Diffusion tensor imaging parameter quantification method, system, device and medium

Abstract

The present invention provides a method, system, device and medium for quantifying diffusion tensor imaging parameters, wherein the method comprises: obtaining dMRI data {S0, S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ; According to the diffusion coding direction g n The spherical harmonic function #imgabs0# is calculated to transform the dMRI data, resulting in a diffusion attenuation image. A spherical harmonic coefficient map is then calculated based on the diffusion attenuation image and the complex conjugate #imgabs1# of the spherical harmonic function. The spherical harmonic coefficient map is then input into the constructed deep learning network to obtain a DTI diffusion tensor field. Based on the DTI diffusion tensor field, a DTI quantization parameter map is calculated. This method can improve the accuracy of DTI diffusion tensor estimation, thereby improving the accuracy of DTI quantization parameters, providing a reliable quantitative basis for precision medicine.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic resonance technology, and in particular relates to a diffusion tensor imaging parameter quantification method, system, equipment and medium. Background Art

[0002] Diffusion Magnetic Resonance Imaging (dMRI) can noninvasively map the diffusion of water molecules in living tissue and is an important clinical examination tool. Diffusion tensor imaging (DTI), a common diffusion MRI modality, requires the acquisition of at least one non-diffusion-weighted image and six diffusion-weighted (DW) images. It is currently widely used in clinical quantitative diagnosis and visualization of white matter fiber tracts. The b-value, representing the diffusion sensitivity coefficient, is a key parameter of DW images. Its magnitude affects the attenuation of DW image signals; higher b-values ​​result in weaker DW image signals. Consequently, DW images are more noisy than conventional structural MRI images. The presence of noise blurs image details, degrades image quality, and leads to inaccurate estimation of the DTI diffusion tensor field. Consequently, this leads to inaccurate calculation of DTI quantitative parameters such as mean diffusivity (MD) and fractional anisotropy (FA), compromising clinical diagnosis and scientific research.

[0003] Therefore, reducing noise in DW images is essential. In clinical practice, the method of repeatedly acquiring multiple DW images and averaging them is often used to improve the signal-to-noise ratio (SNR) of DW images. However, this method prolongs acquisition time, thereby increasing acquisition costs and the probability of subject motion during acquisition. To this end, researchers have proposed a large number of post-processing techniques to reduce the impact of noise on DTI parameter quantification. These post-processing techniques include traditional methods and deep learning approaches. Among them, MPPCA, NLM, and BM3D are traditional methods for denoising DW images. The main drawbacks of traditional methods are long computational time and high algorithmic complexity. Current deep learning methods use deep learning networks to predict noise-free DW images, DTI diffusion tensor fields, or DTI quantification parameters based on noisy DW images. Existing deep learning methods offer better results and faster speed than traditional methods. However, existing deep learning methods are limited by the specified acquisition scheme (diffusion encoding direction, number of directions, and b-value), resulting in poor generalization and limited clinical application. Summary of the Invention

[0004] The present invention aims to provide a diffusion tensor imaging parameter quantification method, system, device and medium, which can improve the accuracy of DTI diffusion tensor field estimation, thereby improving the accuracy of DTI quantification parameters and providing a reliable quantitative basis for precision medicine.

[0005] The present invention is achieved through the following technical solutions:

[0006] A diffusion tensor imaging parameter quantification method comprises the following steps:

[0007] Acquire dMRI data {S0,S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ;

[0008] According to the diffusion coding direction g n , find the spherical harmonic function Y l m (g n );

[0009] Perform data transformation on the dMRI data to obtain the diffusion attenuation image, and then calculate the diffusion attenuation image and the complex conjugate Y of the spherical harmonic function. l m* (g n ) Calculate the spherical harmonic coefficient map;

[0010] The spherical harmonic coefficient map is input into the constructed deep learning network to obtain the DTI diffusion tensor field;

[0011] According to the DTI diffusion tensor field, the DTI quantitative parameter map is calculated.

