Magnetic resonance imaging modality conversion method and system based on wavelet transform and perceptual state space modeling

Abstract

The application relates to the technical field of image processing, in particular to a magnetic resonance imaging (MRI) modality conversion method and system based on wavelet transformation and a perception state space modeling, aiming to solve the problem of realizing MRI modality conversion with consideration of efficiency, modality capture accuracy and input flexibility. The method comprises the following steps: constructing an input feature tensor according to the number of source modalities of MRI images; extracting initial features by an input feature mapping module; performing multi-level discrete wavelet decomposition by a wavelet transformation module to extract high-frequency and low-frequency information; performing layer-by-layer semantic feature coding by a residual encoder module; performing cross fusion and attention enhancement of multi-scale coding features and high-frequency information by a multi-scale cross efficient attention module; performing state space modeling and step-by-step decoding reconstruction of enhanced features and low-frequency information by a state space decoder module; and generating target modality MRI images by an output head module. The MRI modality conversion method and system have high efficiency, high accuracy and input flexibility.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a magnetic resonance imaging mode conversion method and system based on wavelet transform and perceptual state space modeling. Background Technology

[0002] Magnetic Resonance Imaging (MRI), with its superior imaging capabilities of human soft tissues, can clearly present the anatomical structure and physiological and pathological characteristics of organs and tissues, and has been widely used in clinical disease diagnosis, disease assessment, and medical research. In actual clinical imaging, MRI employs various imaging modalities. Based on their unique imaging principles, these modalities differ significantly in imaging contrast, tissue response characteristics, and information emphasis. Images from different modalities provide complementary imaging information support for clinical diagnosis.

[0003] However, in practical clinical applications, limitations such as imaging time, patient condition, medical equipment configuration, and examination costs often make it difficult to simultaneously acquire all the modal MRI images required by the patient. For example, some critically ill patients cannot tolerate prolonged multimodal scanning, and primary healthcare institutions may be unable to perform specific modal MRI examinations due to equipment limitations. Against this backdrop, modality conversion technology, which generates target modal MRI images based on existing available MRI modal images, can effectively compensate for the lack of image information and provide more comprehensive support for clinical diagnosis. Therefore, it has gradually attracted widespread attention in the medical and related technological fields.

[0004] Currently, most existing MRI modality conversion methods rely on deep convolutional neural networks or generative models to achieve feature mapping and image conversion between different modalities. These methods typically construct a network model by stacking multiple convolutional layers to extract local features from the input MRI image and perform pixel-level conversion between modalities based on these extracted features, thus achieving basic modality conversion functionality to a certain extent. However, practical verification has shown that existing MRI modality conversion methods still have many technical shortcomings, making it difficult to fully meet the needs of actual clinical applications: First, in pursuit of higher conversion accuracy, existing methods often require continuously increasing network depth and model parameter scale, which significantly increases the computational complexity of the model and significantly limits inference efficiency, making it impossible to achieve rapid modality conversion and difficult to adapt to the image processing speed requirements in clinical scenarios, creating a technical contradiction between conversion accuracy and computational efficiency; Second, in the feature modeling process, existing methods mostly use a uniform feature representation method to extract and process input image features, lacking explicit differentiation and targeted processing of different spatial frequency information (including high-frequency detail information and low-frequency structural information) in MRI images, resulting in the model's inability to accurately and specifically handle different spatial frequency information (including high-frequency detail information and low-frequency structural information). The methods used for modal conversion are insufficient in their ability to perceive and capture differences in contrast between modalities, tissue edge contours, and fine structures. As a result, the generated target modal images are prone to problems such as structural blurring and contrast distortion, which affects the clinical diagnostic value of the images. Thirdly, some existing methods are highly dependent on the input modality in their model structure design. They are usually designed specifically for a single input modality or a fixed combination of multiple modalities. The adaptability of the models is poor. When the number or type of input modalities changes (such as from a single modal input to a dual modal input, or changing to a different source modality type), the existing models are often difficult to adapt directly. The applicability and scalability are greatly limited, and they cannot meet the diverse single-modal or multi-modal MRI data conversion needs in different clinical scenarios.

[0005] In summary, there is an urgent need for an MRI modality conversion method that can effectively perceive and adapt to intermodal contrast differences while ensuring computational efficiency, and has good input flexibility. This would improve the image quality, structural consistency, and detail fidelity of the modality conversion results, enhance the practical application value of the method, and better meet the needs of clinical diagnosis and medical research. Summary of the Invention

[0006] To address the aforementioned technical problems in the prior art, namely, to achieve MRI modality conversion that balances efficiency, modality capture accuracy, and input flexibility, this application provides an MRI modality conversion method and system based on wavelet transform and perceptual state space modeling.

[0007] In a first aspect of this application, an MRI modality conversion method based on wavelet transform and perceptual state space modeling is provided, comprising:

[0008] The source modal MRI images are used to construct the input feature tensor according to the number of modalities;

[0009] The input feature mapping module is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer and a max pooling layer connected in sequence.

[0010] The input feature tensor is subjected to multi-level discrete wavelet decomposition and convolution mapping using a wavelet transform module to extract the high-frequency and low-frequency information of the source modality MRI image.

[0011] The initial features are downsampled and semantic features are encoded layer by layer using a residual encoder module to obtain multi-scale encoded features.

[0012] A multi-scale cross-efficiency attention module is used to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information at the corresponding scales to obtain enhanced features;

[0013] The state-space decoder module is used to perform state-space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features;

[0014] The high-dimensional features are mapped to the target space using the output head module to generate a target modal MRI image.

[0015] In a second aspect of this application, an MRI modality conversion system based on wavelet transform and perceptual state space modeling is provided, comprising:

[0016] The input processing module is used to construct an input feature tensor from the source modal MRI images according to the number of modalities;

[0017] The input feature mapping module is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer connected in sequence.

[0018] The wavelet transform module is used to perform multi-level discrete wavelet decomposition and convolution mapping on the input feature tensor to extract the high-frequency and low-frequency information of the source modality MRI image.

[0019] The residual encoder module is used to perform layer-by-layer downsampling and semantic feature encoding on the initial features to obtain multi-scale encoded features;

[0020] A multi-scale cross-efficiency attention module is used to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information of the corresponding scale to obtain enhanced features;

[0021] The state-space decoder module is used to perform state-space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features;

[0022] The output head module is used to map the high-dimensional features to the target space to generate a target modal MRI image.

