Flexible diffusion tensor estimation method and system based on q-space coordinate guidance

Abstract

The present invention relates to the technical field of fast diffusion tensor estimation, and discloses a flexible diffusion tensor estimation method and system based on q-space coordinate guidance, comprising: grouping each q-space coordinate in the input data with its corresponding DW image according to a group embedding fusion strategy; extracting multi-scale features of the DW image in each group of data through a shared encoder; embedding the multi-scale features with the corresponding q-space coordinates through an efficient embedding mechanism to obtain embedded features; fusing and reconstructing the embedded features through a decoder to output a high-quality diffusion tensor. The flexible and fast diffusion tensor estimation method based on q-space coordinate guidance provided by the present invention eliminates the reliance on a fixed diffusion sampling scheme, and solves the problem that previous methods fail to fully consider the spatial correlation between adjacent voxel signals. The present invention achieves better results in terms of flexibility and reconstruction quality.

Description

Technical Field

[0001] The present invention relates to the technical field of fast diffusion tensor estimation, and in particular to a flexible diffusion tensor estimation method and system based on q-space coordinate guidance. Background Art

[0002] Diffusion Magnetic Resonance Imaging (dMRI) is a rapidly developing imaging technique that detects the motion of water molecules in an applied diffusion gradient magnetic field. dMRI can generate diffusion-weighted (DW) images and further calculate the diffusion tensor (DT).

[0003] Traditional algorithms theoretically require only six DW images and one non-DW image to estimate the diffusion tensor. However, due to the typically low signal-to-noise ratio of clinical DW images, traditional algorithms struggle to provide accurate DT estimates on these images. Studies have shown that to achieve statistical rotational invariance, at least 30 measurements along evenly distributed directions are necessary. Therefore, to achieve a reasonably accurate DT estimate, a scan duration of 10 to 30 minutes per patient is typically required. However, prolonged scans can cause patient discomfort and may also introduce motion artifacts. Therefore, efficiently and accurately estimating the diffusion tensor while using fewer DW images has become an important research topic in diffusion MRI image processing.

[0004] Some researchers have used the powerful representation learning capabilities of deep learning to estimate high-quality diffusion tensors using six DW images with fixed diffusion gradient directions and b-values ​​and one non-DW image, thereby reducing the number of required DW images and scanning time. For example, DeepDTI first denoises low-quality DW images through a neural network to generate high-quality images, and then uses a traditional tensor model to fit a high-quality DT. SuperDTI uses a neural network to learn the mapping relationship between DW images and DT-derived parameters, rather than directly estimating DT. TransDTI bypasses the traditional tensor model that is sensitive to noise and error, and adopts a flexible Transformer model to achieve end-to-end cross-modal high-quality DT estimation.

[0005] Although these deep learning methods can reconstruct high-quality diffusion tensors from a small number of DW images, they still face some challenges in practical applications. Specifically, these methods usually rely on a specific q-space sampling scheme that must be consistent with the training data, otherwise the quality of the estimated diffusion tensor will deteriorate.

[0006] While traditional diffusion tensor estimation methods have good generalization performance, their estimation accuracy decreases when the number of diffusion gradient directions is small. To address this issue, DIFFnet introduces a matrix (Qmatrix) generated by projecting and quantizing q-space coordinates onto the corresponding diffusion signals as network input, thereby improving generalization. However, DIFFnet fails to fully account for the spatial correlation between adjacent voxel signals, and the quality of the diffusion tensor estimation still significantly decreases when the number of diffusion gradient directions is reduced. Given the inherent continuity and structural characteristics of DW images, this neglect of spatial information affects the robustness and accuracy of the derived parameters of the estimated DT. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by the present invention is that the existing diffusion tensor estimation method relies on a specific q-space sampling scheme, has limited generalization performance, and does not fully utilize the spatial information of adjacent voxels, resulting in a significant decrease in estimation quality under low sampling conditions.