[0012] Furthermore, the step of performing data transformation on the dMRI data includes:

[0013] Formula (1) is used to transform the dMRI image data. Formula (1) is:

[0014]

[0015] Where, {S(g n ), n=1,2,…,N} are diffusion attenuation images independent of b value, S0 and S b (g n ) are the non-diffusion-weighted image and diffusion-weighted image in dMRI data, b is the diffusion-weighted image S b (g n ) corresponding to the b value, b train is the b value of the training data of the deep learning network.

[0016] Furthermore, according to the diffusion attenuation image and the complex conjugate Y of the spherical harmonics l m* (g n ) The steps for calculating the spherical harmonic coefficients map include:

[0017] Complex conjugate Y based on diffusion attenuation image and spherical harmonics l m* (g n ), using the generalized Fourier transform, the spherical harmonic coefficients corresponding to each voxel are calculated voxel by voxel. The calculation formula of the spherical harmonic coefficients is as follows:

[0018]

[0019] Where, are the spherical harmonic coefficients, Y l m* (r) Spherical harmonic function Y l m The complex conjugate of (r), r is the coordinate on the unit sphere, r∈R 3 , s(r) is the spherical signal, indicating the direction of different diffusion gradients g n The corresponding diffusion signal attenuation value, s(r)∈R.

[0020] Furthermore, the deep learning network includes:

[0021] An encoder is used to extract features of the spherical harmonic coefficient graph to obtain a feature image;

[0022] The decoder is used to decode the feature image extracted by the encoder to obtain a decoded feature map;

[0023] The skip connection layer connects the feature image output by the encoder with the decoded feature image output by the decoder.

[0024] Furthermore, the DTI quantitative parameter map includes at least a FA map and a MD map;

[0025] The steps of calculating the DTI quantization parameter map according to the DTI diffusion tensor field include:

[0026] Perform eigenvalue decomposition on the diffusion tensor D of the DTI diffusion tensor field to obtain eigenvalues ​​λ1, λ2, and λ3, and

[0027] The FA map is calculated according to formula (3), and the MD map is calculated according to formula (4), where formula (3) is as follows:

[0028]

[0029] Formula (4) is as follows:

[0030]

[0031] Furthermore, the deep learning network training process includes the following steps:

[0032] (1) Training data preparation: obtain multiple dMRI data and process each dMRI data. The data processing process is as follows:

[0033] (1-1) Select a diffusion-weighted image with a preset b value and a corresponding non-diffusion-weighted image and perform preprocessing. The preprocessing process includes image denoising, Gibbs artifact removal, and bias field correction;

[0034] (1-2) For the preprocessed diffusion-weighted image and the corresponding non-diffusion-weighted image, the weighted linear least squares method is used to perform pixel-by-pixel fitting to obtain the DTI diffusion tensor field, and the obtained DTI diffusion tensor field is used as the real diffusion tensor field of the dMRI data. And the true value of the network output when the dMRI data is used as the training dataset;

[0035] (1-3) Apply the DTI model to the DTI diffusion tensor field and reconstruct the dMRI data using the preset diffusion gradient table to obtain a diffusion-weighted image. The formula of the DTI model is:

[0036]

[0037] Where s b is the diffusion weighted signal, s0 is the non-diffusion weighted signal, s0∈R, g is the diffusion gradient direction, g∈R 3×1 , D is the DTI diffusion tensor, D∈R 3×3 , is a positive definite symmetric matrix;

[0038] (1-4) Add Rice noise to the reconstructed dMRI data:

[0039]

[0040] Where I is the real noise-free dMRI data after reconstruction, N1 and N2 are independent and identically distributed Gaussian noises;

[0041] (1-5) For dMRI data I with Rice noise added noise , using formula (1) and formula (2) to obtain the corresponding spherical harmonic coefficient map, and using the obtained spherical harmonic coefficient map as the input when the dMRI data is used as the training data set;

[0042] (2) Deep learning network training;

[0043] (2-1) Randomly select n cases from multiple dMRI data as training sets, and the remaining cases as validation sets;

[0044] (2-2) Use the training set to train the deep learning network until the network converges and the validation set loss reaches a stable level.

