Method for generating quantitative susceptibility map and method for training generation model

Abstract

The embodiment of the application provides a quantitative susceptibility image generation method and a training method of a generation model, a T1 image of a head of a target object directly obtainable by a magnetic resonance device is preprocessed to obtain a T1 image of a brain of the target object, the T1 image of the brain of the target object is input into a quantitative susceptibility image generation model which is pre-trained, the quantitative susceptibility image generation model can fuse a T1 image feature and a brain template feature, and thus outputs a quantitative susceptibility image of the brain of the target object. The widely popular T1 image can be directly used to generate the quantitative susceptibility image, the use threshold of the quantitative susceptibility imaging technology is significantly reduced, and the conversion of the quantitative susceptibility imaging technology to the clinical routine application is facilitated. The fusion of the T1 image feature and the brain template feature can be learned in a partitioned and targeted manner, accurate quantitative values are ensured in different brain regions, and the quantitative susceptibility image generation precision is improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method for generating quantitative magnetic susceptibility images and a method for training a generation model. Background Technology

[0002] Quantitative susceptibility mapping (QSM) is a technique used to quantitatively measure the distribution of magnetic susceptibility in tissues. Magnetic susceptibility is an inherent physical property of matter, primarily originating from the myelin sheath and iron deposition in brain tissue. Abnormal iron deposition in the brain is closely related to the development and progression of neurodegenerative diseases such as Parkinson's disease and Alzheimer's disease. Therefore, quantitative magnetic susceptibility images obtained through QSM can provide important imaging tools and insights for the diagnosis, pathophysiological research, and progression monitoring of these diseases.

[0003] Currently, obtaining a quantitative magnetic susceptibility image usually requires using a magnetic resonance imaging device to acquire the raw phase data of a specific multi-echo gradient echo sequence (such as a GRE or SWI sequence), and then going through a series of complex post-processing steps, including phase unwrapping, background field removal, and dipole inversion, before the quantitative magnetic susceptibility map can be finally reconstructed.

[0004] However, this process has specific requirements for the scanning sequence, and the post-processing is complex and computationally intensive, which limits the popularization and application of QSM in routine clinical examinations. Summary of the Invention

[0005] This application provides a method for generating a quantitative magnetic susceptibility image and a method for training a generation model, which simplifies the generation process of the quantitative magnetic susceptibility image and enhances its application value.

[0006] In a first aspect, embodiments of this application provide a method for generating a quantitative magnetic susceptibility image, the method comprising: Acquire a T1 image, which is obtained by scanning the head of the target object using a magnetic resonance imaging device; The T1 image is preprocessed to obtain a preprocessed T1 image; the preprocessing includes skull removal. The preprocessed T1 image is input into the quantitative magnetic susceptibility image generation model, which then performs the following processing: feature extraction is performed on the preprocessed T1 image to obtain T1 features; the T1 features and brain template features are fused to obtain fused features, wherein the brain template features are obtained by extracting features from the brain template based on the feature extraction method performed on the preprocessed T1 image; the fused features are encoded and decoded to output the quantitative magnetic susceptibility image of the target object's brain; the quantitative magnetic susceptibility image generation model is a trained deep learning model.

[0007] Secondly, embodiments of this application provide a training method for a quantitative magnetic susceptibility image generation model, the method comprising: Multiple sample pairs are acquired, each of which contains a T1 image sample and a QSM data sample obtained by scanning the head of the same object using a magnetic resonance imaging device. The T1 image samples are preprocessed to obtain preprocessed T1 image samples; The QSM data samples are preprocessed to obtain quantitative magnetic susceptibility image labels; An initial quantitative magnetic susceptibility image generation model is trained based on the preprocessed T1 image samples and the brain template. The initial quantitative magnetic susceptibility model is used to perform the following processing: extracting features from the preprocessed T1 image samples and the brain template using the same feature extraction method to obtain T1 feature samples and brain template feature samples; fusing the T1 feature samples and brain template feature samples to obtain fused feature samples; encoding and decoding the fused feature samples to output the initial quantitative magnetic susceptibility image corresponding to the T1 image samples. Based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample, the loss value is obtained; The initial quantitative magnetic susceptibility image generation model is iteratively trained based on the loss value until it converges, thus obtaining the quantitative magnetic susceptibility image generation model.

[0008] Thirdly, embodiments of this application provide an electronic device, including: a memory, a processor, and a communication interface; wherein, the memory stores executable code, and when the executable code is executed by the processor, the processor performs the method described in the first aspect.

[0009] Fourthly, embodiments of this application provide an electronic device, including: a memory, a processor, and a communication interface; wherein, the memory stores executable code, and when the executable code is executed by the processor, the processor performs the method as described in the second aspect.

[0010] Fifthly, embodiments of this application provide a non-transitory machine-readable storage medium storing executable code, which, when executed by a processor of an electronic device, enables the processor to at least implement the method described in the first aspect.

[0011] In a sixth aspect, embodiments of this application provide a non-transitory machine-readable storage medium storing executable code, which, when executed by a processor of an electronic device, enables the processor to at least implement the method described in the second aspect.

[0012] In a seventh aspect, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, can implement the method described in the first aspect.

[0013] Eighthly, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, can implement the method described in the second aspect.