[0009] To solve the above technical problems, the present invention provides the following technical solution: a flexible diffusion tensor estimation method based on q-space coordinate guidance, comprising: grouping each q-space coordinate in the input data with its corresponding DW image according to a group embedding and fusion strategy; extracting multi-scale features of the DW image in each group of data through a shared encoder; embedding the multi-scale features with the corresponding q-space coordinates through an efficient embedding mechanism to obtain embedded features; and fusing and reconstructing the embedded features through a decoder to output a high-quality diffusion tensor.

[0010] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the input data includes a q-space sampling scheme and its corresponding DW image.

[0011] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the q-space sampling scheme is composed of multiple q-space coordinates, each q-space coordinate contains a diffusion gradient direction and a b value, and can generate a corresponding DW image.

[0012] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the shared encoder adopts a continuous downsampling structure composed of a dual-branch residual dense module to perform multi-scale feature extraction on the DW image of each set of input data.

[0013] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the decoder adopts a continuous upsampling structure composed of a dual-branch residual dense module that is symmetrical to the shared encoder.

[0014] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the dual-branch residual dense module consists of three parts: multi-scale feature fusion, long jump connection and short jump connection. The dual-branch structure is introduced to process features of different scales, and long and short jump connections are used to enhance feature transfer and fusion.

[0015] As a preferred solution of the flexible diffusion tensor estimation method based on q-space coordinate guidance described in the present invention, the efficient embedding mechanism includes embedding and fusing the q-space coordinates with the multi-scale feature map of the DW image through a linear transformation through a q-space coordinate embedding (QCE) module, so that the feature map can dynamically adapt to different q-space sampling conditions, thereby enhancing the generalization ability of the model. The linear transformation is expressed as:

[0016] ,

[0017] in, and The q-space coordinates and extracted multi-scale features The adjustment parameter scale and displacement parameter are obtained. is the embedded feature after embedding the fused q space coordinates.

[0018] A flexible diffusion tensor estimation system based on q-space coordinate guidance using any of the methods described in the present invention, wherein: a grouping module divides input data into groups, each group containing a q-space coordinate and its corresponding DW image; an extraction module extracts multi-scale features of each group of DW images through a shared encoder; an embedding module embeds and fuses the multi-scale features with the corresponding q-space coordinates through an efficient embedding mechanism; and an output module performs same-level feature fusion on the embedded multi-scale features of each group to generate multi-scale fused features, and reconstructs and outputs a high-quality diffusion tensor.

[0019] A computer device comprises: a memory and a processor; the memory stores a computer program, comprising: the steps of implementing any one of the methods of the present invention when the processor executes the computer program.

[0020] A computer-readable storage medium stores a computer program thereon, comprising: steps of implementing any one of the methods of the present invention when the computer program is executed by a processor.

[0021] Beneficial effects of the present invention: The flexible and fast diffusion tensor estimation method based on q-space coordinate guidance provided by the present invention eliminates the dependence on a fixed diffusion sampling scheme, and uses a group embedding fusion strategy to enable the model to fully learn the association between different q-space coordinates and their corresponding DW images. Using an efficient embedding mechanism, through a linear adjustment feature map method, efficient embedding of each q-space coordinate and its corresponding DW image is achieved. Using a dual-branch residual dense module, the ability to extract DW image features and the ability to reconstruct the diffusion tensor are significantly enhanced, solving the problem that previous methods failed to fully consider the spatial correlation between adjacent voxel signals. The present invention achieves better results in terms of flexibility and reconstruction quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 A flowchart of a flexible diffusion tensor estimation method based on q-space coordinate guidance provided by one embodiment of the present invention;

[0024] Figure 2 An overall framework diagram of a flexible diffusion tensor estimation method based on q-space coordinate guidance provided by one embodiment of the present invention;

[0025] Figure 3 A structural diagram of a dual-branch residual dense module of a flexible diffusion tensor estimation method based on q-space coordinate guidance provided by one embodiment of the present invention;

[0026] Figure 4 A structural diagram of a q-space coordinate embedding module of a flexible diffusion tensor estimation method based on q-space coordinate guidance provided by one embodiment of the present invention;