[0045] Furthermore, the loss function of the deep learning network is:

[0046]

[0047] Where, They are the true diffusion tensor field corresponding to the training data and the diffusion tensor field predicted by the deep learning network, respectively.

[0048] The present invention also provides a diffusion tensor imaging parameter quantification system, comprising:

[0049] Acquisition module, used to acquire dMRI data {S0,S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ;

[0050] Obtaining module, used to encode the direction g according to the diffusion n , find the spherical harmonic function Y l m (g n );

[0051] The processing module is used to perform data transformation on the dMRI data to obtain the diffusion attenuation image and calculate the complex conjugate Y of the diffusion attenuation image and the spherical harmonic function. l m* (g n ) Calculate the spherical harmonic coefficient map;

[0052] The input module is used to input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field;

[0053] The calculation module is used to calculate the DTI quantitative parameter map according to the DTI diffusion tensor field.

[0054] The present invention also discloses an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0055] The present invention also discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0056] Compared with the existing technology, the beneficial effects of the present invention are: using deep learning networks and spherical harmonic functions to quantify diffusion tensor imaging parameters is faster than traditional methods and can obtain more accurate DTI quantification parameters; and compared with other diffusion tensor parameter quantization methods based on deep learning networks, it is not restricted to a specific acquisition scheme and can accurately estimate the diffusion tensor field of dMRI data from different acquisition schemes, with strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of the steps of the diffusion tensor imaging parameter quantification method of the present invention;

[0058] Figure 2 This is the overall framework diagram of the diffusion tensor imaging parameter quantification method of the present invention;

[0059] Figure 3 Schematic diagram of the structure of the deep learning network in the diffusion tensor imaging parameter quantification method of the present invention;

[0060] Figure 4 The experimental method comparison results for the simulation data of Example 1 are shown, showing the six element maps, FA map and MD map of the DTI diffusion tensor field and the corresponding error map;

[0061] Figure 5 The experimental method comparison results of high-resolution real brain data in Example 2 are shown, showing the six element maps of the DTI diffusion tensor field, FA map and MD map, and local magnification map;

[0062] Figure 6 Schematic diagram of the modules of the diffusion tensor imaging parameter quantification system of the present invention;

[0063] Figure 7 This is a schematic structural diagram of an electronic device according to an embodiment of the present invention;

[0064] Figure 8 This is a schematic block diagram of the structure of an embodiment of a computer-readable storage medium of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0066] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0067] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are used only to distinguish the description and should not be understood as indicating or implying relative importance.

[0068] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0069] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the inventive product is usually placed when in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they should not be understood as limiting the present invention.

[0070] See also Figure 1 and Figure 2 , Figure 1 is a flowchart of the steps of the diffusion tensor imaging parameter quantification method of the present invention, Figure 2 The overall framework diagram of the diffusion tensor imaging parameter quantification method of the present invention is as follows:

[0071] S1, obtain dMRI data {S0, S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n;

[0072] S2, according to the diffusion coding direction g n , find the spherical harmonic function Y l m (g n );

[0073] S3, perform data transformation on the dMRI data to obtain the diffusion attenuation image, and calculate the diffusion attenuation image and the complex conjugate Y of the spherical harmonic function. l m* (g n ) Calculate the spherical harmonic coefficient map;

[0074] S4, input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field;

[0075] S5. Calculate a DTI quantization parameter map based on the DTI diffusion tensor field.

[0076] In the above step S1, the dMRI data {S0,S b (g n ), n=1,2,…,N}, S0 represents the non-diffusion weighted image, b value is 0, S b (g n ) indicates that the diffusion coding direction is g n The diffusion weighted image collected at this time has a b value greater than 0 and a diffusion coding direction g n ∈R 3 ; {S0,S b (g n ), n=1,2,…,N} means that the acquired dMRI data contains one non-diffusion-weighted image and N diffusion-weighted images, and the N diffusion-weighted images are S b (g1), S b (g2), ..., S b (g N ), and the diffusion coding directions corresponding to the N diffusion weighted images are g1, g2, ..., g N Furthermore, N is greater than or equal to 6, that is, the acquired dMRI data contains at least 6 diffusion-weighted images.

