminator; The first image and the second image are input into the generator to generate a predicted acoustic parameter distribution map; The discriminator is used to distinguish between the predicted acoustic parameter distribution map and the actually obtained reference acoustic parameter distribution map; The generator and the discriminator are alternately optimized and trained by combining the least squares adversarial loss function, the pixel-level reconstruction loss function, and the perceptual loss function until an equilibrium state is reached.

[0010] Optionally, constructing the generative model includes: The generator employs a dual-stream encoder structure to extract features from the first image and the second image, respectively. In the downsampling path of the encoder, a feature fusion module with an attention mechanism is introduced to fuse the features extracted from the two images. The synthesis network is constructed using residual blocks, and skip connections are used to connect the corresponding layers of the encoder and decoder.

[0011] Optionally, the generative model is any one of a diffusion probability model, a variational autoencoder, a generative model based on the Transformer architecture, or an autoregressive generative model.

[0012] Optionally, the first image is a computed tomography image, and the second image is a magnetic resonance imaging (MRI) image, wherein the MRI image includes at least one of T1-weighted imaging, T2-weighted imaging, or liquid attenuation inversion recovery sequence imaging.

[0013] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0014] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention acquires at least a first image providing bone density information and a second image providing soft tissue structure information, spatially aligns the two, extracts cranial tissue regions and various soft tissue regions, and obtains a tissue mask map. Based on the acoustic parameters mapped to different tissue regions using this mask map, a reference acoustic parameter distribution map is obtained. This overcomes the limitation of existing technologies that treat brain soft tissue as a homogeneous medium, establishing a non-uniform acoustic parameter model encompassing multiple tissues such as the skull, gray matter, white matter, and cerebrospinal fluid. This significantly improves the accuracy of phase error calculation when ultrasound penetrates non-uniform cranial tissue, making the simulated focal zone localization and energy distribution calculation more accurate.

[0016] This invention constructs a generative model that achieves end-to-end mapping from medical images to acoustic parameter distribution maps, and uses the first image, the second image, and the reference acoustic parameter distribution map as training data. This avoids the tedious manual segmentation, multimodal registration, and pixel-by-pixel physical calculations required in traditional processes, reducing the modeling process from hours or even days to seconds or minutes. This provides technical support for rapid simulation planning, enabling more efficient iteration and optimization of transcranial focused ultrasound simulation planning.

[0017] This invention takes a target medical image as input and uses a trained generative model to generate a target acoustic parameter distribution map. Since the model has learned the potential mapping relationships between different modal images and acoustic parameters during the training phase, only a single modal image is needed as input during the inference phase to predict the acoustic parameter map of the entire tissue, including skull information. This significantly reduces the additional radiation risks and data acquisition costs associated with computed tomography scans of subjects, and improves the versatility of the simulation method. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the modeling process for the acoustic parameters of the brain according to an embodiment of the present invention; Figure 2 This is a diagram of the generated network architecture according to an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 This embodiment provides a medical image-driven method for modeling cranial ultrasound simulation parameters, including the following steps: Acquire multimodal medical images of the brain, wherein the multimodal medical images include at least a first image providing bone density information and a second image providing soft tissue structure information; Spatially align the first image and the second image to obtain the aligned image; Based on the aligned image, the skull tissue region and various soft tissue regions are extracted to obtain a tissue mask map; Based on the tissue mask map, acoustic parameters are mapped to different tissue regions to obtain a reference acoustic parameter distribution map; Construct a generative model that realizes end-to-end mapping from medical images to acoustic parameter distribution maps; The generative model is trained using the first image, the second image, and the reference acoustic parameter distribution map as training data. Input the target medical image to be processed, and use the trained generative model to generate the target acoustic parameter distribution map, which is used for cranial ultrasound simulation.

[0022] like Figure 1 As shown, as a specific implementation method, the following steps are included: Step 1: Acquire multimodal medical images: Multimodal medical images of the brain were acquired from the same subject. The first image was a computed tomography (CT) image, used to provide bone density information; the second image was a magnetic resonance imaging (MRI) image, specifically T1-weighted imaging (T1-MRI), used to provide soft tissue structure information. The original images were stored in NIfTI format.