[0027] Figure 5 A qualitative comparison diagram of different methods of the flexible diffusion tensor estimation method based on q-space coordinate guidance in terms of FA, MD, AD, and RD provided in the second embodiment of the present invention;

[0028] Figure 6 A comparison diagram of detailed anatomical information of different methods for the flexible diffusion tensor estimation method based on q-space coordinate guidance provided by the second embodiment of the present invention;

[0029] Figure 7The second embodiment of the present invention provides a flexible diffusion tensor estimation method based on q-space coordinate guidance, which generates whole-brain fiber bundles and main white matter bundle maps from 6 DW images using different methods. DETAILED DESCRIPTION

[0030] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0031] Example 1, with reference to Figure 1-Figure 4 , which is an embodiment of the present invention, provides a flexible diffusion tensor estimation method based on q-space coordinate guidance, including:

[0032] S1: According to the group embedding fusion strategy, each q-space coordinate in the input data is grouped with its corresponding DW image.

[0033] Furthermore, the input data includes a q-space sampling scheme and its corresponding DW image.

[0034] It should be noted that the q-space sampling scheme consists of multiple q-space coordinates, each of which contains a diffusion gradient direction and a b value, and can generate a corresponding DW image.

[0035] At the same time, the input data consists of a non-DW image and six DW images and its q-space sampling scheme , where the q space coordinates With DW image Corresponding . ,in represents the diffusion gradient direction, Represents the b value. The output of the network is six different elements of the diffusion tensor .

[0036] Furthermore, the traditional method uses the channel stitching method to take all DW data as input data. The problem solved by the present invention is to embed the q-space coordinates to achieve flexible diffusion tensor estimation. Since each q-space coordinate corresponds to a DW image, the channel stitching method will cause the model to be unable to learn this correspondence, so a group embedding fusion strategy is designed.

[0037] The input data is divided into groups, each group contains a q-space coordinate and its corresponding DW image.

[0038] It should be noted that the six DW images With the same non-DW image Splicing in the channel dimension to obtain the input of six groups of shared parameter encoders (SE) :

[0039] , ,

[0040] Among them, B represents the batch size of the input data, 2 represents the number of channels of the data (one non-DW image and one DW image), H, W, and D represent the height, width, and depth of the data respectively. is the set of real numbers.

[0041] Because each set of inputs needs to be extracted through a shared encoder to extract multi-scale features and embed the q-space coordinates, in the implementation process, the six sets of input data are spliced ​​on the batch channel B to obtain the input of the shared parameter encoder The same operation is performed on the q space coordinates to obtain the module input of QCE .

[0042] S2: Extract multi-scale features of the DW image in each set of data through a shared encoder.

[0043] Furthermore, the multi-scale features of each input data are extracted through the encoder with shared parameters .

[0044] In this way, the shared parameters can be fully utilized to improve the efficiency of feature extraction, while ensuring that the feature extraction process of each DW image is correlated with each other.

[0045] ,

[0046] ,

[0047] in, is the feature map output by the first module, represents the j-th dual-branch residual dense module operation in the shared encoder SE, is the downsampling operation.

[0048] , represents the feature map output by the j-th module, It is the feature map output by the previous module. In the present invention, the encoder with shared parameters consists of three consecutive down-sampling dual-branch residual dense modules.

[0049] It should be noted that the shared encoder adopts a continuous downsampling structure composed of a dual-branch residual dense module to perform multi-scale feature extraction on the DW image of each set of input data.

[0050] It should also be noted that the dual-branch residual dense module is used to learn the structural features and detail features of DW images at different scales, such as Figure 3 As shown. The dual-branch residual dense module consists of three parts: multi-scale feature fusion, long skip connection and short skip connection. Specifically, in the multi-scale feature fusion part, the dual-branch residual dense module first uses two convolutions with kernel sizes of 3 and 5 to extract the detail information and structural information of the DW image respectively, and then uses a convolution with a kernel size of 1 to perform channel reduction on these multi-scale information, and finally fuses the features through a convolution with a kernel size of 3. The dual-branch residual dense module not only uses long skip connections, but also short skip connections to share features of different scales so that the model can effectively capture local and global information. In the present invention, short skip connections are fused with multi-scale features to form a dual-branch residual (DR) module.