[0077] In the above step S2, according to the diffusion coding direction g of the dMRI data n , the spherical harmonic function Y of the corresponding diffusion coding direction can be calculated using formula (3) l m (g n ), formula (3) is as follows:

[0078]

[0079] Where, P l m is the companion Legendre polynomial, Y l m are called spherical harmonics of l and m, where l is a natural number and m = -l, -(l-1),…,0,…,(l-1), For r∈R 3 The representation in spherical coordinate system. Preferably, the order l is 0 and 2, that is, the spherical harmonic function is obtained

[0080] In step S3, the step of performing data transformation on the dMRI data includes:

[0081] S31. Use formula (1) to transform the dMRI data. Formula (1) is:

[0082]

[0083] In the formula, {S(g n ), n=1,2,…,N} are diffusion attenuation images independent of b value, S0 and S b (g n ) are the non-diffusion-weighted image (b=0) and diffusion-weighted image (b>0) in the dMRI test data, b is the diffusion-weighted image S b (g n ) corresponding to the b value, b train is the b value of the training data of the deep learning network.

[0084] In step S31, the diffusion weighted images S b (g n ) is divided by the non-diffusion weighted image S0, and then the diffusion weighted image S b (g n ) is subjected to power operation on the corresponding b value to obtain the corrected diffusion attenuation image that is independent of the b value. b (g n ) The corresponding b value is the diffusion sensitivity factor, which is controlled by the parameters of the applied gradient field strength and the unit is s / mm 2 The larger the b value is, the smaller the signal of the diffusion weighted image is. In this embodiment, the training data b=1000s / mm 2 , that is, b train =1000s / mm 2 Because different MRI acquisition equipment or acquisition centers may have different b-values ​​(b-value is a parameter that can be set on the machine, similar to parameters such as TR or TE), the commonly used b-value for DTI acquisition is 800s / mm. 2 , 1000s / mm 2Or 1500s / mm 2 Different b values ​​result in different diffusion weighted signals of the same tissue, and the b value used in the training data is 1000s / mm 2 ,In order to ensure that the trained network can be applied to dMRI data with other b values, ,we use formula (1) to perform data transformation on dMRI data.

[0085] In step S3, the complex conjugate Y of the diffusion attenuation image and the spherical harmonic function is calculated. l m* (g n ) The steps for calculating the spherical harmonic coefficients map include:

[0086] S32, complex conjugate Y based on diffusion attenuation image and spherical harmonics l m* (g n ), using the generalized Fourier transform, the spherical harmonic coefficients corresponding to each voxel are calculated voxel by voxel. The calculation formula of the spherical harmonic coefficients is as follows:

[0087]

[0088] Where, are the spherical harmonic coefficients, Y l m* (r) is the spherical harmonic function Y l m The complex conjugate of (r), r is the coordinate on the unit sphere, r∈ 3 , s(r) is a spherical signal, indicating different diffusion gradient directions (r = n ) corresponds to the diffusion signal attenuation value, s(r)∈R.

[0089] In the above step S32, the spherical harmonic coefficient map is a map composed of spherical harmonic coefficients. Each voxel corresponds to a spherical harmonic coefficient (scalar), and the spherical harmonic coefficient corresponds to a map, which is called a spherical harmonic coefficient map. Specifically, in this embodiment, the order l = 2, so the spherical harmonic coefficient calculated for each voxel based on the non-diffusion weighted image and the diffusion weighted image is a 1*6 vector, so the spherical harmonic coefficient map is six maps, such as Figure 2 In addition, s(r) is the diffusion attenuation signal, and r corresponds to the corresponding diffusion coding direction g n , is the diffusion attenuation signal of all diffusion coding directions and Y l m* The integral form of (r). Formula (2) is described by the signal, while formula (1) describes the entire image, so it is called a diffusion attenuation signal.