[0023] As an alternative implementation, the input images are not limited to T1-MR and conventional CT. Image modalities that provide similar anatomical or density information, such as T2-MR, FLAIR sequences, dual-energy CT, and CBCT, can also be used as input data in this embodiment.

[0024] Step 2: Spatially align the first and second images: Using CT images as a reference space, a rigid registration algorithm is employed to translate and rotate MRI images, aligning them to the spatial coordinate system of the CT images. This ensures spatial consistency between the CT and MRI images in terms of anatomical structure. After alignment, isotropic resampling is performed on both CT and MRI images to unify the spatial resolution to 1mm × 1mm × 1mm, eliminating the impact of inconsistent slice thicknesses. Simultaneously, for the CT images, erosion, extraction of maximum connected components, and dilation operations are used to remove artifacts from non-biological tissues such as the head frame and coils.

[0025] Step 3: Based on the aligned image, extract the skull tissue region and various soft tissue regions to obtain a tissue mask image: (1) Extract the skull tissue region: Extract the skull tissue region from the aligned CT image based on the preset pixel threshold (Henness unit value greater than 300) to obtain the skull mask.

[0026] (2) Extraction of multiple soft tissue regions: Using a whole-brain automatic segmentation tool (such as FreeSurfer), multiple soft tissue regions, including gray matter, white matter and cerebrospinal fluid regions, are extracted from the aligned MRI images to obtain the corresponding soft tissue mask.

[0027] The skull mask and the soft tissue masks are combined to form a tissue mask map containing identifiers of different tissue categories.

[0028] As an additional implementation method, the extraction of skull and soft tissue is not limited to thresholding or specific segmentation networks. Any image processing algorithm capable of accurately identifying tissue boundaries, including but not limited to region growing, level sets, graph cut, and various deep learning-based semantic segmentation models (such as U-Net, Transformer architecture, etc.), can be applied to this step.

[0029] Step 4: Based on the tissue mask map, map the corresponding acoustic parameters to different tissue regions to obtain a reference acoustic parameter distribution map. Specifically: Based on the Hounsfield Unit (HU) value from CT scans, the sound velocity and density of the skull are calculated using the following formula: (1) (2) (3) Where HU represents the Hounsfield Unit in CT images, φ represents the skull porosity calculated based on the HU value, and ρ represents the density distribution of the skull. min and ρ max These represent the preset minimum and maximum skull density values, respectively, which are 1000 kg / m³. 3 and 1950kg / m 3 ;c represents the speed of sound propagation in the skull, c min and c max These represent the preset minimum and maximum values ​​of the sound velocity in the skull, which are 1500 m / s and 3100 m / s respectively.

[0030] Based on the soft tissue mask category, standard acoustic constants are set for each of these soft tissues.

[0031] The sound velocity and density values ​​calculated from the skull tissue region are spatially merged with the standard sound velocity constants and standard density constants assigned to each soft tissue region to obtain a reference acoustic parameter distribution map covering the entire brain. The reference acoustic parameter distribution map includes a sound velocity distribution map and a density distribution map.

[0032] Step 5: Construct a generative model that achieves end-to-end mapping from medical images to acoustic parameter distribution maps; train the generative model using the first image, the second image, and the reference acoustic parameter distribution map as training data. Specifically: This embodiment employs a multimodal input, feature fusion-based generative adversarial network (GAN) to automate acoustic parameter modeling. Its network architecture is as follows: Figure 2 As shown.

[0033] The generative adversarial network (GAN) comprises a generator and a discriminator. The generator employs a dual-stream encoder based on a U-Net structure to independently extract features from computed tomography (CT) and magnetic resonance imaging (MRI) images. A feature fusion module is included in each downsampling path of the encoder. This module uses a squeeze-and-excitation block (SEB) structure, incorporating an attention mechanism to adaptively learn and fuse complementary information from the two types of images.