[0051] ,

[0052] in, Represents the features after multi-scale feature fusion, For input, represents the Relu activation layer, 、 and Represent convolution operations with kernel sizes of 1, 3, and 5, respectively. Indicates that the added feature maps are concatenated in the channel dimension.

[0053] DRD is composed of two DR modules and long jump connections.

[0054] ,

[0055] in, represents the multi-scale residual module, For input, is the output feature.

[0056] S3: Embed the multi-scale features with the corresponding q-space coordinates through an efficient embedding mechanism to obtain embedded features.

[0057] Each q-space coordinate is embedded into the multi-scale features extracted from the corresponding DW image through the q-space coordinate embedding module. Fusion .

[0058] ,

[0059] in, Representatives of the QCE operation is performed based on the characteristics of

[0060] It should be noted that in order to efficiently integrate the q-space sampling scheme with the multi-scale features extracted from the corresponding DW image for multimodal embedding, the present invention proposes a q-space coordinate embedding module, such as Figure 4 shown.

[0061] The QCE module is set on the jump connection between the encoder and the decoder. Multi-scale features extracted from the corresponding DW image by DRD As input, the channel dimension of DW feature is linearly transformed. Efficient embedding, obtaining .

[0062] Specifically, QCE first extracts the multi-scale features Extract overall features through Global Average Pooling (GAP) and Global Max Pooling (GMP) and salient features , realize DW image and q space coordinates Bidirectional learning between After passing through a shared multilayer perceptron (MLP) Splicing; then generate adjustment coefficients through MLP and That is, scale and displacement parameters; finally, by adjusting the coefficient and Perform linear transformation on multi-scale features. QCE feature fusion The calculation is:

[0063] ,

[0064] ,

[0065] in, represents the average split operation, represents MLP, represents global average pooling, Represents global maximum pooling.

[0066] S4: The embedded features are fused and reconstructed through the decoder to output a high-quality diffusion tensor.

[0067] After embedding the q-space coordinates, the multi-scale features of the same level are concatenated by channel, and then the number of channels of the concatenated features is matched to the decoder of the same level through convolution operations. This method associates the fused features of each DW image and ensures that the input decoder features match the encoder features, thereby facilitating the subsequent decoding and reconstruction process.

[0068] ,

[0069] in, Representatives of the Characteristics of the operation Perform fusion operation because ,in It is obtained after the QCE operation of the i-th DW image. It represents the splicing operation of the features after embedding the q-space coordinates. During the implementation process, the Reshape operation is used to perform channel splicing operation on the fused features.

[0070] This embodiment also provides a flexible diffusion tensor estimation system based on q-space coordinate guidance, including a grouping module that divides input data into groups, each group containing a q-space coordinate and its corresponding DW image; an extraction module that extracts multi-scale features from each group of DW images through a shared encoder; an embedding module that embeds and fuses the multi-scale features with the corresponding q-space coordinates through an efficient embedding mechanism; and an output module that fuses the embedded multi-scale features of each group at the same level to generate multi-scale fused features and reconstruct and output a high-quality diffusion tensor.

[0071] Example 2, reference Figure 5-Figure 7 , which is an embodiment of the present invention, provides a flexible diffusion tensor estimation method based on q-space coordinate guidance. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0072] The proposed method was compared with previous methods for reconstructing high-quality diffusion tensors from a small number of DW images using a fixed q-space sampling scheme, namely DeepDTI and TransDTI. In addition, to demonstrate the effectiveness of the proposed method in estimating the diffusion tensor under a flexible q-space sampling scheme, comparative experiments were conducted with weighted linear least squares (WLS) and DIFFnet. In order to compare the proposed algorithm with related research algorithms, the norm of the difference between the predicted diffusion tensor and the reference diffusion tensor was first used as an evaluation metric. Specifically, represents the predicted diffusion tensor, The reference diffusion tensor estimation error is defined as , where index Refers to six tensor elements.