[0023] The MRI modality conversion method and system based on wavelet transform and perceptual state space modeling provided in this application decomposes the input MRI image into high-frequency detail components and low-frequency structural components at different scales through multi-level discrete wavelet transform technology. Subsequently, differentiated processing and precise fusion are performed on the characteristics of the two types of components, thereby selectively preserving and reconstructing high-frequency detail information such as texture and edges that are important for clinical diagnosis, while ensuring the overall consistency of core low-frequency structures such as organ contours. This effectively avoids problems such as structural blurring and detail loss that are prone to occur in traditional methods, thus significantly improving the generation of target modal images. Overall quality was significantly improved, enhancing the model's ability to perceive and fuse key modal features, improving the modeling accuracy of complex nonlinear mapping relationships, and ensuring the clinical application value of images. By designing a multi-scale cross-efficiency attention module, the deep semantic features extracted by the residual encoder and the corresponding high-frequency detail features obtained through wavelet decomposition were deeply cross-fused in both spatial and channel dimensions, achieving adaptive feature enhancement. Through this mechanism, the model can dynamically focus on anatomical regions crucial for MRI modality transformation, precisely strengthening the representation ability of discriminative features while effectively suppressing the interference of redundant information, thereby improving image quality. The model's accuracy in modeling complex nonlinear mapping relationships between modalities ensures the accuracy and effectiveness of the target modal features after conversion. In the decoding and reconstruction stage, a perceptual state space modeling module is introduced. This module, through parallel spatial and channel state space models, can efficiently model global long-range dependencies within feature maps. This allows the overall contextual information of the input image to be fully referenced and utilized during image reconstruction, avoiding the limitations of traditional convolutional neural networks in terms of limited receptive field and difficulty in capturing long-distance feature associations. This effectively improves the model's ability to model global contextual information and long-range dependencies, overcoming the inherent limitations of traditional convolutional neural networks. The invention overcomes limitations by ultimately generating a target modal MRI image with superior global consistency in anatomical structural integrity and contrast rationality. Through modular structural design, it can flexibly adapt to different input configurations without modifying the core structure, meeting different clinical application scenarios such as single-modal input or dual-modal input. It effectively solves the problems of strong dependence on input modality and limited applicability and scalability of traditional methods, and can meet diverse clinical MRI modality conversion needs. It also has high computational efficiency, optimizes resource utilization, and balances image conversion accuracy and inference speed, significantly improving the practical value and promotion prospects of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the MRI mode conversion method based on wavelet decomposition and perceptual state space modeling provided in this embodiment of the application.

[0026] Figure 2 A schematic diagram of the overall framework of the MRI modality conversion method based on wavelet decomposition and perceptual state space modeling provided in the embodiments of this application;

[0027] Figure 3 This is a schematic diagram of the wavelet transform module provided in an embodiment of this application;

[0028] Figure 4 This is a schematic diagram of the structure of the multi-scale cross-efficiency attention module provided in the embodiments of this application;

[0029] Figure 5 This is a schematic diagram of the structure of the state space decoder provided in the embodiments of this application;

[0030] Figure 6 This is a structural block diagram of one embodiment of the MRI modality conversion system based on wavelet transform and perceptual state space modeling provided in this application. Detailed Implementation

[0031] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0032] The following will describe in detail, with reference to the accompanying drawings, an MRI modality conversion method based on wavelet transform and perceptual state space modeling according to an embodiment of this application. Figure 1 This is a schematic diagram of the MRI mode conversion method based on wavelet decomposition and perceptual state space modeling provided in an embodiment of this application. Figure 1 As shown, the MRI mode conversion method based on wavelet decomposition and perceptual state space modeling in the first embodiment of this application includes:

[0033] Step S101: The source modal MRI images are used to construct an input feature tensor according to the number of modalities;

[0034] Step S102: The input feature mapping module is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer and a max pooling layer connected in sequence.

[0035] Step S103: Use the wavelet transform module to perform multi-level discrete wavelet decomposition and convolution mapping on the input feature tensor to extract the high-frequency and low-frequency information of the source modality MRI image;

[0036] Step S104: Use the residual encoder module to perform layer-by-layer downsampling and semantic feature encoding on the initial features to obtain multi-scale encoded features;

[0037] Step S105: Use a multi-scale cross-efficiency attention module to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information of the corresponding scale to obtain enhanced features;

[0038] Step S106: Use the state space decoder module to perform state space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features;

[0039] Step S107: Use the output head module to map the high-dimensional features to the target space to generate a target modal MRI image.

[0040] Specifically, in step S101, different channel configurations are used to organize the source modal MRI images according to the different number of input modalities: if the input source modal MRI image is a single-source modal MRI image, the source modal MRI image is input as a single-channel grayscale image, the number of channels is set to 1, and a single-channel input feature tensor is generated to characterize the MRI image information under a single modality; if the input source modal MRI image is a dual-source modal MRI image... and One of the dual-source modal MRI images Mapped to the R channel of the input tensor The G channel of the input tensor is mapped to the input tensor, and the B channel is set to a constant 1 to generate a three-channel input feature tensor. This allows the bimodal information to be organized in a unified manner in the form of a color image while maintaining spatial consistency.

[0041] Specifically, in step S102, the input feature mapping module performs channel mapping and initial feature extraction using the input feature tensor as input to obtain initial features. The input feature mapping module can be composed of an initial 7×7 convolutional layer, a batch normalization layer, a ReLU activation function layer, and a max pooling layer connected in sequence.

[0042] Specifically, in step 103, the wavelet transform module employs a three-level discrete wavelet decomposition and convolution mapping, using the input feature tensor as the input to the wavelet transform module, to extract the high-frequency and low-frequency information of the source modality MRI image. Corresponding to the number of levels in the discrete wavelet decomposition, the high-frequency and low-frequency information are each in three layers, with the number of layers corresponding one-to-one with the number of levels in the three-layer discrete wavelet decomposition. Figure 3 As shown, the first-level discrete wavelet transform corresponds to the first layer of low-frequency information and the first layer of high-frequency information, the second-level discrete wavelet transform corresponds to the second layer of low-frequency information and the second layer of high-frequency information, and the third-level discrete wavelet transform corresponds to the third layer of low-frequency information and the third layer of high-frequency information.

[0043] Specifically, in step S103, the wavelet transform module can employ a two-dimensional separable filtering method, performing low-pass filtering and high-pass filtering in the row and column directions respectively, combined with downsampling, to decompose the input step by step. Specifically, the wavelet transform module obtains four sub-band features from the input feature tensor through three-level discrete wavelet decomposition, denoted as HH, HL, LH, and LL, where LL is the low-frequency sub-band, and HH, HL, and LH are the high-frequency sub-bands.

[0044] ,

[0045] ,

[0046] ,

[0047] ,

[0048] in, Let be the pixel value of the two-dimensional MRI image represented by the input feature tensor at coordinates (m,n), where i is the coordinate index of the sub-band feature in the row direction, j is the coordinate index of the sub-band feature in the column direction, m is the pixel position index of the input feature tensor in the row direction, n is the pixel position index of the input feature tensor in the column direction, h is the low-pass filter coefficient used to extract approximate information, g is the high-pass filter coefficient used to extract detail information, and the low-frequency sub-band LL and high-frequency sub-bands HH, HL, and LH can respectively represent the downsampling results after applying different combinations of low-pass and high-pass filters to the input in the row and column directions. In step S103, during the three-level decomposition process, only the low-frequency sub-band LL obtained from the previous level of decomposition is subjected to discrete wavelet transform, thereby forming a progressively recursive multi-scale decomposition structure, that is, the th The discrete wavelet decomposition of the first level is divided into the second level. Discrete wavelet decomposition of the low-frequency subband:

[0049] ,

[0050] DWT stands for Discrete Wavelet Decomposition. Through the three-level discrete wavelet transform operation described above, the source modal MRI image is decomposed into multiple sub-band features at different scales. Then, the sub-bands obtained from each level of decomposition are further segmented. Specifically, the LL sub-band obtained from each level of decomposition is selected as the high-frequency sub-band for that level, and the HH, HL, and LH sub-bands obtained from each level of decomposition are uniformly used as the low-frequency sub-bands for that level, thus completing the high- and low-frequency classification of the wavelet sub-bands. After completing the high- and low-frequency sub-band segmentation, convolutional mapping is performed on the high-frequency and low-frequency sub-bands at each level. For example, 3×3 convolution is used to perform feature mapping on the high-frequency and low-frequency sub-bands at each level, resulting in three layers of high-frequency information and three layers of low-frequency information. The number of channels for each layer of low-frequency information is set to be consistent with the output channel number of the corresponding layer of the residual encoder module, and the number of channels for each layer of high-frequency information is uniformly set to 16.