[0014] In the quantitative magnetic susceptibility image generation method provided in this application embodiment, when it is necessary to generate a quantitative magnetic susceptibility map of the target object's brain, the T1 image of the target object's head, which can be directly obtained from an MRI scanner, can be preprocessed to obtain a T1 image of the target object's brain. This T1 image is then input into a pre-trained quantitative magnetic susceptibility image generation model. The model can fuse T1 image features with brain template features to output a quantitative magnetic susceptibility image of the target object's brain. By constructing a deep learning-based generation model, the widely available T1 image can be directly used to generate a quantitative magnetic susceptibility image, avoiding the limitations of specific scanning sequences and complex post-processing required to obtain such images. This significantly lowers the barrier to entry for quantitative magnetic susceptibility imaging technology and facilitates its transformation into routine clinical applications. In addition, since the brain contains many brain regions with different anatomical structures and functions, the magnetic susceptibility (iron deposition) change patterns of different brain regions in different diseases may vary. In this embodiment, fusing T1 image features with brain template features can enable targeted learning by region, ensuring that the generated quantitative magnetic susceptibility image has accurate quantitative values ​​in different brain regions and improving the accuracy of quantitative magnetic susceptibility image generation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for generating a quantitative magnetic susceptibility image, as provided in an embodiment of this application; Figure 2 A schematic diagram illustrating the principle of feature fusion provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the principle of a quantitative magnetic susceptibility image generation model provided in this application embodiment; Figure 4 A flowchart illustrating a training method for a quantitative magnetic susceptibility image generation model provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of an electronic device provided in this embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. In addition, the timing of the steps in the following method embodiments is only an example and not a strict limitation.

[0018] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to large language models or other models) comply with relevant laws and standards.

[0019] First, the terms or concepts involved in the embodiments of this application will be explained: Brain template: A digital map that divides the brain into regions (i.e., brain regions) with specific anatomical structures and / or functional definitions. Each brain region is typically labeled with a unique integer label.

[0020] Cross-attention mechanism: In this attention mechanism, the query vector (Q) and the key vector (K) and value vector (V) originate from different feature sources. By calculating attention weights across features, one feature focuses on another and extracts information, thereby achieving deep fusion between features.

[0021] Multi-head attention mechanism: A deep learning mechanism for modeling dependencies between sequences or features. Its core idea is to map input information to multiple subspaces in parallel through multiple sets of different linear transformations (i.e., multiple "heads"), and independently compute attention weights in each subspace. This allows the model to learn information simultaneously from different representation subspaces (e.g., different levels, different types of relationships), thus capturing more comprehensive and complex associations between features.

[0022] This application provides a method for generating quantitative magnetic susceptibility images. Utilizing widely available clinical T1 images (also known as T1-weighted images), a deep learning model deeply integrated with prior knowledge of brain anatomy directly generates brain region-specific quantitative magnetic susceptibility images. This avoids the limitations of specific scanning sequences and complex post-processing required to obtain quantitative magnetic susceptibility images, significantly lowering the barrier to entry for quantitative magnetic susceptibility imaging technology and facilitating its transition to routine clinical application. It also ensures that the generated quantitative magnetic susceptibility images have accurate quantitative values ​​for different brain regions, improving the accuracy of quantitative magnetic susceptibility image generation.

[0023] The method for generating quantitative magnetic susceptibility images provided in this application can be applied to devices for generating quantitative magnetic susceptibility images, such as computers, servers, personal computers, and medical devices.

[0024] The execution process of the quantitative magnetic susceptibility image generation method provided in this application embodiment is described in detail below with reference to the accompanying drawings. Optionally, this method can be... Figure 1 A flowchart illustrating a method for generating a quantitative magnetic susceptibility image provided in this application embodiment is shown below. Figure 1 As shown, the method includes the following steps: 101. Obtain the T1 image. The T1 image is obtained by scanning the head of the target object using a magnetic resonance imaging device.

[0025] 102. Preprocess the T1 image to obtain the preprocessed T1 image; the preprocessing includes skull removal.

[0026] 103. The preprocessed T1 image is input into the quantitative magnetic susceptibility image generation model, so that the quantitative magnetic susceptibility image generation model performs the following processing: feature extraction is performed on the preprocessed T1 image to obtain T1 features; the T1 features and brain template features are fused to obtain fused features, wherein the brain template features are obtained by extracting features from the brain template based on the feature extraction method of the preprocessed T1 image; the fused features are encoded and decoded to output the quantitative magnetic susceptibility image of the target object's brain; the quantitative magnetic susceptibility image generation model is a trained deep learning model.

[0027] In practical applications, T1 images are obtained by scanning the head of the target subject using an MRI scanner. T1 images are one of the most common and widely used structural imaging sequences in MRI, and are extensively acquired clinically to visualize the anatomical structures of the brain. Their signal intensity primarily reflects the longitudinal relaxation time of tissues, with clear contrast between gray matter, white matter, and cerebrospinal fluid.

[0028] Because MRI scanners scan the head of the target subject, the resulting T1 images may contain information unrelated to the brain's magnetic susceptibility distribution, such as structures like the skull and neck. Therefore, preprocessing includes, but is not limited to, craniectomy. Cranectomy refers to obtaining image regions representing the brain from the T1 images.

[0029] The aforementioned quantitative magnetic susceptibility image generation model is a pre-trained deep learning model. For example, the model's architecture could be an attention-based architecture (Transformer), a convolutional neural network, or a Mamba architecture, where the Transformer could be a sliding window (Swin) Transformer. Its internal processing flow follows steps 1031-1033 to achieve the mapping from the T1 image to the quantitative magnetic susceptibility image: Step 1031: Extract features from the preprocessed T1 image to obtain T1 features.

[0030] The preprocessed T1 image is downsampled and feature-encoded through operations such as convolution to gradually extract its multi-level and multi-scale visual features, ultimately resulting in a high-dimensional feature tensor, denoted as the T1 feature. Downsampling can also be referred to as downsampling.

[0031] In practical applications, quantitative magnetic susceptibility image generation models typically include brain template features. These brain template features are obtained by extracting features from the brain template using the same feature extraction method as in step 1031. This facilitates feature fusion between the two types of features.