[0034] The generator's synthesis network is constructed using residual blocks to deepen the network layers and learn the complex nonlinear mapping from medical images to acoustic parameter distribution maps, while avoiding the vanishing gradient problem. Furthermore, the generator employs skip connections to directly pass the detailed features extracted from each layer of the encoder to the corresponding layers of the decoder, enabling the reconstruction of images to simultaneously preserve the detailed information of both computed tomography (CT) and magnetic resonance imaging (MRI) images.

[0035] The discriminator employs spectral normalization to maintain the stability of the generative adversarial network training.

[0036] To guide the model in generating high-quality images, the overall objective function Ltotal for generating G in this embodiment is jointly driven by three parts: least squares adversarial loss, pixel-wise reconstruction loss, and perceptual loss.

[0037] First, to address the gradient vanishing problem that is easily caused by using sigmoid cross-entropy loss in the original GAN, and to further improve the stability of the training process, this embodiment adopts the loss architecture of least squares generative adversarial network (LSGAN). The adversarial loss of generator G... Defined as: (4) The loss function aims to make the generated image G(x) produced by the generator G deceive the discriminator D as much as possible, where E x This represents the mathematical expectation of the data distribution of the input sample x.

[0038] Secondly, to ensure that the generated image maintains a high degree of consistency with the real target image (Ground Truth, y) in terms of global topological structure and low-frequency information, this embodiment introduces a pixel-level reconstruction loss based on the L1 norm. Defined as: (5) Among them, E x,y This represents the mathematical expectation of the joint data distribution of the sample pair (x,y).

[0039] However, relying solely on pixel-level loss often leads to the loss of high-frequency details in the generated image, resulting in an overly smooth visual effect. To compensate for this deficiency, enhance the visual realism of the image, and preserve complex texture information, this embodiment further introduces perceptual loss. Specifically, the generated image and the real image are respectively input into a pre-trained VGG-19 network to extract deep feature representations, and the L1 distance between the two in the feature space is calculated: (6) in, This indicates the feature extraction operation of the pre-trained VGG-19 network.

[0040] Therefore, the total loss is: (7) in, and These represent the weight hyperparameters for pixel-level reconstruction loss and perceptual loss, respectively, and are used to balance the contribution of each loss term to the overall objective.

[0041] In each training iteration, the generator parameters are first fixed, and the first image and the second image are input into the generator to obtain the predicted acoustic parameter distribution map; the predicted acoustic parameter distribution map is then compared with the reference acoustic parameter distribution map. Figure 1 The discriminator is then fed into the generator, its loss function is calculated, and its parameters are updated. Subsequently, with the discriminator's parameters fixed, the generator's total loss function is calculated, and its parameters are updated. This alternating optimization process is repeated until the generator and discriminator reach an equilibrium state, meaning the discriminator can no longer effectively distinguish between the predicted acoustic parameter distribution and the true reference acoustic parameter distribution. At this point, training is complete.

[0042] Step Six: After training is complete, input the target medical image to be processed into the trained generator, and a target acoustic parameter distribution map for cranial ultrasound simulation can be generated within seconds.

[0043] As an additional implementation, the generative model can be replaced with any deep learning architecture capable of learning a nonlinear mapping from multimodal medical image features to a personalized acoustic parameter field. Specifically, the generative model includes, but is not limited to: diffusion probability models (such as the denoising diffusion probability model DDPM, the latent diffusion model LatentDiffusion), variational autoencoders and their variants (such as the conditional variational autoencoder CVAE, the vector quantization variational autoencoder VQ-VAE), generative models based on the Transformer architecture (such as the Vision Transformer generator), or autoregressive generative models.

[0044] Compared with the prior art, this embodiment has the following beneficial effects: First, this embodiment overcomes the limitation of single-tissue modeling in existing transcranial focused ultrasound simulation modeling. Utilizing high-density skull information from computed tomography (CT) images and fine soft tissue structure information from magnetic resonance imaging (MRI), a non-uniform acoustic parameter model incorporating various tissues such as the skull, gray matter, white matter, and cerebrospinal fluid is established, overcoming the previous limitation of treating brain soft tissue as a homogeneous medium. T1-weighted MRI is used to assign acoustic parameters to subdivided tissues, significantly improving the accuracy of phase error calculation when ultrasound penetrates non-uniform cranial tissue, resulting in more accurate simulation focal zone localization and energy distribution calculation.