[0051] Specifically, in step S104, the residual encoder module can adopt a ResNet34 network structure and include four cascaded residual encoder layers, such as... Figure 2 As shown. The residual encoder module takes the initial features output in step S102 as input and performs feature encoding processing sequentially through multiple residual blocks, i.e., cascaded four residual encoder layers, according to the ResNet34 network hierarchy. Each residual encoder layer adopts the residual connection method of identity mapping and is composed of convolution operations and nonlinear mapping. Specifically, in step S104, the nth layer encoded features output by the nth residual encoder layer are downsampled through a convolution operation with a stride setting and used as the input of the (n+1)th residual encoder layer, n=1,2,3. For example, the second layer encoded features output by the second residual encoder layer are downsampled through a convolution operation with a stride setting and used as the input of the third residual encoder layer, and so on. The input of the first residual encoder layer is the initial features output by the input feature mapping module, thereby obtaining the encoded features output by each layer of the residual encoder, such as... Figure 2 As shown. The coding features of each layer are the multi-scale coding features. The spatial resolution of each layer of coding features decreases layer by layer according to the ResNet34 structure. Specifically, except for the first residual encoder, the convolution stride of each residual encoder is set to 2, so that the width and height of the output coding features are reduced to half of the previous layer, thereby achieving a 1 / 2 ratio decrease in spatial resolution layer by layer. The number of channels of each layer of coding features is adjusted according to the predefined configuration of ResNet34. Specifically, as the spatial resolution decreases, the number of channels of the coding features increases proportionally. For example, starting from 64 channels in the first residual encoder, the subsequent residual encoders are adjusted to 128, 256 and 512 channels respectively to compensate for the loss of feature representation caused by the reduction of spatial information.

[0052] Specifically, in step S105, the multi-scale cross-efficiency attention module performs cross-fusion and attention enhancement on the multi-scale coding features and the high-frequency features of the corresponding scale. That is, it performs cross-fusion and attention enhancement on the coding features of each layer and the high-frequency features of the corresponding layer. Specifically, it performs cross-fusion and attention enhancement on the coding features of the first layer and the high-frequency features of the first layer, and so on. However, the coding features output by the fourth residual encoder layer are not used as input to the multi-scale cross-efficiency attention module and do not participate in the calculation of the multi-scale cross-efficiency attention module.

[0053] Specifically, in step S105, the multi-scale cross-efficiency attention module performs the same processing procedure on the encoded features of each layer and the high-frequency features of the corresponding layer number, such as... Figure 2 As shown, for each layer of encoded features and the corresponding high-frequency features of the layer number, such as Figure 4 As shown, the multi-scale cross-efficiency attention module can include the following processing steps:

[0054] The coding features of each layer other than the coding features output by the 4th residual encoder layer and the high-frequency information of the corresponding scale are added by channel, and then channel mapping is performed by 1×1 convolution to obtain intermediate features;

[0055] After performing layer normalization on the intermediate features of each layer, two parallel x-axis branches, a first x-axis branch and a second x-axis branch, are constructed along the x-axis direction of the intermediate features. The first x-axis branch is sequentially subjected to 1×7 convolution and average pooling operations to obtain the pooling result of the first x-axis branch. Similarly, the second x-axis branch is sequentially subjected to 1×15 dilated convolution and max pooling operations to obtain the pooling result of the second x-axis branch. The pooling result of the first x-axis branch is then input into a mapping structure, which can specifically be a dimension transpose (interchanging the spatial dimension and the channel dimension), followed by a 1×1 convolution to achieve linearity. The combined mapping, after passing through the mapping structure and being processed by the Sigmoid activation function, is multiplied element-wise with the pooling result of the first x-axis branch to obtain the preprocessed result of the first x-axis branch. The pooling result of the second x-axis branch is then input into the mapping structure, processed by the activation function, and multiplied element-wise with the pooling result of the second x-axis branch to obtain the preprocessed result of the second x-axis branch. The preprocessed results of the first and second x-axis branches are then added and fused, and finally, a 1×1 convolution is performed to obtain the intermediate attention features in the x-axis direction of each layer.

[0056] Simultaneously, after performing layer normalization on the intermediate features of each layer, two parallel y-axis branches, a first y-axis branch and a second y-axis branch, are constructed along the y-axis direction of the intermediate features. The first y-axis branch is sequentially subjected to 7×1 convolution and average pooling operations to obtain the first y-axis branch pooling result. Similarly, the second y-axis branch is sequentially subjected to 15×1 dilated convolution and max pooling operations to obtain the second y-axis branch pooling result. The first y-axis branch pooling result is input into the aforementioned mapping structure and processed by the aforementioned activation function, then multiplied element-wise with the first y-axis branch pooling result to obtain the first y-axis branch preprocessing result. The second y-axis branch pooling result is then input into the aforementioned mapping structure and processed by the aforementioned activation function, then multiplied element-wise with the second y-axis branch pooling result to obtain the second y-axis branch preprocessing result. Finally, the first and second y-axis branch preprocessing results are added and fused, and then subjected to 1×1 convolution to obtain the y-axis direction attention intermediate features of each layer.

[0057] Linear mapping is performed on the intermediate features of the attention along the x-axis of each layer to generate the corresponding keys. and value Linear mapping is performed on the corresponding y-axis attention intermediate features to generate the corresponding x-axis attention intermediate features. Furthermore, a linear mapping is performed on the intermediate features of the attention along the y-axis of each layer to generate the corresponding key. and value And perform linear mapping on the corresponding x-axis direction attention intermediate features to generate the corresponding y-axis direction attention intermediate features. ;Will and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the x-axis direction attention output features of each layer, and then... and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the attention output features in the y-axis direction of each layer;

[0058] The x-axis and y-axis attention output features of each layer are mapped through 1×1 convolutions and then added element-wise with the intermediate features to obtain the enhanced features of each layer.