[0032] Optionally, the brain template used in this application, such as the Anatomical Automatic Labeling (AAL) template, is used to provide prior knowledge of the brain's anatomical structure for the quantitative magnetic susceptibility image generation model.

[0033] In one possible embodiment, the brain template includes, but is not limited to, the AAL template.

[0034] Optionally, taking the brain template as the AAL template as an example, since the integer labels in the AAL template, such as 1, 2, 3, up to 90, have no inherent meaning, but when the AAL template is input into the quantitative magnetic susceptibility image generation model, the model may incorrectly interpret the magnitude of the label number F value as the importance ranking of brain regions, posing a risk of misleading the quantitative magnetic susceptibility image generation model. Therefore, a trainable parameter table can be pre-established. The pre-established parameter table contains the correspondence between each brain region and an N-dimensional vector. The N-dimensional vector corresponding to each brain region in the pre-established parameter table is the model parameter adjusted during the training of the initial quantitative magnetic susceptibility image generation model. Replacing the label values ​​of each brain region in the AAL template with the corresponding N-dimensional vectors in the pre-established parameter table yields the updated AAL template; where N is an integer greater than 1. Correspondingly, the brain template features are obtained by extracting features from the updated AAL template based on the feature extraction method of the preprocessed T1 image.

[0035] Optionally, the parameter table has dimensions [A, N], where A is the total number of brain regions. For example, for the AAL template in the MNI space, its dimensions are (91, 109, 91), with 90 brain regions, so A = 90. D is the embedding dimension. Each row in the parameter table is a D-dimensional vector, uniquely representing a brain region. During processing, each voxel of the brain template is traversed, and based on its integer label value, the corresponding D-dimensional vector is found in the parameter table and replaced. In this way, the AAL template with dimensions (91, 109, 91) is transformed into a feature map with dimensions (D, 91, 109, 91).

[0036] In this way, the model transforms anatomical prior knowledge from discrete labels into learnable continuous vector representations, avoiding misjudgments and allowing the model to automatically optimize the feature representations of each brain region during training.

[0037] It should be noted that it is also possible to choose not to perform downsampling and directly perform subsequent feature fusion at high resolution. However, this will increase the computational load and reduce processing efficiency.

[0038] In addition, depending on the computing resources and the need for feature details, different downsampling methods can be flexibly adopted, such as pooling operations, dilated convolutions, or deeper convolutional blocks.

[0039] Step 1032: Fuse the T1 features and brain template features to obtain fused features.

[0040] When performing feature fusion, attention-based feature fusion or other feature fusion methods can be used. The following is an example of using an attention-based feature fusion method: Linear projection of T1 features yields the T1 vector. Linear projection of brain template features yields the brain region vector.

[0041] Optionally, T1 features can be linearly projected through a linear layer, and correspondingly, brain template features can be linearly projected through a linear layer.

[0042] Subsequently, based on the multi-head attention mechanism, fused features were obtained according to the T1 vector and brain region vector.

[0043] In one possible implementation of feature fusion based on a multi-head attention mechanism, the T1 key vector, T1 value vector, and brain region query vector are input into the first multi-head attention layer to obtain fused features. The brain template features and T1 features are processed separately through linear layers to generate brain region query vectors for the brain template features, and T1 key vectors and T1 value vectors for the T1 features. These vectors are then input into the multi-head attention mechanism. Since the query vectors, key vectors, and value vectors originate from different sources, the attention mechanism models the association between the brain template features and the T1 features, fusing anatomical prior information from the brain template into the T1 features, thus completing feature fusion based on voxel-level correlation.

[0044] Another possible implementation of feature fusion based on multi-head attention mechanism achieves feature fusion through two stages. The first stage involves inputting the T1 key vector, T1 value vector, and brain region query vector into the first multi-head attention layer to obtain the first feature. This process is similar to the possible implementation mentioned above and will not be elaborated further here.

[0045] The second stage involves concatenating the T1 query vector with the brain region query vector to obtain the concatenated query vector. Then, the T1 key vector is concatenated with the brain region key vector to obtain the concatenated key vector. Finally, the T1 value vector is concatenated with the brain region value vector to obtain the concatenated value vector. These concatenated query vectors, key vectors, and value vectors are then input into the second multi-head attention layer to obtain the second feature.

[0046] It is evident that the execution of the first and second phases is not sequential; they are usually executed simultaneously.

[0047] Next, the first and second features are concatenated to obtain the fused features. The concatenation of the first and second features can be done along a dimension. Feature fusion is achieved through two different methods, realizing T1 generation based on brain template feature constraints and obtaining the fused features.

[0048] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the principle of feature fusion provided in an embodiment of this application. The AAL template feature corresponds to the brain template feature in the above embodiment. Figure 2 The left side of the middle section illustrates the fusion principle of the first stage, and the multi-head attention mechanism used corresponds to the first multi-head attention mechanism mentioned above. Figure 2 The right side of the middle section illustrates the fusion principle of the second stage, and the multi-head attention mechanism used corresponds to the second multi-head attention mechanism mentioned above.

[0049] Step 1033: Encode and decode the fusion features to output a quantitative magnetic susceptibility image of the target brain.

[0050] In this embodiment, the fused high-dimensional abstract features are progressively reconstructed into a quantitative magnetic susceptibility image. During the encoding and decoding process, a generative model architecture can be used, such as an encoder and decoder structure. The encoder typically consists of multiple downsampling blocks, used to further compress the fused features and extract higher-level semantic information. The decoder consists of multiple upsampling blocks, progressively restoring the spatial dimensions and converting the high-level semantic information back into the image pixel space.