[0090] In step S4 above, a deep learning network is pre-constructed and trained using pre-set training data. The deep learning network inputs the spherical harmonic coefficient map, and the deep learning network outputs the DTI diffusion tensor field. Therefore, the calculated spherical harmonic coefficient map is input into the pre-constructed deep learning network, and the DTI diffusion tensor field is obtained through the deep learning network. The deep learning network can be a U-net network, which is a U-shaped network.

[0091] Furthermore, the deep learning network includes:

[0092] An encoder is used to extract features of the spherical harmonic coefficient map to obtain a feature image;

[0093] A decoder, configured to decode the feature image extracted by the encoder to obtain a decoded feature image;

[0094] Skip connection layer, s(r)∈R.

[0095] The deep learning network consists of an encoder and a decoder, that is, an encoding path and a decoding path. The encoder consists of a 3D convolutional layer and a 3D maximum pooling layer. The decoder consists of a 3D convolutional layer and a 3D upsampling layer. The convolution kernel size of the 3D convolutional layer is 3×3×3, and the activation function is ReLU. Skip connection layers are used to connect features of the same dimension in the encoder and decoder, allowing for channel-wise fusion of features of the same dimension in the encoder and decoder. It should be noted that the type and layout of the deep learning network described above are for illustration and not limitation; any other suitable configuration may be used.

[0096] Furthermore, the deep learning network training process includes the following steps:

[0097] (1) Training data preparation: multiple dMRI data were obtained. The dMRI data used for training came from the diffusion magnetic resonance imaging data of healthy adults in the Human Connectome Project (HCP). 20 dMRI data were selected from the data for training and validation of the deep learning network. Each dMRI data was processed. The data processing process is as follows:

[0098] (1-1) Select a diffusion-weighted image with a preset b value and the corresponding non-diffusion-weighted image and perform preprocessing; the preprocessing process includes image denoising, Gibbs artifact removal, and bias field correction; the preset value can be 1000s / mm 2 ;

[0099] (1-2) For the preprocessed diffusion-weighted image and the corresponding non-diffusion-weighted image, the linear least squares method is used to perform pixel-by-pixel fitting to obtain the DTI diffusion tensor field, and the obtained DTI diffusion tensor field is used as the real diffusion tensor field of the dMRI data. And the true value of the network output when the dMRI data is used as the training dataset;

[0100] (1-3) Apply the DTI model to the DTI diffusion tensor field and reconstruct the dMRI data using a preset diffusion gradient table to obtain a diffusion-weighted image. Ensure that the obtained diffusion-weighted image satisfies the DTI model. The preset diffusion gradient table can be given 15 diffusion encoding directions and b = 1000 s / mm 2 , the formula of the DTI model is:

[0101]

[0102] Where s b is the diffusion weighted signal, s0 is the non-diffusion weighted signal, s0∈R, g is the diffusion gradient direction, g∈R 3×1 , D is the DTI diffusion tensor, D∈R 3×3 , is a positive definite symmetric matrix; the diffusion weighted signal forms a diffusion weighted image and the non-diffusion weighted signal forms a non-diffusion weighted image;

[0103] (1-4) Add Rice noise to the reconstructed dMRI data:

[0104]

[0105] Where I is the real noise-free dMRI data after reconstruction, N1 and N2 are independent and identically distributed Gaussian noises;

[0106] (1-5) For dMRI data I with Rice noise added noise , use formula (1) and formula (2) to obtain the corresponding spherical harmonic coefficient map, and use the obtained spherical harmonic coefficient map as the input of the network when dMRI data is used as the training data set;

[0107] (2) Deep learning network training;

[0108] (2-1) Randomly select n cases from multiple dMRI data as training sets, and the remaining cases as validation sets; if 20 cases of dMRI data were obtained in step (1-1), then randomly select 16 cases of dMRI data from the 20 cases of dMRI data as training sets, and the remaining 4 cases of dMRI data as validation sets;

[0109] (2-2) Use the training set to train the deep learning network until the network converges and the validation set loss reaches a stable level.