[0059] Specifically, in step S106, corresponding to the structure of the residual encoder module, the state space decoder module includes four cascaded state space decoder layers, such as... Figure 2 As shown, the state-space decoder module performs state-space modeling and step-by-step decoding and reconstruction on the enhanced features output by the multi-scale cross-efficiency attention module at each layer and the low-frequency features at the corresponding scale (i.e., the corresponding layer number), to obtain high-dimensional features. Specifically, as... Figure 2 and Figure 5 As shown, the multi-scale cross-efficiency attention module can include the following processing steps:

[0060] The fourth-layer state-space encoder layer is used to perform state-space modeling and step-by-step decoding reconstruction on the fourth-layer residual encoder layer output fourth-layer encoded features, outputting fourth-layer decoding stage features. This includes: sequentially passing the fourth-layer encoded features through a convolutional mapping structure consisting of two 3×3 convolutions, batch normalization, and ReLU activation functions to obtain intermediate mapped features; upsampling the intermediate mapped features to obtain fourth-layer decoding stage features. The upsampling can be performed using bilinear interpolation, with the upsampling factor set to 2.

[0061] The (n+1)th layer decoding stage features output from the (n+1)th layer state space encoder layer, the low-frequency information corresponding to the scale (i.e., the corresponding layer number) of the nth layer state space encoder layer, and the enhancement features corresponding to the scale (i.e., the corresponding layer number) are added together and used as the nth layer spatial encoding input of the nth layer state space encoder layer, where n=3,2,1. For example, the 3rd layer spatial encoding input of the 3rd layer state space encoder layer is the sum of the 4th layer decoding stage features output from the 4th layer state space encoder layer, the 3rd layer low-frequency information, and the 3rd layer enhancement features.

[0062] Using the nth-layer state-space encoder layer to perform state-space modeling and step-by-step decoding and reconstruction on the nth-layer spatial encoded input, the output of the nth-layer decoding stage features includes:

[0063] The spatial encoding input of the nth layer is sequentially subjected to 1×1 convolution, layer normalization and GELU activation to obtain the state space modeling input features;

[0064] The input features of the state space modeling are input into the dual-branch state space, that is, input into the spatial state space model branch and the contrast state space model branch respectively for parallel modeling, to obtain the output features of the spatial state space model and the output features of the contrast state space model. The parallel spatial state space model branch and the contrast state space model branch are used to model the spatial long-range dependency of the features and capture the contrast information between the features, respectively.

[0065] The output features of the spatial state spatial model are processed based on feature maps. Specifically, a 3×3 convolution operation is performed on the output features of the spatial state spatial model to extract local spatial features. Then, layer normalization is applied to stabilize the numerical distribution. Next, a non-linear transformation is performed through a GELU activation function layer. Then, a 1×1 convolution operation is performed to integrate the features between channels. Finally, a Sigmoid activation function layer is used to map the feature values ​​between 0 and 1 to generate attention weight coefficients. Then, the attention weight coefficients obtained from the above processing are multiplied element-wise with the output features of the contrast state spatial model at the corresponding spatial positions and channel dimensions. The result of the multiplication operation is then added to the output features of the spatial state spatial model using residuals to obtain the fused features.

[0066] The fused features are simultaneously input into three parallel convolutional branches for parallel convolution. Specifically, the first convolutional branch performs 5×1 and 1×5 convolutions sequentially, the second convolutional branch performs 3×3 convolutions sequentially, and the third convolutional branch performs 1×7 and 7×1 convolutions sequentially. The convolutional outputs of each convolutional branch are batch normalized, and the features of each convolutional branch after batch normalization are added element-wise at the corresponding positions to obtain the branch fused features. After GELU activation processing, the branch fused features are sequentially processed by 1×1 convolution for channel mapping and upsampling to obtain the features of the nth layer decoding stage.

[0067] Finally, the first-level decoding stage features output by the first-level state space encoder layer are used as the high-dimensional features output by the state space decoder module.

[0068] The process of inputting the state-space modeling input features into the spatial state-space model branch and the contrast state-space model branch for parallel modeling to obtain the spatial state-space model output features and the contrast state-space model output features may include the following steps:

[0069] The spatial state space model branch flattens the state space modeling input features according to a preset spatial scanning order. The spatial scanning order adopts a spiral scanning order. The flattening process is as follows: starting from the upper left corner of the two-dimensional feature map corresponding to the state space modeling input features, it traverses horizontally to the right to the boundary, and then performs a spiral scan layer by layer in a clockwise direction until the entire two-dimensional feature map is traversed. After the traversal is completed, the features at the corresponding pixel positions are arranged sequentially according to the spatial scanning order to form a one-dimensional sequence. The one-dimensional sequence is then input into the spatial state space model branch for state update and output calculation. According to the index relationship corresponding to the spatial scanning order, the sequence output by the spatial state space model branch is rearranged and restored to a two-dimensional feature representation to obtain the spatial state space model output features.

[0070] The contrast state space model branch flattens the input features of the state space modeling in a channel-first manner, unfolding each channel feature into a one-dimensional sequence, and constructing forward and reverse sequences based on the one-dimensional sequence. The forward and reverse sequences are then input into the contrast state space model branch for state update and output calculation, respectively. The output of the contrast state space model branch corresponding to the reverse sequence is then reversed and added to the corresponding position of the output of the contrast state space model branch corresponding to the forward sequence, restoring it to a two-dimensional feature representation, thus obtaining the output features of the contrast state space model.

[0071] Specifically, in step S107, the output head module maps the high-dimensional features to the target space to generate the target modal MRI image. Step S107 may include the following processing steps: performing feature refinement on the high-dimensional features, which includes sequentially performing 3×3 convolution, batch normalization, and LeakyReLU activation to obtain refined features; performing upsampling on the refined features and then performing output mapping to obtain mapped features, wherein the upsampling adopts bilinear interpolation and the upsampling factor is set to 2; the output mapping includes sequentially performing 3×3 convolution, LeakyReLU activation, and 1×1 convolution, which maps the number of channels of the refined features to the number of channels corresponding to the target modal MRI image; and performing hyperbolic tangent function processing on the mapped features to generate the target modal MRI image.

[0072] In one possible implementation, the MRI modality conversion method based on wavelet transform and perceptual state space modeling provided in this application embodiment may further include: training an input feature mapping module, a wavelet transform module, a residual encoder module, a multi-scale cross-efficiency attention module, a state space decoder module, and an output head module using a multi-constraint joint loss function, wherein the joint loss function may include:

[0073] Pixel-level L1 loss function is used to constrain the numerical precision of an image, as shown below:

[0074] ,

[0075] in, This represents the total number of pixels in the image. The ground truth image of the target modality MRI image represents the first... The value of each pixel. This indicates the number of elements in the generated target modality MRI image (predicted image). The value of each pixel;

[0076] Multi-scale structural similarity loss functions are used to maintain the consistency of anatomical structures in MRI images, as shown below:

[0077] ,

[0078] in, This indicates the total number of assessment scales. These represent the brightness, contrast, and structure comparison factors calculated at different resolution scales, respectively. The exponential weights of each component are given. and These represent the ground truth image and the predicted image, respectively.

[0079] Gradient consistency loss function is used to enhance the ability to preserve edges and details, as shown below:

[0080] ,

[0081] in, This represents the gradient operator of the image in the horizontal direction. This represents the gradient operator of the image in the vertical direction. represents the L1 norm, which enhances the reconstruction of tissue boundaries and minute structures by constraining the rate of change of pixels; E() represents the mathematical expectation.

[0082] The contrast statistical loss function is used to constrain global and local grayscale contrast characteristics, as shown below:

[0083] ,

[0084] in, The global standard deviation of an image is used to measure the overall contrast range. and This represents the local mean value calculated using a sliding window (such as mean filtering). and This indicates the local texture features after removing the low-frequency background.