[0051] For example, the encoder and decoder structure can be as follows: The encoder sequentially includes: a first Swing Transformer module, a first downsampling module, a second Swing Transformer module, and a second downsampling module. The encoder is connected to the decoder. The decoder sequentially includes: a third Swing Transformer module, a first upsampling module, a fourth Swing Transformer module, a second upsampling module, a fifth Swing Transformer module, a third upsampling module, a sixth Swing Transformer module, a fourth upsampling module, and a convolutional module. The fused features are input to the encoder, and the decoder outputs a quantitative magnetic susceptibility image.

[0052] In summary, the solution provided by the embodiments of this application allows for the generation of a quantitative magnetic susceptibility map of the target brain when it is necessary to generate such a map. This can be achieved by preprocessing the T1 images of the target head, which are directly obtainable from MRI equipment, to obtain a T1 image of the target brain. This T1 image is then input into a pre-trained quantitative magnetic susceptibility image generation model. The model fuses T1 image features with brain template features to output a quantitative magnetic susceptibility image of the target brain. By constructing a deep learning-based generation model, widely available T1 images can be directly used to generate quantitative magnetic susceptibility images, avoiding the limitations of specific scanning sequences and complex post-processing required to obtain such images. This significantly lowers the barrier to entry for quantitative magnetic susceptibility imaging technology and facilitates its transformation into routine clinical applications. In addition, since the brain contains many brain regions with different anatomical structures and functions, the magnetic susceptibility (iron deposition) change patterns of different brain regions in different diseases may vary. In this embodiment, fusing T1 image features with brain template features can enable targeted learning by region, ensuring that the generated quantitative magnetic susceptibility image has accurate quantitative values ​​in different brain regions and improving the accuracy of quantitative magnetic susceptibility image generation.

[0053] In one possible embodiment, step 102 can be implemented by the following steps 1021-1024: Step 1021: Redirect the T1 image to the reference space to obtain the redirected T1 image.

[0054] The T1 image is retargeted to a standard reference space, resulting in a retargeted T1 image. This reference space is the one used by the standard brain template. Retargeting the T1 image to this reference space facilitates subsequent spatial alignment with the brain template.

[0055] Optionally, the reference space can be the MNI space or the Talairach space.

[0056] Step 1022: Perform bias field correction processing on the redirected T1 image to obtain the corrected T1 image.

[0057] In magnetic resonance imaging (MRI), due to factors such as inhomogeneity in the radio frequency coil, image intensity exhibits slow spatial variations, known as the bias field, which may interfere with image analysis. Therefore, bias field correction is necessary for the redirected image to eliminate intensity differences caused by magnetic field inhomogeneity. Correcting this inhomogeneity improves the uniformity of image quality, which is beneficial for the model to learn stable features.

[0058] Step 1023: Perform craniotomy on the corrected T1 image to obtain a brain T1 image.

[0059] By precisely segmenting and removing non-brain tissue structures such as the scalp and skull, only the brain parenchyma is preserved. T1 images of the brain are obtained.

[0060] Step 1024: Register the brain T1 image to the reference space to obtain the preprocessed T1 image.

[0061] The T1 images of the brain obtained after skull removal are registered to a reference space to ensure that the brain structure of the target subject is spatially aligned with the brain template, thereby ensuring that the anatomical labels extracted from the template can be accurately mapped to the corresponding locations in the target subject's brain.

[0062] In one possible embodiment, after registration in step 1024, the registered brain T1 image can be intensity normalized. For example, the Min-Max normalization method can be used to linearly scale the image pixel values ​​to the [0,1] interval to obtain the preprocessed T1 image. Normalization can accelerate the training convergence process of deep learning models and improve the numerical stability of the models.

[0063] In this embodiment, through the above preprocessing process, the final preprocessed T1 image is an image with uniform intensity in a standard space, containing only brain tissue and with a standardized intensity range, which can be directly used for model input.

[0064] The following example illustrates the processing procedure of the quantitative magnetic susceptibility image generation model described above.

[0065] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the principle of a quantitative magnetic susceptibility image generation model provided in an embodiment of this application. The quantitative magnetic susceptibility image generation model includes a brain template fusion unit, an encoder, and a decoder connected sequentially. In practical applications, T1 images are input into the quantitative magnetic susceptibility image generation model and first processed by the brain template fusion unit. The brain template fusion unit includes a brain template fusion module, which can be the aforementioned... Figure 2 The principle structure is shown. The trainable parameter table corresponds to the pre-established parameter table in the above embodiments.

[0066] In some scenarios, it is necessary to train a quantitative magnetic susceptibility image generation model. The following example... Figure 4 The illustrated embodiment provides a training method for a quantitative magnetic susceptibility image generation model. This training method can be executed alone or in combination with the quantitative magnetic susceptibility image generation method of the above embodiment. When executed in combination with the quantitative magnetic susceptibility image generation method of the above embodiment, it can be performed before step 101. Figure 4 The training method for the quantitative magnetic susceptibility image generation model is shown.

[0067] Figure 4 A flowchart illustrating a training method for a quantitative magnetic susceptibility image generation model provided in this application embodiment is shown below. Figure 4 As shown, the method includes the following steps: 401. Obtain multiple sample pairs, each of which contains a T1 image sample and a QSM data sample obtained by scanning the head of the same object using a magnetic resonance imaging device.

[0068] 402. Preprocess the T1 image sample to obtain the preprocessed T1 image sample.

[0069] 403. Preprocess the QSM data samples to obtain quantitative magnetic susceptibility image labels.