[0045] Second, this embodiment overcomes the shortcomings of existing methods in terms of low computational efficiency. It utilizes the nonlinear feature mapping capability of generative adversarial networks to learn the direct transformation logic from raw medical images to acoustic parameter spaces (including sound velocity and density). By adopting an end-to-end generation approach, it avoids the tedious manual segmentation, multimodal registration, and pixel-by-pixel physical calculations of traditional processes, shortening the modeling process from hours or even days to seconds or minutes. This provides technical support for rapid intraoperative simulation planning, enabling faster iteration and optimization of preoperative simulation planning for transcranial focused ultrasound.

[0046] Third, this embodiment overcomes the shortcomings of existing methods that over-rely on multimodal imaging. By jointly training paired image data, the model captures the potential correlations between different modalities. During the inference phase, through feature mapping using deep learning, the trained model can predict the acoustic parameter map of the entire tissue containing skull information with only single-modal image input (e.g., only magnetic resonance imaging or only low-dose computed tomography scan), thereby reducing the data acquisition cost for subjects, reducing ionizing radiation exposure, and improving the universality of the simulation method.

[0047] Example 2 This embodiment provides another method for medical image-driven cranial ultrasound simulation parameter modeling, including the following steps.

[0048] Step 1: Multimodal medical image acquisition and preprocessing: (1) Data acquisition: Computed tomography (CT) images and T1-weighted magnetic resonance imaging (T1-MRI) images of the same subject were acquired. The raw images were stored in medical image format (such as NIfTI format).

[0049] (2) Isotropic resampling: Using medical image processing tools, CT images and MRI images are resampled to the same spatial resolution, such as isotropic voxels of 1 mm × 1 mm × 1 mm, to eliminate the influence of inconsistent acquisition slice thickness.

[0050] (3) Spatial registration: Using CT images as the reference space, a rigid registration algorithm is used to translate and rotate MRI images to ensure that the anatomical structures of the two are completely overlapped.

[0051] (4) Artifact removal: For CT images, morphological operations such as erosion, extraction of the largest connected component, and dilation are used to remove non-biological tissue artifacts such as head frame and coils, so as to reduce the interference of noise on model training.

[0052] Step 2: Construction of the Cranial Acoustic Parameter Dataset: (1) Mask extraction: The skull mask is extracted from the CT images using the Henle unit threshold method (Henle unit value greater than 300); the MRI images are processed using a whole-brain automatic segmentation tool (e.g., FreeSurfer) to extract soft tissue masks including gray matter, white matter, and cerebrospinal fluid.

[0053] (2) Physical parameter mapping: The Henle unit values ​​of CT images are converted into the sound velocity and density values ​​of the skull using the mapping formula; according to the type of soft tissue mask, standard acoustic physical constants, including sound velocity constant and density constant, are assigned to different soft tissue brain regions.

[0054] (3) Spatially merge the mapped skull parameters and soft tissue parameters to generate a three-dimensional acoustic parameter matrix covering the entire brain.

[0055] Step 3: Building and training the network architecture: Reference Figure 2 The process of building and training an end-to-end generative model includes the following sub-steps: (1) Dual-stream encoder design: The density features of CT images and the tissue features of MRI images are extracted by using separate encoders.

[0056] (2) Feature calibration and fusion: During the downsampling process, an attention mechanism is introduced through the Squeeze-and-Excitation Block (SEB) to adaptively learn and enhance complementary information between different modalities.

[0057] (3) Use residual blocks to build synthetic networks to increase model depth to learn complex nonlinear mappings while avoiding the gradient vanishing problem.