[0085] The MRI modality conversion method based on wavelet transform and perceptual state space modeling provided in this application decomposes high- and low-frequency components of an image through multi-level discrete wavelet transform. After differential fusion processing, it can accurately preserve high-frequency details such as texture and edges, and ensure the consistency of low-frequency structures such as organ contours, avoiding the structural blurring and detail loss problems of traditional methods. At the same time, it enhances the perceptual fusion capability of modal features and the accuracy of nonlinear mapping modeling, significantly improving the overall quality and clinical value of the generated image. Through a multi-scale cross-efficient attention module, it deeply fuses the semantic features extracted by the residual encoder with the corresponding scale high-frequency features, dynamically focuses on key anatomical regions, strengthens discriminative features, suppresses redundant information, and improves intermodal features. The nonlinear mapping modeling accuracy ensures the accuracy and effectiveness of target modal features, improving the accuracy of modal feature fusion and the effectiveness of target features. A perceptual state space modeling module is introduced in the decoding and reconstruction stage, efficiently modeling long-range dependencies through parallel and channel state space models. This fully utilizes global contextual information, avoids the limited receptive field problem of traditional CNNs, and generates target images with better consistency in anatomical structure and contrast, enhancing global modeling capabilities. Modular design flexibly adapts to clinical scenarios such as single / dual modal input, solving the problems of insufficient adaptability and scalability of traditional methods. The overall architecture is computationally efficient and resource-efficient, balancing conversion accuracy and inference speed, significantly improving practical value and prospects for widespread application.

[0086] A second aspect of this application provides an MRI modality conversion system based on wavelet transform and perceptual state space modeling. Figure 6 This paper illustrates a structural block diagram of one embodiment of the MRI modality conversion system based on wavelet transform and perceptual state space modeling provided in this application. Figure 6 As shown, the MRI modality conversion system based on wavelet transform and perceptual state space modeling according to the second embodiment of this application includes:

[0087] Input processing module 601 is used to construct an input feature tensor from the source modal MRI images according to the number of modalities;

[0088] The input feature mapping module 602 is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer and a max pooling layer connected in sequence.

[0089] Wavelet transform module 603 is used to perform multi-level discrete wavelet decomposition and convolution mapping on the input feature tensor to extract high-frequency and low-frequency information of the source modality MRI image;

[0090] The residual encoder module 604 is used to perform layer-by-layer downsampling and semantic feature encoding on the initial features to obtain multi-scale encoded features;

[0091] The multi-scale cross-efficiency attention module 605 is used to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information of the corresponding scale to obtain enhanced features;

[0092] The state space decoder module 606 is used to perform state space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features;

[0093] The output head module 607 is used to map the high-dimensional features to the target space to generate a target modal MRI image.

[0094] Specifically, the input processing module is configured to: if the source modal MRI image is a single-source modal MRI image, input the source modal MRI image as a single-channel grayscale image, set the number of channels to 1, and generate a single-channel input feature tensor; if the source modal MRI image is a dual-source modal MRI image... and One of the dual-source modal MRI images Mapped to the R channel of the input tensor Map the input tensor to the G channel and set the B channel to a constant 1 to generate a three-channel input feature tensor.

[0095] Specifically, in the wavelet transform module, the multi-level discrete wavelet decomposition and convolution mapping is a three-level discrete wavelet decomposition and convolution mapping. The high-frequency information and the low-frequency information are each in three layers, and the number of layers of the high-frequency information and the low-frequency information corresponds one-to-one with the number of levels of the three-layer discrete wavelet decomposition. The wavelet transform module is specifically used for:

[0096] The input feature tensor is subjected to three-level discrete wavelet decomposition to obtain four sub-band features, denoted as HH, HL, LH, and LL, where LL is the low-frequency sub-band, and HH, HL, and LH are the high-frequency sub-bands. The discrete wavelet decomposition adopts a two-dimensional separable filtering method. The discrete wavelet decomposition of the first level is divided into the second level. Discrete wavelet decomposition of the low-frequency subband: DWT stands for Discrete Wavelet Decomposition;

[0097] The LL subbands obtained from each level of decomposition are selected as the high-frequency subbands of each level, and the HH, HL and LH subbands obtained from each level of decomposition are uniformly used as the low-frequency subbands of each level.

[0098] Convolutional mapping is performed on the high-frequency subbands and low-frequency subbands at each level to obtain the three layers of high-frequency information and the three layers of low-frequency information.

[0099] In one possible implementation, the residual encoder module is a ResNet34 network structure and includes four cascaded residual encoder layers. Specifically, the residual encoder module is used for:

[0100] The nth layer coding feature output by the nth residual encoder layer is downsampled by a convolution operation with stride setting and used as the input of the (n+1)th residual encoder layer, where n=1,2,3. The input of the 1st residual encoder layer is the initial feature, thus obtaining the coding features of each layer output by each residual encoder layer.

[0101] The encoded features of each layer are the multi-scale encoded features. The spatial resolution of each layer of encoded features decreases layer by layer according to the ResNet34 structure. Specifically, except for the residual encoder of the first layer, each residual encoder reduces the width and height of the output encoded features to half that of the previous layer by setting the convolution stride to 2, thereby achieving a 1 / 2 ratio decrease in spatial resolution layer by layer. The number of channels of each layer of encoded features is adjusted according to the predefined configuration of ResNet34. Specifically, as the spatial resolution decreases, the number of channels of each layer of encoded features increases proportionally. For example, starting from 64 channels in the first layer of encoded features, the subsequent layers of encoded features are adjusted to 128, 256 and 512 channels respectively to compensate for the loss of feature representation caused by the reduction of spatial information.

[0102] Specifically, the multi-scale cross-efficiency attention module is used for:

[0103] The coding features of each layer other than the coding features output by the 4th residual encoder layer and the high-frequency information of the corresponding scale are added by channel, and then channel mapping is performed by 1×1 convolution to obtain intermediate features. The corresponding scale is the corresponding layer number.

[0104] After performing layer normalization on the intermediate features of each layer, two parallel first x-axis branches and second x-axis branches are constructed along the x-axis direction of the intermediate features. The first x-axis branches and second x-axis branches are x-axis branches. The first x-axis branch is sequentially subjected to 1×7 convolution and average pooling operations to obtain the first x-axis branch pooling result, and the second x-axis branch is sequentially subjected to 1×15 dilated convolution and max pooling operations to obtain the second x-axis branch pooling result. The first x-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the first x-axis branch pooling result to obtain the first x-axis branch preprocessing result. The second x-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the second x-axis branch pooling result to obtain the second x-axis branch preprocessing result. The first x-axis branch preprocessing result and the second x-axis branch preprocessing result are added and fused, and then subjected to 1×1 convolution to obtain the x-axis direction attention intermediate features of each layer.