[0070] 404. Train an initial quantitative magnetic susceptibility image generation model based on the preprocessed T1 image samples and brain template; the initial quantitative magnetic susceptibility model is used for the following processing: using the same feature extraction method, extract features from the preprocessed T1 image samples and brain template respectively to obtain T1 feature samples and brain template feature samples; fuse the T1 feature samples and brain template feature samples to obtain fused feature samples; encode and decode the fused feature samples to output the initial quantitative magnetic susceptibility image corresponding to the T1 image samples.

[0071] 405. Obtain the loss value based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample.

[0072] 406. Iteratively train the initial quantitative magnetic susceptibility image generation model based on the loss value until the initial quantitative magnetic susceptibility image generation model converges, thus obtaining the quantitative magnetic susceptibility image generation model.

[0073] This embodiment involves similar concepts and steps to the above embodiments, and will not be repeated here.

[0074] In this embodiment, T1 image samples and QSM data samples obtained by scanning the heads of different objects using a magnetic resonance imaging device are acquired. That is, T1 image samples and QSM data samples corresponding to the same object head are a sample pair.

[0075] Because MRI scanners scan the heads of different subjects, the resulting T1-weighted image samples may contain information unrelated to the brain's magnetic susceptibility distribution, such as structures like the skull and neck. Therefore, preprocessing of T1-weighted image samples includes, but is not limited to, craniotomy. Craniotomy refers to extracting image regions representing the brain from the T1-weighted image samples.

[0076] In one possible embodiment, step 402 can be implemented by the following steps 4021-4024: Step 4021: Redirect the T1 image sample to the reference space to obtain the redirected T1 image sample.

[0077] The T1 image samples are retargeted to a standard reference space to obtain retargeted T1 image samples. This reference space is the one used by the standard brain template. Retargeting the T1 image samples to the reference space facilitates subsequent spatial alignment with the brain template.

[0078] Optionally, the reference space can be the MNI space or the Talairach space.

[0079] Step 4022: Perform bias field correction processing on the redirected T1 image sample to obtain the corrected T1 image sample.

[0080] In magnetic resonance imaging (MRI), due to factors such as inhomogeneity in the radio frequency coil, image intensity exhibits slow spatial variations, known as the bias field, which may interfere with image analysis. Therefore, bias field correction is necessary for the redirected image to eliminate intensity differences caused by magnetic field inhomogeneity. Correcting this inhomogeneity improves the uniformity of image quality, which is beneficial for the model to learn stable features.

[0081] Step 4023: Perform skull stripping on the corrected T1 image samples to obtain brain T1 image samples.

[0082] By precisely segmenting and removing non-brain tissue structures such as the scalp and skull, only the brain parenchyma is preserved. T1 images of the brain are obtained.

[0083] Step 4024: Register the brain T1 image samples to the reference space to obtain preprocessed T1 image samples.

[0084] The T1 images of the brain obtained after skull removal are registered to a reference space. This ensures that the brain structure of the target subject is spatially aligned with the brain template, thereby guaranteeing that the anatomical labels extracted from the template can be accurately mapped to the corresponding locations in the target subject's brain.

[0085] In one possible embodiment, after registration in step 4024, the registered brain T1 image samples can be intensity normalized. For example, the Min-Max normalization method can be used to linearly scale the image pixel values ​​to the [0,1] interval to obtain preprocessed T1 image samples. Normalization can accelerate the training convergence process of deep learning models and improve the numerical stability of the models.

[0086] Through the above preprocessing process, the final preprocessed T1 image sample is an image with uniform intensity in standard space, containing only brain tissue and with a standardized intensity range, which can be directly used as model input.

[0087] QSM data samples are obtained from magnetic resonance imaging (MRI) scans; therefore, preprocessing is required to obtain quantitative magnetic susceptibility image labels. Optionally, the preprocessing of QSM data samples can include: performing multi-echo phase synthesis and phase unwrapping; removing the background field, extracranial tissue, and large-scale inhomogeneous magnetic fields, while preserving the tissue field; calculating the magnetic susceptibility to obtain quantitative magnetic susceptibility image labels. Alternatively, the Dipole Inversion method can be used to calculate the magnetic susceptibility.

[0088] The aforementioned initial quantitative susceptibility image generation model is a deep learning model. For example, the model architecture could be based on an attention mechanism, a convolutional neural network, or a Mamba architecture, where the Transformer could be a SwinTransformer. During training, its internal processing follows steps 4041-4043 to achieve the mapping from T1 image samples to the initial quantitative susceptibility image: Step 4041: Extract features from the preprocessed T1 image samples to obtain T1 feature samples.

[0089] The preprocessed T1 image samples are downsampled and feature-encoded through operations such as convolution to gradually extract multi-level and multi-scale visual features, ultimately resulting in a high-dimensional feature tensor, denoted as the T1 feature sample. Downsampling can also be referred to as downsampling.

[0090] In practical applications, quantitative magnetic susceptibility image generation models typically include brain template feature samples. These brain template feature samples are obtained by extracting features from the brain template using the same feature extraction method as in step 4041. This facilitates feature fusion between the two types of features.

[0091] Optionally, the brain template used in this application, such as the Anatomical Automatic Labeling (AAL) template, is used to provide prior knowledge of the brain's anatomical structure for the quantitative magnetic susceptibility image generation model.

[0092] In one possible embodiment, the brain template includes an AAL template.