[0058] (4) Objective function setting: The adversarial loss of least squares generative adversarial network is adopted to improve training stability; the pixel-level first norm reconstruction loss and the perceptual loss based on VGG-19 network are combined to ensure that the generated parameter map conforms to physical reality in both macroscopic values ​​and fine structure.

[0059] (5) Stabilization mechanism: Spectral normalization is used in the discriminator to maintain training balance.

[0060] Step Four: Reasoning Stage Once the model is trained, it supports single-modal or dual-modal image-driven inference and can generate an acoustic parameter matrix for cranial ultrasound simulation within seconds.

[0061] Example 3 This embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0062] Example 4 This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0063] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for modeling cranial ultrasound simulation parameters driven by medical images, characterized in that, Includes the following steps: Acquire multimodal medical images of the brain, wherein the multimodal medical images include at least a first image providing bone density information and a second image providing soft tissue structure information; Spatially align the first image and the second image to obtain the aligned image; Based on the aligned image, the skull tissue region and various soft tissue regions are extracted to obtain a tissue mask map; Based on the tissue mask map, acoustic parameters are mapped to different tissue regions to obtain a reference acoustic parameter distribution map; Construct a generative model that realizes end-to-end mapping from medical images to acoustic parameter distribution maps; The generative model is trained using the first image, the second image, and the reference acoustic parameter distribution map as training data. Input the target medical image to be processed, and use the trained generative model to generate the target acoustic parameter distribution map, which is used for cranial ultrasound simulation.

2. The method according to claim 1, wherein spatial alignment of the first image and the second image is performed, comprising: The second image is transformed to the spatial coordinate system of the first image using a rigid registration algorithm, so that the first image and the second image are spatially consistent in terms of anatomical structure.

3. The method according to claim 1, wherein based on the aligned image, the skull tissue region and various soft tissue regions are extracted to obtain a tissue mask map, comprising: The skull tissue region is extracted from the first image based on a preset pixel threshold. Based on a preset automatic segmentation algorithm, multiple soft tissue regions are extracted from the aligned second image. These multiple soft tissue regions include gray matter regions, white matter regions, and cerebrospinal fluid regions.

4. The method according to claim 1, wherein acoustic parameters are mapped to different tissue regions based on the tissue mask map to obtain a reference acoustic parameter distribution map, comprising: For the skull tissue region, the sound velocity and density values ​​of each voxel are determined according to the pixel values ​​of each voxel in the first image and based on a preset physical mapping relationship. For the various soft tissue regions, according to the type of each soft tissue region, a preset standard sound velocity constant and standard density constant are assigned to the voxels in the corresponding region; The sound velocity and density values ​​calculated from the skull tissue region are spatially merged with the standard sound velocity constants and standard density constants assigned to each soft tissue region to obtain a reference acoustic parameter distribution map.

5. The method according to claim 1, wherein the generative model is trained using the first image, the second image, and the reference acoustic parameter distribution map as training data, comprising: The generative model includes a generator and a discriminator; The first image and the second image are input into the generator to generate a predicted acoustic parameter distribution map; The discriminator is used to distinguish between the predicted acoustic parameter distribution map and the actually obtained reference acoustic parameter distribution map; The generator and the discriminator are alternately optimized and trained by combining the least squares adversarial loss function, the pixel-level reconstruction loss function, and the perceptual loss function until an equilibrium state is reached.

6. The method according to claim 5, wherein constructing the generative model comprises: The generator employs a dual-stream encoder structure to extract features from the first image and the second image, respectively. In the downsampling path of the encoder, a feature fusion module with an attention mechanism is introduced to fuse the features extracted from the two images. The synthesis network is constructed using residual blocks, and skip connections are used to connect the corresponding layers of the encoder and decoder.

7. The method according to claim 1, wherein the generative model is any one of a diffusion probability model, a variational autoencoder, a generative model based on a Transformer architecture, or an autoregressive generative model.

8. The method according to claim 1, wherein the first image is a computed tomography image, the second image is a magnetic resonance image, and the magnetic resonance image includes at least one of T1-weighted imaging, T2-weighted imaging, or liquid attenuation inversion recovery sequence imaging.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images