[0105] After performing layer normalization on the intermediate features of each layer, two parallel y-axis branches, a first y-axis branch and a second y-axis branch, are constructed along the y-axis direction of the intermediate features. The first y-axis branch and the second y-axis branch are y-axis branches. The first y-axis branch is sequentially subjected to 7×1 convolution and average pooling operations to obtain the first y-axis branch pooling result. The second y-axis branch is sequentially subjected to 15×1 dilated convolution and max pooling operations to obtain the second y-axis branch pooling result. The first y-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the first y-axis branch pooling result to obtain the first y-axis branch preprocessing result. The second y-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the second y-axis branch pooling result to obtain the second y-axis branch preprocessing result. The first y-axis branch preprocessing result and the second y-axis branch preprocessing result are added and fused, and then subjected to 1×1 convolution to obtain the y-axis direction attention intermediate features of each layer.

[0106] Linear mapping is performed on the intermediate features of the x-axis attention in each layer to generate corresponding keys. and value Linear mapping is performed on the corresponding y-axis direction attention intermediate features to generate a corresponding x-axis direction attention intermediate feature. Furthermore, a linear mapping is performed on the intermediate features of the y-axis attention in each layer to generate corresponding keys. and value Linear mapping is performed on the corresponding x-axis direction attention intermediate features to generate the corresponding y-axis direction attention intermediate features. ;Will and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the x-axis direction attention output features of each layer, and then... and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the attention output features in the y-axis direction of each layer;

[0107] The x-axis attention output features and y-axis attention output features of each layer are mapped by 1×1 convolution, and then added element-wise with the intermediate features to obtain the enhanced features of each layer.

[0108] In one possible implementation, the state space decoder module includes four cascaded state space decoder layers. Specifically, the state space decoder module is used for:

[0109] The fourth layer state-space encoder layer performs state-space modeling and step-by-step decoding and reconstruction on the fourth layer residual encoder layer outputting the fourth layer decoding stage features. This includes: sequentially passing the fourth layer encoded features through a convolutional mapping structure consisting of two 3×3 convolutions, batch normalization, and ReLU activation functions to obtain intermediate mapping features; and upsampling the intermediate mapping features to obtain the fourth layer decoding stage features. The upsampling adopts bilinear interpolation with an upsampling factor of 2.

[0110] The (n+1)th layer decoding stage features output by the (n+1)th layer state space encoder layer, the low-frequency information of the corresponding scale of the nth layer state space encoder layer, and the enhancement features of the corresponding scale are added together and used as the nth layer spatial encoding input of the nth layer state space encoder layer, where n=3,2,1, and the corresponding scale is the corresponding layer number.

[0111] The nth-layer state-space encoder layer is used to perform state-space modeling and step-by-step decoding and reconstruction on the nth-layer spatial encoded input, outputting the features of the nth-layer decoding stage, including:

[0112] The nth layer spatial encoding input is sequentially subjected to 1×1 convolution, layer normalization and GELU activation to obtain the state space modeling input features;

[0113] The state space modeling input features are respectively input to the spatial state space model branch and the contrast state space model branch for parallel modeling to obtain the spatial state space model output features and the contrast state space model output features.

[0114] After performing feature map-based processing on the output features of the spatial state spatial model, specifically, 3×3 convolution operations are performed sequentially on the output features of the spatial state spatial model to extract local spatial features. Then, layer normalization is performed to stabilize the numerical distribution. Next, nonlinear transformation is performed through the GELU activation function layer. Then, 1×1 convolution operation is performed to integrate the features between channels. Finally, the Sigmoid activation function layer is used to map the feature values ​​between 0 and 1 to generate attention weight coefficients. Then, the processed attention weight coefficients are multiplied element-wise with the output features of the contrast state spatial model at the corresponding spatial positions and channel dimensions. Finally, the residuals are added to the output features of the spatial state spatial model to obtain the fused features.

[0115] The fused features are simultaneously input into three parallel convolutional branches for parallel convolution. Specifically, the first convolutional branch performs 5×1 and 1×5 convolutions sequentially, the second convolutional branch performs 3×3 convolutions sequentially, and the third convolutional branch performs 1×7 and 7×1 convolutions sequentially. Batch normalization is performed on the convolutional outputs of each convolutional branch, and the features of each convolutional branch after batch normalization are added element-wise at corresponding positions to obtain the branch fused features. After GELU activation processing is applied to the branch fused features, channel mapping and upsampling processing are performed sequentially using 1×1 convolutions to obtain the features of the nth layer decoding stage.

[0116] Finally, the high-dimensional features are the features of the first-layer decoding stage output by the first-layer state space encoder layer.

[0117] Specifically, the spatial state space model branch is used to: flatten the state space modeling input features according to a preset spatial scanning order, wherein the spatial scanning order adopts a spiral scanning order; the flattening process is as follows: starting from the upper left corner of the two-dimensional feature map corresponding to the state space modeling input features, traversing horizontally to the right to the boundary, and then spirally scanning inward layer by layer in a clockwise direction until the entire two-dimensional feature map is traversed; after the traversal is completed, the features at the corresponding pixel positions are arranged sequentially according to the spatial scanning order to form a one-dimensional sequence, and the one-dimensional sequence is input into the spatial state space model branch for state update and output calculation; according to the index relationship corresponding to the spatial scanning order, the sequence output by the spatial state space model branch is rearranged and restored to a two-dimensional feature representation to obtain the output features of the spatial state space model;

[0118] Specifically, the contrast state space model branch is used to: flatten the input features of the state space modeling in a channel-first manner, unfold each channel feature into a one-dimensional sequence, and construct a forward arrangement sequence and a reverse arrangement sequence based on the one-dimensional sequence; input the forward arrangement sequence and the reverse arrangement sequence into the contrast state space model branch for state update and output calculation, reverse the output corresponding to the reverse arrangement sequence, and add it to the corresponding position of the output corresponding to the forward arrangement sequence to restore it to a two-dimensional feature representation, thereby obtaining the output features of the contrast state space model.

[0119] Specifically, the output header module is used for:

[0120] The high-dimensional features are subjected to feature refinement processing, which includes sequential 3×3 convolution operation, batch normalization processing, and LeakyReLU activation operation to obtain refined features;

[0121] After upsampling the refined features, output mapping is performed to obtain the mapped features. The upsampling adopts bilinear interpolation and the upsampling factor is set to 2. The output mapping includes sequential 3×3 convolution operation, LeakyReLU activation operation and 1×1 convolution operation. The output mapping maps the number of channels of the refined features to the number of channels corresponding to the target modality MRI image.

[0122] The mapping features are processed using a hyperbolic tangent function to generate the target modality MRI image.

[0123] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be referred to the corresponding processes in the foregoing method embodiments, and therefore will not be repeated here.

[0124] The MRI modality conversion system based on wavelet transform and perceptual state space modeling provided in this application decomposes high- and low-frequency components of an image through multi-level discrete wavelet transform and differentially fuses them. This accurately preserves details, stabilizes structure, avoids the shortcomings of traditional methods, enhances feature perception and modeling accuracy, and improves the quality and clinical value of generated images. The multi-scale cross-efficiency attention module deeply fuses semantics and high-frequency features, focuses on key regions, and suppresses redundancy, ensuring the effectiveness of target modality features. The perceptual state space module is introduced in the decoding stage to efficiently model long-range dependencies, overcome the limitations of traditional CNNs, and improve the global consistency of the image. The modular design adapts to single / dual-modal input scenarios, and the overall architecture is computationally efficient, balancing accuracy and speed, significantly improving practical value and prospects for promotion.