[0093] Optionally, taking the brain template as the AAL template as an example, since the integer labels in the AAL template, such as 1, 2, 3, up to 90, have no inherent meaning, but when the AAL template is input into the quantitative magnetic susceptibility image generation model, the model may incorrectly interpret the magnitude of the label number F value as the importance ranking of brain regions, posing a risk of misleading the quantitative magnetic susceptibility image generation model. Therefore, a trainable parameter table can be pre-established. The pre-established parameter table contains the correspondence between each brain region and an N-dimensional vector. The N-dimensional vector corresponding to each brain region in the pre-established parameter table is the model parameter adjusted during the training of the initial quantitative magnetic susceptibility image generation model. The label values ​​of each brain region in the AAL template are replaced with the corresponding N-dimensional vectors in the pre-established parameter table to obtain the updated AAL template; where N is an integer greater than 1. The brain template feature samples are obtained by extracting features from the updated AAL template using the same feature extraction method as in step 4041.

[0094] Optionally, the parameter table has dimensions [A, N], where A is the total number of brain regions. For example, for the AAL template in the MNI space, its dimensions are (91, 109, 91), with 90 brain regions, so A = 90. D is the embedding dimension. Each row in the parameter table is a D-dimensional vector, uniquely representing a brain region. During processing, each voxel of the brain template is traversed, and based on its integer label value, the corresponding D-dimensional vector is found in the parameter table and replaced. In this way, the AAL template with dimensions (91, 109, 91) is transformed into a feature map with dimensions (D, 91, 109, 91).

[0095] In this way, the model transforms anatomical prior knowledge from discrete labels into learnable continuous vector representations, avoiding misjudgments and allowing the model to automatically optimize the feature representations of each brain region during training.

[0096] It should be noted that it is also possible to choose not to perform downsampling and directly perform subsequent feature fusion at high resolution. However, this will increase the computational load and reduce processing efficiency.

[0097] In addition, depending on the computing resources and the need for feature details, different downsampling methods can be flexibly adopted, such as pooling operations, dilated convolutions, or deeper convolutional blocks.

[0098] Step 4042: Perform feature fusion on the T1 feature samples and the brain template feature samples to obtain fused features.

[0099] In this embodiment, feature fusion can be performed using an attention-based mechanism, or other feature fusion methods can be used. The following is an example of using an attention-based mechanism for feature fusion: Linear projection of the T1 feature samples yields the T1 vector. Linear projection of the brain template feature samples yields the brain region vector.

[0100] Optionally, T1 feature samples can be linearly projected using a linear layer, and correspondingly, brain template feature samples can be linearly projected using a linear layer.

[0101] Subsequently, based on the multi-head attention mechanism, fused features were obtained according to the T1 vector and brain region vector.

[0102] In one possible implementation of feature fusion based on a multi-head attention mechanism, the T1 key vector, T1 value vector, and brain region query vector are input into the first multi-head attention layer to obtain fused features. The brain template feature samples and T1 feature samples are processed separately through a linear layer to generate brain region query vectors for the brain template feature samples, as well as T1 key vectors and T1 value vectors for the T1 feature samples. These vectors are then input into the multi-head attention mechanism. Since the three vectors originate from different sources, the attention mechanism models the association between the brain template feature samples and the T1 feature samples, fusing anatomical prior information from the brain template into the T1 feature samples, thus completing feature fusion based on voxel-level correlation.

[0103] Another possible implementation of feature fusion based on multi-head attention mechanism achieves feature fusion through two stages. The two stages are described below.

[0104] The first stage involves inputting the T1 key vector, T1 value vector, and brain region query vector into the first multi-head attention layer to obtain the first feature. This process is similar to that described in one possible implementation above and will not be repeated here.

[0105] The second stage involves concatenating the T1 query vector with the brain region query vector to obtain the concatenated query vector. Then, the T1 key vector is concatenated with the brain region key vector to obtain the concatenated key vector. Finally, the T1 value vector is concatenated with the brain region value vector to obtain the concatenated value vector. These concatenated query vectors, key vectors, and value vectors are then input into the second multi-head attention layer to obtain the second feature.

[0106] It is evident that the execution of the first and second phases is not sequential; they are usually executed simultaneously.

[0107] Next, the first and second features are concatenated to obtain the fused features. The concatenation of the first and second features can be done along a dimension. Feature fusion is achieved through two different methods, realizing T1 generation based on brain template feature sample constraints, and obtaining the fused features.

[0108] Step 4043: Encode and decode the fused features to output the initial quantitative magnetic susceptibility image.

[0109] In this embodiment, the fused, high-dimensional abstract features are reconstructed into an initial quantitative magnetic susceptibility image. During the encoding and decoding process, a generative model architecture can be used, such as an encoder and decoder structure. The encoder typically consists of multiple downsampling blocks, used to further compress the fused features and extract higher-level semantic information. The decoder consists of multiple upsampling blocks, progressively restoring the spatial dimensions and converting the high-level semantic information back into the image pixel space.

[0110] For example, the encoder and decoder structure can be as follows: The encoder sequentially includes: a first SwinTransformer module, a first downsampling module, a second SwinTransformer module, and a second downsampling module. The encoder is connected to the decoder. The decoder sequentially includes: a third SwinTransformer module, a first upsampling module, a fourth SwinTransformer module, a second upsampling module, a fifth SwinTransformer module, a third upsampling module, a sixth SwinTransformer module, a fourth upsampling module, and a convolutional module. The fused features are input to the encoder, and the decoder outputs an initial quantitative magnetic susceptibility image.

[0111] Subsequently, based on the initial quantitative susceptibility image corresponding to the T1 image sample and the corresponding quantitative susceptibility image label, a loss value is obtained. The initial quantitative susceptibility image generation model is iteratively trained based on the loss value until it converges, thus obtaining the quantitative susceptibility image generation model.

[0112] Optionally, the loss value can be calculated as follows: Calculate the structural loss value based on the initial quantitative susceptibility image and the corresponding quantitative susceptibility image label for the T1 image sample. Determine the voxel regression loss value based on the preset weights of each brain region, the initial quantitative susceptibility image, and the corresponding quantitative susceptibility image label for the T1 image sample. Add the structural loss value and the voxel regression loss value to obtain the final loss value.