[0073] The present invention calculates the eigendecomposition of the tensor in each voxel and uses their standard definitions to calculate the fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) from the eigenvalues. For each voxel, the error of FA, MD, AD, and RD is defined as the absolute error between the predicted result and the reference result. The present invention calculates the angle between the principal eigenvectors of the predicted tensor and the reference tensor as the angular error between the two tensors. The mean absolute difference (MAD) is used to quantitatively compare FA, MD, AD, RD, and the principal eigenvectors of the tensor. In addition, the present invention also provides a visual evaluation of the fiber bundle map.

[0074] Table 1 shows the errors of the proposed method in estimating the tensor, FA, MD, AD, RD, and the direction of the main eigenvectors for the HCP dataset using fixed and flexible q-space sampling schemes. For each of these variables, the average error is calculated for each test subject (only the part without CSF is taken).

[0075] As shown in Table 1, on all 20 test objects from the HCP dataset, the present method achieved lower estimation errors than all other methods. DIFFnet uses DW images to directly generate derivative parameters such as FA, MD, AD and RD, rather than reconstructing the diffusion tensor. Compared with other methods on a dataset with a fixed q-space sampling scheme, the method proposed in the present invention achieved the best in all measurement indicators. From the experimental results on a dataset with a flexible q-space sampling scheme, it can be seen that DeepDTI and TransDTI have significantly decreased in all indicators. Although DIFFnet is designed based on a flexible sampling scheme, it cannot estimate a high-quality diffusion tensor using only 6 DW images. The method of the present invention reaches the best on a dataset obtained using a random diffusion sampling scheme, and exceeds the method using data obtained using a fixed diffusion sampling scheme.

[0076] Table 1 Quantitative results

[0077] ,

[0078] Figure 5The visualization results of FA, MD, AD, and RD generated using six DW images using a traditional tensor fitting method (WLS), DeepDTI, TransDTI, DIFFnet, and the method of the present invention are shown. The WLS results, using all 90 DW images plus 18 unweighted images, were used as the ground truth. Because traditional methods are very sensitive to noise, WLS yields the worst results. DeepDTI's results are still affected by noise because it uses a noise-sensitive traditional method to fit the diffusion tensor. TransDTI exhibits a grid effect, which reduces the reconstruction quality and visualization effect. Notably, DIFFnet cannot estimate a high-quality diffusion tensor using six DW images, while the method of the present invention is not only applicable to data with flexible diffusion schemes, but can also reconstruct a high-quality diffusion tensor using only six DW images. The results of the method of the present invention are less noisy than those of other methods, and the details are more similar to the reference image. The present invention displays the MAD of the quantitative parameters generated by different methods below the residual plot, and the results show that the method of the present invention outperforms other methods in both visualization and quality.

[0079] Figure 6 The fractional anisotropy maps from different methods are shown. In terms of visual quality, the estimates from our method are more accurate than those from the other methods, visually closer to the labels and with less noise. To observe the effect in more detail, we zoomed in on a small area for a closer comparison, as shown in the yellow box. In this zoomed-in view, the principal fiber directions estimated by our method (shown as sticks) are as coherent as the labels. In contrast, the other three estimates all have very incoherent fibers.

[0080] In addition to the voxel-wise performance comparison, the present invention also performs tractography analysis. Figure 7 The diffusion tensors obtained by different methods were used to generate whole-brain fiber tracts using the FACT fiber tracking algorithm, and 20 representative major white matter tracts were identified using Automated Fiber Quantification (AFQ) software. Compared with the WLS method, DeepDTI, TransDTI, and the method of the present invention were visually closer to the labels. Compared with other methods, the method of the present invention performed better in terms of visual quality. In the results of major white matter tracts, the visual quality of each white matter tract of the method of the present invention was more similar to the label, while other methods lacked fiber tracts due to estimation errors. Notably, the estimation of the arcuate white matter tract by the method of the present invention was basically consistent with the label, while DeepDTI and TransDTI had estimation errors.