[0125] It should be noted that the MRI modality conversion system based on wavelet transform and perceptual state space modeling provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of this application can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of this application are only for distinguishing the various modules or steps and are not considered as an improper limitation of this application.

[0126] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0127] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0128] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0129] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0130] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0131] 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, and should all be included within the protection scope of this application.

Claims

1. A magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling, characterized in that, include: The source modal magnetic resonance imaging images are used to construct the input feature tensor according to the number of modes; The input feature mapping module is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer and a max pooling layer connected in sequence. The input feature tensor is subjected to multi-level discrete wavelet decomposition and convolution mapping using a wavelet transform module to extract the high-frequency and low-frequency information of the source modal magnetic resonance imaging image. The initial features are downsampled and semantic features are encoded layer by layer using a residual encoder module to obtain multi-scale encoded features. A multi-scale cross-efficiency attention module is used to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information at the corresponding scales to obtain enhanced features; The state-space decoder module is used to perform state-space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features; The high-dimensional features are mapped to the target space using the output head module to generate a target modal magnetic resonance imaging image.

2. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 1, characterized in that, The step of constructing an input feature vector from the source modal magnetic resonance imaging images according to the number of modalities includes: If the source modal magnetic resonance imaging image is a single source modal magnetic resonance imaging image, the source modal magnetic resonance imaging image is used as a single-channel grayscale image input, the number of channels is set to 1, and a single-channel input feature tensor is generated. If the source modal magnetic resonance imaging image is a dual-source modal magnetic resonance imaging image and One of the dual-source modal magnetic resonance imaging images Mapped to the R channel of the input tensor Map the input tensor to the G channel and set the B channel to a constant 1 to generate a three-channel input feature tensor.

3. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 1, characterized in that, The multi-level discrete wavelet decomposition and convolution mapping is a three-level discrete wavelet decomposition and convolution mapping, with three layers for both the high-frequency and low-frequency information. The number of layers for the high-frequency and low-frequency information corresponds one-to-one with the level of the three-layer discrete wavelet decomposition. The step of using a wavelet transform module to perform multi-level discrete wavelet decomposition and convolution mapping on the input feature tensor to extract the high-frequency and low-frequency information of the source modal magnetic resonance imaging image includes: The input feature tensor is decomposed into four sub-band features by three-level discrete wavelet decomposition, denoted as HH, HL, LH and LL, where LL is the low-frequency sub-band and HH, HL and LH are the high-frequency sub-bands. ， ， ， ， in, Let be the pixel value of the two-dimensional magnetic resonance imaging image represented by the input feature tensor at coordinates (m,n), i be the coordinate index of the sub-band feature in the row direction, j be the coordinate index of the sub-band feature in the column direction, m be the pixel position index of the input feature tensor in the row direction, n be the pixel position index of the input feature tensor in the column direction, h be the low-pass filter coefficient used to extract approximate information, and g be the high-pass filter coefficient used to extract detail information. The discrete wavelet decomposition adopts a two-dimensional separable filtering method. The discrete wavelet decomposition of the first level into the second level is as follows: Discrete wavelet decomposition of the low-frequency subband: DWT stands for Discrete Wavelet Decomposition; The LL subbands obtained from each level of decomposition are selected as the high-frequency subbands of each level, and the HH, HL and LH subbands obtained from each level of decomposition are uniformly used as the low-frequency subbands of each level. Convolutional mapping is performed on the high-frequency subbands and low-frequency subbands at each level to obtain the three layers of high-frequency information and the three layers of low-frequency information.

4. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 3, characterized in that, The residual encoder module is a ResNet34 network structure and includes four cascaded residual encoder layers. The residual encoder module is used to perform layer-by-layer downsampling and semantic feature encoding on the initial features to obtain multi-scale encoded features, including: The nth layer coding feature output by the nth residual encoder layer is downsampled by a convolution operation with stride setting and used as the input of the (n+1)th residual encoder layer, where n=1,2,3. The input of the 1st residual encoder layer is the initial feature, thus obtaining the coding features of each layer output by each residual encoder layer. The coding features of each layer are the multi-scale coding features, wherein the spatial resolution of each coding feature decreases by 1 / 2 layer by layer according to the ResNet34 structure setting, and the number of channels of each coding feature increases by 2 times layer by layer according to the predefined configuration of ResNet34.

5. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 4, characterized in that, The method employs a multi-scale cross-efficiency attention module to cross-fuse and enhance the attention of the multi-scale encoded features and the corresponding high-frequency information at each scale, resulting in enhanced features, including: The coding features of each layer other than the coding features output by the 4th residual encoder layer and the high-frequency information of the corresponding scale are added by channel, and then channel mapping is performed by 1×1 convolution to obtain intermediate features. The corresponding scale is the corresponding layer number. After performing layer normalization on the intermediate features of each layer, two parallel first x-axis branches and second x-axis branches are constructed along the x-axis direction of the intermediate features. The first x-axis branches and second x-axis branches are x-axis branches. The first x-axis branch is sequentially subjected to 1×7 convolution and average pooling operations to obtain the first x-axis branch pooling result, and the second x-axis branch is sequentially subjected to 1×15 dilated convolution and max pooling operations to obtain the second x-axis branch pooling result. The first x-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the first x-axis branch pooling result to obtain the first x-axis branch preprocessing result. The second x-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the second x-axis branch pooling result to obtain the second x-axis branch preprocessing result. The first x-axis branch preprocessing result and the second x-axis branch preprocessing result are added and fused, and then subjected to 1×1 convolution to obtain the x-axis direction attention intermediate features of each layer. After performing layer normalization on the intermediate features of each layer, two parallel y-axis branches, a first y-axis branch and a second y-axis branch, are constructed along the y-axis direction of the intermediate features. The first y-axis branch and the second y-axis branch are y-axis branches. The first y-axis branch is sequentially subjected to 7×1 convolution and average pooling operations to obtain the first y-axis branch pooling result. The second y-axis branch is sequentially subjected to 15×1 dilated convolution and max pooling operations to obtain the second y-axis branch pooling result. The first y-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the first y-axis branch pooling result to obtain the first y-axis branch preprocessing result. The second y-axis branch pooling result is input into the mapping structure and processed by the activation function, and then multiplied element-wise with the second y-axis branch pooling result to obtain the second y-axis branch preprocessing result. The first y-axis branch preprocessing result and the second y-axis branch preprocessing result are added and fused, and then subjected to 1×1 convolution to obtain the y-axis direction attention intermediate features of each layer. Linear mapping is performed on the intermediate features of the x-axis attention in each layer to generate corresponding keys. and value Linear mapping is performed on the corresponding y-axis direction attention intermediate features to generate a corresponding x-axis direction attention intermediate feature. Furthermore, a linear mapping is performed on the intermediate features of the y-axis attention in each layer to generate corresponding keys. and value Linear mapping is performed on the corresponding x-axis direction attention intermediate features to generate the corresponding y-axis direction attention intermediate features. ;Will and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the x-axis direction attention output features of each layer, and then... and After performing the multiplication operation, the result is multiplied by... Perform multiplication to obtain the attention output features in the y-axis direction of each layer; The x-axis attention output features and y-axis attention output features of each layer are mapped by 1×1 convolution, and then added element-wise with the intermediate features to obtain the enhanced features of each layer.

6. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 5, characterized in that, The state space decoder module includes four cascaded state space decoder layers. The state space decoder module is used to perform state space modeling and step-by-step decoding and reconstruction on the enhanced features and the corresponding scale of the low-frequency information to obtain high-dimensional features, including: The fourth-layer state-space encoder layer is used to perform state-space modeling and step-by-step decoding reconstruction on the fourth-layer residual encoder layer outputting the fourth-layer decoding stage features. This includes: sequentially passing the fourth-layer encoded features through a convolutional mapping structure consisting of two 3×3 convolutions, batch normalization, and ReLU activation functions to obtain intermediate mapping features; and upsampling the intermediate mapping features to obtain the fourth-layer decoding stage features. The upsampling adopts bilinear interpolation and the upsampling factor is set to 2. The (n+1)th layer decoding stage features output by the (n+1)th layer state space encoder layer, the low-frequency information of the corresponding scale of the nth layer state space encoder layer, and the enhancement features of the corresponding scale are added together and used as the nth layer spatial encoding input of the nth layer state space encoder layer, where n=3,2,1, and the corresponding scale is the corresponding layer number. The nth-layer state-space encoder layer is used to perform state-space modeling and step-by-step decoding and reconstruction on the nth-layer spatial encoded input, outputting the features of the nth-layer decoding stage, including: The nth layer spatial encoding input is sequentially subjected to 1×1 convolution, layer normalization and GELU activation to obtain the state space modeling input features; The state space modeling input features are respectively input to the spatial state space model branch and the contrast state space model branch for parallel modeling to obtain the spatial state space model output features and the contrast state space model output features. After processing the output features of the spatial state spatial model sequentially with 3×3 convolution, layer normalization, GELU activation function, 1×1 convolution and Sigmoid activation function, the fused features are then multiplied element-wise with the output features of the contrast state spatial model at the corresponding spatial position and channel dimension, and then the residuals are added to the output features of the spatial state spatial model to obtain the fused features. The fused features are simultaneously input into three parallel convolutional branches for parallel convolution. Specifically, the first convolutional branch performs 5×1 and 1×5 convolutions sequentially, the second convolutional branch performs 3×3 convolutions sequentially, and the third convolutional branch performs 1×7 and 7×1 convolutions sequentially. Batch normalization is performed on the convolutional outputs of each convolutional branch, and the features of each convolutional branch after batch normalization are added element-wise at corresponding positions to obtain the branch fused features. After GELU activation processing is applied to the branch fused features, channel mapping and upsampling processing are performed sequentially using 1×1 convolutions to obtain the features of the nth layer decoding stage. The high-dimensional features are the features of the first-layer decoding stage output by the first-layer state space encoder layer.

7. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 6, characterized in that, The step of inputting the state-space modeling input features into the spatial state-space model branch and the contrast state-space model branch for parallel modeling, to obtain the spatial state-space model output features and the contrast state-space model output features, includes: The spatial state model branch flattens the state space modeling input features according to a preset spatial scanning order. The spatial scanning order adopts a spiral scanning order. The flattening process is as follows: starting from the upper left corner of the two-dimensional feature map corresponding to the state space modeling input features, it traverses horizontally to the right to the boundary, and then performs a spiral scan layer by layer in a clockwise direction until the entire two-dimensional feature map is traversed. After the traversal is completed, the features at the corresponding pixel positions are arranged sequentially according to the spatial scanning order to form a one-dimensional sequence. The one-dimensional sequence is then input into the spatial state model branch for state update and output calculation. According to the index relationship corresponding to the spatial scanning order, the sequence output by the spatial state model branch is rearranged and restored to a two-dimensional feature representation to obtain the output features of the spatial state model. The contrast state space model branch flattens the input features of the state space modeling in a channel-first manner, unfolding each channel feature into a one-dimensional sequence, and constructing a forward and reverse arrangement sequence based on the one-dimensional sequence. After inputting the forward and reverse arrangement sequences into the contrast state space model branch for state update and output calculation, the output corresponding to the reverse arrangement sequence is reversed and added to the corresponding position of the output corresponding to the forward arrangement sequence to restore the two-dimensional feature representation, thus obtaining the output features of the contrast state space model.

8. The magnetic resonance imaging mode conversion method based on wavelet transform and sensing state space modeling according to claim 1, characterized in that, The step of using the output head module to map the high-dimensional features to the target space and generate a target modal magnetic resonance imaging image includes: The high-dimensional features are subjected to feature refinement processing, which includes sequential 3×3 convolution operation, batch normalization processing, and LeakyReLU activation operation to obtain refined features; After upsampling the refined features, output mapping is performed to obtain the mapped features. The upsampling adopts bilinear interpolation and the upsampling factor is set to 2. The output mapping includes sequential 3×3 convolution operation, LeakyReLU activation operation and 1×1 convolution operation. The output mapping maps the number of channels of the refined features to the number of channels corresponding to the target modal magnetic resonance imaging image. The mapping features are processed by a hyperbolic tangent function to generate the target modal magnetic resonance imaging image.

9. The magnetic resonance imaging mode conversion method based on wavelet transform and perceptual state space modeling according to any one of claims 1-8, characterized in that, Also includes: The input feature mapping module, wavelet transform module, residual encoder module, multi-scale cross-efficiency attention module, state space decoder module, and output head module are trained using a multi-constraint joint loss function. The joint loss function includes: pixel-level L1 loss function, multi-scale structural similarity loss function, gradient consistency loss function, and contrast statistical loss function. The pixel-level L1 loss function is used to constrain the numerical accuracy of the image. The multi-scale structural similarity loss function is used to maintain the anatomical consistency of the magnetic resonance imaging image. The gradient consistency loss function is used to enhance the edge and detail preservation ability. The contrast statistical loss function is used to constrain the global and local grayscale contrast characteristics.

10. A magnetic resonance imaging mode conversion system based on wavelet transform and sensing state space modeling, characterized in that, include: The input processing module is used to construct an input feature tensor from the source modal magnetic resonance imaging images according to the number of modes. The input feature mapping module is used to perform channel mapping and initial feature extraction on the input feature tensor to obtain initial features. The input feature mapping module consists of an initial convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer connected in sequence. The wavelet transform module is used to perform multi-level discrete wavelet decomposition and convolution mapping on the input feature tensor to extract the high-frequency and low-frequency information of the source modal magnetic resonance imaging image. The residual encoder module is used to perform layer-by-layer downsampling and semantic feature encoding on the initial features to obtain multi-scale encoded features; A multi-scale cross-efficiency attention module is used to cross-fuse and enhance the attention of the multi-scale encoded features and the high-frequency information of the corresponding scale to obtain enhanced features; The state-space decoder module is used to perform state-space modeling and step-by-step decoding and reconstruction on the enhanced features and the low-frequency information at the corresponding scale to obtain high-dimensional features; The output head module is used to map the high-dimensional features to the target space to generate a target modal magnetic resonance imaging image.

Images