[0113] Furthermore, the loss value can be calculated according to the following formula (1).

[0114] Formula (1) in, The value represents the voxel regression loss. This is the structural loss value. Optional, is the voxel regression loss value. It can be obtained through the following formula (2).

[0115] Formula (2) in, The magnetic susceptibility of voxel x in the initial quantitative magnetic susceptibility map image. Let |x| be the magnetic susceptibility of voxel x in the quantitative magnetic susceptibility image label, and |·| be the L1 loss. The weights of voxel x. This indicates a summation.

[0116] For the 90 brain regions of the AAL template, a 90-dimensional weight can be generated. Each value corresponds to a weight for a brain region. When a voxel in the initial quantitative susceptibility image belongs to a certain brain region, the weight value corresponding to that brain region is extracted and weighted accordingly. This brain region constraint ensures consistency between the initial quantitative susceptibility image and the initial quantitative susceptibility image label. If all brain regions are considered to contribute equally during the generation process, all weights can be set to the same value, for example, 1. If certain brain regions are considered more important, such as those of greater interest in certain diseases, the meaning of brain regions in the AAL template can be reviewed, and weights can be set according to the importance of the brain regions. That is, the more important the brain region, the higher its weight, so that the model pays more attention to the generation quality of that brain region, achieving generation based on anatomical information constraints.

[0117] This constraint is used to maintain the overall contrast and structural consistency between the generated and true magnetic susceptibility maps. Optionally, it can be calculated using 1-SSIM, where SSIM is the structural similarity coefficient.

[0118] In summary, the scheme provided in the embodiments of this application obtains multiple sample pairs, each containing a T1 image sample and a QSM data sample obtained by scanning the head of the same object using a magnetic resonance imaging (MRI) device. The T1 image samples are preprocessed to obtain preprocessed T1 image samples. The QSM data samples are preprocessed to obtain quantitative magnetic susceptibility image labels. An initial quantitative magnetic susceptibility image generation model is trained based on the preprocessed T1 image samples and brain region templates. The initial quantitative magnetic susceptibility model is used for the following processing: feature extraction is performed on the preprocessed T1 image samples and brain region templates using the same feature extraction method to obtain T1 feature samples and brain region template feature samples; the T1 feature samples and brain region template feature samples are fused to obtain fused feature samples. The fused feature samples are encoded and decoded to output the initial quantitative magnetic susceptibility image corresponding to the T1 image samples. A loss value is obtained based on the initial quantitative magnetic susceptibility image and the quantitative magnetic susceptibility image labels corresponding to the T1 image samples. The initial quantitative magnetic susceptibility image generation model is iteratively trained based on the loss value until it converges, thus obtaining the quantitative magnetic susceptibility image generation model. By constructing a deep learning-based generation model, it is possible to directly utilize widely available T1 image samples to generate quantitative magnetic susceptibility images, avoiding the limitations of specific scanning sequences and complex post-processing required to obtain quantitative magnetic susceptibility images. This significantly lowers the barrier to entry for quantitative magnetic susceptibility imaging technology and facilitates its transformation into routine clinical applications. Furthermore, since the brain contains numerous brain regions with diverse anatomical structures and functions, the magnetic susceptibility (iron deposition) variation patterns in different brain regions may differ in different diseases. In this embodiment, fusing T1 image sample features with brain template feature samples allows for targeted learning in different regions, ensuring that the generated quantitative magnetic susceptibility images have accurate quantitative values ​​in different brain regions and improving the accuracy of quantitative magnetic susceptibility image generation.

[0119] In some embodiments, during the training of the initial quantitative susceptibility image generation model, multiple sample pairs can be divided into a training set, a validation set, and a test set. The training set is used to implement the above-described process of training the initial quantitative susceptibility image generation model. During training, the validation set is used to evaluate the model performance, and the optimal model parameters are selected to obtain the quantitative susceptibility image generation model. During testing, the test set is input into the model with these parameters to evaluate the model's ability to actually generate quantitative susceptibility maps.

[0120] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 5 As shown, in practice, this electronic device includes a memory 21 and a processor 22.

[0121] Memory 21 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0122] The processor 22, coupled to the memory 21, is used to execute the computer program in the memory 21 to implement the quantitative magnetic susceptibility image generation method or the quantitative magnetic susceptibility image generation model training method provided in the foregoing embodiments.

[0123] Furthermore, such as Figure 5 As shown, the electronic device also includes other components such as a communication component 23, a display 24, a power supply component 25, and an audio component 26. Figure 5 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 5 The components shown are as follows. The electronic device in this embodiment can be a terminal device such as a desktop computer, laptop computer, smartphone, or IoT device, or a server device such as a conventional server, cloud server, or server array.

[0124] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0126] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0127] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0128] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0129] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a processor, the processor is able to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. In addition, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, so that the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device can be implemented as a means to implement the corresponding functions in the above method embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating a quantitative magnetic susceptibility image, characterized in that, The method includes: Acquire a T1 image, which is obtained by scanning the head of the target object using a magnetic resonance imaging device; The T1 image is preprocessed to obtain a preprocessed T1 image; the preprocessing includes skull removal. The preprocessed T1 image is input into the quantitative magnetic susceptibility image generation model, which then performs the following processing: feature extraction is performed on the preprocessed T1 image to obtain T1 features; the T1 features and brain template features are fused to obtain fused features, wherein the brain template features are obtained by extracting features from the brain template based on the feature extraction method performed on the preprocessed T1 image; the fused features are encoded and decoded to output the quantitative magnetic susceptibility image of the target object's brain; the quantitative magnetic susceptibility image generation model is a trained deep learning model.