[0081] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0082] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0083] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, and then editing, interpreting, or processing in another suitable manner as necessary, and then storing it in a computer memory.

[0084] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A flexible diffusion tensor estimation method based on q-space coordinate guidance for diffusion magnetic resonance imaging image processing, characterized in that: include: According to the group embedding fusion strategy, the input data is divided into groups, and each q-space coordinate in the input data is grouped with the DW image corresponding to the coordinate; Each group includes a q-space coordinate and its corresponding DW image, wherein the input data includes a q-space sampling scheme and the DW image corresponding to the coordinate; the q-space sampling scheme is composed of multiple q-space coordinates, each q-space coordinate includes a diffusion gradient direction and a b value, and can generate a corresponding DW image; A shared encoder is used to extract multi-scale features of the DW image in each set of data. The shared encoder uses a continuous downsampling structure consisting of a dual-branch residual dense module to extract multi-scale features from the DW image of each set of input data. The dual-branch residual dense module consists of three parts: multi-scale feature fusion, long skip connections, and short skip connections. It introduces a dual-branch structure to process features at different scales, and uses long and short skip connections to enhance feature transfer and fusion. The multi-scale features are embedded with the corresponding q-space coordinates through an embedding mechanism to obtain embedded features. The q-space coordinate embedding module is set on the jump connection between the shared encoder and decoder. The q-space coordinate embedding is specifically as follows: first, the extracted multi-scale features are subjected to global average pooling and global maximum pooling to extract overall features and significant features, realizing bidirectional learning between the DW image and the q-space coordinates; secondly, the extracted overall features are passed through a shared multi-layer perceptron (MLP) and concatenated with the q-space coordinates; then, the MLP generates adjustment coefficients, i.e., scale and displacement parameters; finally, the multi-scale features are linearly transformed by the adjustment coefficients. The q-space coordinates are embedded by linearly transforming the channel dimension of the DW features to obtain embedded features. The embedded features are fused and reconstructed through the decoder to output a high-quality diffusion tensor; wherein the multi-scale features of the same level after embedding the q-space coordinates are spliced ​​through channels, and then the number of channels of the spliced ​​features is matched with the decoder of the same level through a convolution operation; The decoder adopts a continuous upsampling structure consisting of a dual-branch residual dense module that is symmetrical to the shared encoder.

2. The flexible diffusion tensor estimation method based on q-space coordinate guidance according to claim 1, characterized in that: The embedding mechanism includes, Through the q-space coordinate embedding module, the q-space coordinates are embedded and fused with the multi-scale feature map of the DW image through linear transformation, so that the feature map can dynamically adapt to different q-space sampling conditions. The linear transformation is expressed as: in, is the adjustment parameter scale, is the displacement parameter, is the multi-scale feature output by the j-th dual-branch residual dense module, is the embedded feature after embedding the fused q-space coordinates, and the embedded feature is obtained after the i-th DW image is embedded in the q-space coordinates.

3. A flexible diffusion tensor estimation system based on q-space coordinate guidance using the method according to any one of claims 1-2, for use in processing diffusion magnetic resonance imaging images, characterized in that: include, The grouping module divides the input data into groups, each group contains a q-space coordinate and a DW image corresponding to the coordinate; The extraction module extracts multi-scale features of each set of DW images through a shared encoder; Embedding module, which embeds and fuses multi-scale features with corresponding q-space coordinates through an embedding mechanism; The output module fuses the embedded multi-scale features of each group at the same level to generate multi-scale fusion features and reconstruct and output high-quality diffusion tensors.

4. A computer device comprising: memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the flexible diffusion tensor estimation method based on q-space coordinate guidance are implemented as described in any one of claims 1-2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the flexible diffusion tensor estimation method based on q-space coordinate guidance are implemented.

Images