[0110] In the above steps (1) and (2), when performing deep network training, the spherical harmonic coefficient map obtained in step (1) based on the dMRI data in the training set and the real diffusion tensor field of the dMRI data are used. A first data pair is formed, and the deep learning network is trained with the first data pair until the network converges. Then the spherical harmonic coefficient map obtained in step (1) of the dMRI data of the validation set and the true diffusion tensor field of the dMRI data are obtained. A second data pair is formed and the deep learning network is validated with the second data pair until the validation set loss reaches a stable level. Furthermore, the loss function of the deep learning network is the mean square error (MSE) between the DTI diffusion tensor field predicted by the deep learning network and the actual DTI diffusion tensor field after the DTI model is applied:

[0111]

[0112] Where, They are the true diffusion tensor field corresponding to the training data and the diffusion tensor field predicted by the deep learning network, respectively.

[0113] The DTI diffusion tensor D∈R corresponding to each voxel in the image 3×3 The constructed field map, where M×N×L is the size of dMRI data.

[0114] Furthermore, the DTI quantization parameter map includes at least an FA map and an MD map. In step S5, the step of calculating the DTI quantization parameter map according to the DTI diffusion tensor field includes:

[0115] S51. Perform eigenvalue decomposition on the diffusion tensor D of the DTI diffusion tensor field to obtain eigenvalues ​​λ1, λ2, and λ3, and calculate the FA map according to formula (3), and calculate the MD map according to formula (4), wherein formula (3) is as follows:

[0116]

[0117] Formula (4) is as follows:

[0118]

[0119] In the above step S51, the DTI diffusion tensor field D field ∈R M×N×L×3×3 The DTI diffusion tensor D∈R corresponding to each voxel in the image 3×3 The field map is constructed, where M×N×L is the size of the dMRI data. The diffusion tensor D is a symmetric positive definite matrix composed of six unknown parameters; the six unknown parameters of any diffusion tensor D in the DTI diffusion tensor field {Dxx ,D xy ,D xz ,D yy ,D yz ,D zz The positive definite symmetric matrix composed of} is expressed as:

[0120]

[0121] Therefore, the diffusion tensor D positive definite symmetric matrix is ​​eigendecomposed to obtain eigenvalues ​​λ1, λ2, and λ3, and the eigenvalues ​​λ1, λ2, and λ3 are combined to calculate the FA map and MD map, thereby obtaining the DTI quantization parameter map.

[0122] To demonstrate the generalization capability of the diffusion tensor imaging parameter quantization method of the present invention (hereinafter referred to as the present method), two embodiments are used for illustration below.

[0123] Example 1

[0124] This method is compared with existing deep learning methods (dMRI data is used as input to the deep learning network) and the traditional denoising method MPPCA (MPPCA is used for denoising, and then WLLS is used to estimate the pixel-by-pixel tensor fitting of the denoised dMRI data). This example simulates the diffusion MRI data of healthy adults from HCP. The simulated data includes a b = 0 s / mm 2 , 15 b = 1000s / mm 2 , with a noise level of 0.02, but with gradient directions that differ from the 15 gradient directions in the training data, serving as test data for this method. This method uses the same deep learning network structure as existing deep learning methods. The evaluation metric is the root mean square error (RMSE).

[0125] See also Figure 4 , Figure 4 The experimental method comparison results of the simulation data of Example 1 are shown, showing the six element maps, FA map and MD map of the DTI diffusion tensor field and the corresponding errors. Figure 4 As shown in the figure, the six element maps of the DTI diffusion tensor field and the FA and MD maps obtained by this method contain the least noise and the details closest to the true values, and have the lowest RMSE value, which proves that this method has the best effect in denoising and parameter estimation, and has strong generalization ability.

[0126] Example 2,

[0127] This method was compared with the MPPCA algorithm (MPPCA was used for denoising, and then WLLS was used to perform pixel-by-pixel tensor fitting estimation on the denoised dMRI data). Existing deep learning methods are trained on 15 diffusion coding directions and cannot be directly applied to the real data (containing 12 diffusion coding directions), so they were not compared with existing deep learning methods. The data used in this embodiment is real data, which is high-resolution dMRI data of the brain of healthy volunteers (resolution is 0.9mm×0.9mm×0.9mm), containing a b=0s / mm 2 , 12 b = 800s / mm 2 .