2. The method according to claim 1, characterized in that, The preprocessing of the T1 image to obtain the preprocessed T1 image includes: The T1 image is redirected to the reference space to obtain the redirected T1 image; The redirected T1 image is subjected to bias field correction processing to obtain the corrected T1 image; The corrected T1 image was subjected to skull stripping to obtain a brain T1 image; The brain T1 image is registered to the reference space to obtain the preprocessed T1 image.

3. The method according to claim 1, characterized in that, The process of fusing the T1 features and brain template features to obtain fused features includes: The T1 vector is obtained by linearly projecting the T1 feature. Linear projection of the brain template features yields brain region vectors; Based on the multi-head attention mechanism, fusion features are obtained according to the T1 vector and the brain region vector.

4. The method according to claim 3, characterized in that, The T1 vector includes: a T1 key vector and a T1 value vector; the brain region vector includes: a brain region query vector; the fusion feature obtained based on the multi-head attention mechanism, according to the T1 vector and the brain region vector, includes: The T1 key vector, the T1 value vector, and the brain region query vector are input into the first multi-head attention layer to obtain fused features.

5. The method according to claim 3, characterized in that, The T1 vector includes: a T1 query vector, a T1 key vector, and a T1 value vector; the brain region vector includes: a brain region query vector, a brain region key vector, and a brain region value vector; the fusion feature obtained based on the multi-head attention mechanism, according to the T1 vector and the brain region vector, includes: The T1 key vector, the T1 value vector, and the brain region query vector are input into the first multi-head attention layer to obtain the first feature; The T1 query vector is concatenated with the brain region query vector to obtain the concatenated query vector; The T1 key vector is concatenated with the brain region key vector to obtain the concatenated key vector; The T1 value vector is concatenated with the brain region value vector to obtain the concatenated value vector; The concatenated query vector, the concatenated key vector, and the concatenated value vector are input into the second multi-head attention layer to obtain the second feature; The first feature and the second feature are concatenated to obtain the fused feature.

6. The method according to claim 1, characterized in that, The method further includes: Multiple sample pairs were acquired, each containing a T1 image sample obtained by scanning the head of the same object using a magnetic resonance imaging device and a QSM data sample. The T1 image samples are preprocessed to obtain preprocessed T1 image samples; The QSM data samples are preprocessed to obtain quantitative magnetic susceptibility image labels; An initial quantitative magnetic susceptibility image generation model is trained based on the preprocessed T1 image samples and the brain template. The initial quantitative magnetic susceptibility model is used to perform the following processing: extracting features from the preprocessed T1 image samples and the brain template using the same feature extraction method to obtain T1 feature samples and brain template feature samples; fusing the T1 feature samples and brain template feature samples to obtain fused feature samples; encoding and decoding the fused feature samples to output the initial quantitative magnetic susceptibility image corresponding to the T1 image samples. Based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample, the loss function is obtained; The initial quantitative magnetic susceptibility image generation model is iteratively optimized based on the loss function until it converges, thus obtaining the quantitative magnetic susceptibility image generation model.

7. A training method for a quantitative magnetic susceptibility image generation model, characterized in that, The method includes: Multiple sample pairs are acquired, each of which contains a T1 image sample and a QSM data sample obtained by scanning the head of the same object using a magnetic resonance imaging device. The T1 image samples are preprocessed to obtain preprocessed T1 image samples; The QSM data samples are preprocessed to obtain quantitative magnetic susceptibility image labels; An initial quantitative magnetic susceptibility image generation model is trained based on the preprocessed T1 image samples and the brain template. The initial quantitative magnetic susceptibility model is used to perform the following processing: extracting features from the preprocessed T1 image samples and the brain template using the same feature extraction method to obtain T1 feature samples and brain template feature samples; fusing the T1 feature samples and brain template feature samples to obtain fused feature samples; encoding and decoding the fused feature samples to output the initial quantitative magnetic susceptibility image corresponding to the T1 image samples. Based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample, the loss value is obtained; The initial quantitative magnetic susceptibility image generation model is iteratively trained based on the loss value until it converges, thus obtaining the quantitative magnetic susceptibility image generation model.

8. The method according to claim 6 or 7, characterized in that, The brain template includes an AAL template; the feature extraction is performed on the preprocessed T1 image sample and the brain template using the same feature extraction method to obtain T1 feature samples and brain template feature samples, including: The label values ​​of each brain region in the anatomically labeled AAL template are replaced with the corresponding D-dimensional vectors in the pre-established parameter table to obtain the updated AAL template; where D is an integer greater than 1; the pre-established parameter table contains the correspondence between each brain region and the D-dimensional vector; the N-dimensional vectors corresponding to each brain region in the pre-established parameter table are model parameters that are adjusted during the training of the initial quantitative magnetic susceptibility image generation model. Using the same downsampling method, feature extraction is performed on the preprocessed T1 image samples and the updated AAL template to obtain T1 feature samples and brain template feature samples.

9. The method according to claim 6 or 7, characterized in that, The loss value is obtained based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample, including: The structural loss value is calculated based on the initial quantitative magnetic susceptibility image corresponding to the T1 image sample and the quantitative magnetic susceptibility image label corresponding to the T1 image sample. Based on the preset weights of each brain region, the initial quantitative susceptibility image corresponding to the T1 image sample, and the quantitative susceptibility image label corresponding to the T1 image sample, the voxel regression loss value is determined. The loss value is obtained by adding the structural loss value and the voxel regression loss value.

10. A computer program product, characterized in that, include: A computer program, when executed by a processor of an electronic device, causes the processor to perform the method as described in any one of claims 1 to 9.

Images