[0128] See also Figure 5 , Figure 5 The experimental method comparison results of high-resolution real brain data in Example 2 are shown, including six element maps of the DTI diffusion tensor field, FA map, MD map and local magnification map. Figure 5 As shown in the figure, the six element maps of the DTI diffusion tensor field and the FA and MD maps obtained using this method have the best denoising effect and image detail preservation, which proves that this method is effective in denoising and DTI parameter quantification and has strong generalization ability.

[0129] See also Figure 6 , Figure 6 The present invention also provides a diffusion tensor imaging parameter quantification system, comprising:

[0130] Acquisition module 1 is used to perform eigendecomposition on the diffusion tensor D positive definite symmetric matrix to obtain eigenvalues ​​λ1, λ2, and λ3, and to combine the eigenvalues ​​λ1, λ2, and λ3 to calculate the FA map and MD map, thereby obtaining the DTI quantitative parameter map;

[0131] Obtaining module 2, for encoding direction g according to diffusion n , find the spherical harmonic function Y l m (g n );

[0132] Processing module 3 is used to perform data transformation on dMRI data to obtain diffusion attenuation image and calculate the diffusion attenuation image and the complex conjugate Y of spherical harmonic function. l m* (g n )Calculate the spherical harmonic coefficients map;

[0133] Input module 4 is used to input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field;

[0134] The calculation module 5 is used to calculate the DTI quantization parameter map according to the DTI diffusion tensor field.

[0135] Acquisition module 1 obtains dMRI data {S0,S b (g n ), n=1,2,…,N}, S0 represents the non-diffusion weighted image, b value is 0, S b (g n ) indicates that the diffusion coding direction is g n The diffusion weighted image collected at this time has a b value greater than 0 and a diffusion coding direction g n ∈R 3 ; {S0,S b (g n ), n=1,2,…,N} means that the acquired dMRI data contains one non-diffusion-weighted image and N diffusion-weighted images, and the N diffusion-weighted images are S b (g1), S b (g2), ..., S b (g N ), and the diffusion coding directions corresponding to the N diffusion weighted images are g1, g2, ..., g N Furthermore, N is greater than or equal to 6, that is, the acquired dMRI data contains at least 6 diffusion-weighted images.

[0136] Module 2 obtains the diffusion coding direction g according to the dMRI data n , the spherical harmonic function value Y of the corresponding diffusion coding direction can be calculated using formula (3) l m (g n ), formula (3) is as follows:

[0137]

[0138] Where, P l m is the companion Legendre polynomial, Y l m It is called the spherical harmonic function of l and m, where l is a natural number and m=-,-(-1),…,0,…,(-1), For r∈ 3 The representation in spherical coordinate system. Preferably, the order l is 0 and 2, that is, the spherical harmonic function is obtained

[0139] Processing module 3 includes:

[0140] The processing submodule is used to perform data transformation on the dMRI data using formula (1), which is:

[0141]

[0142] In the formula, {(g n ), = 1, 2, ...,} is the diffusion attenuation image independent of the b value, S0 and S b (g n ) are the non-diffusion-weighted image (b=0) and diffusion-weighted image (b>0) in the dMRI test data, b is the diffusion-weighted image S b (g n ) corresponding to the b value, b train is the b value of the training data of the deep learning network.

[0143] The processing submodules use the diffusion weighted image S b (g n ) is divided by the non-diffusion weighted image S0, and then the diffusion weighted image S b (g n ) is subjected to power operation on the corresponding b value to obtain the corrected diffusion attenuation image that is independent of the b value. b (g n ) The corresponding b value is the diffusion sensitivity factor, which is controlled by the parameters of the applied gradient field strength and the unit is s / mm 2 The larger the b value is, the smaller the signal of the diffusion weighted image is. In this embodiment, the training data b=1000 / mm 2 , that is, b train =1000 / mm 2 Because different MRI acquisition equipment or acquisition centers may have different b-values ​​(b-value is a parameter that can be set on the machine, similar to parameters such as TR or TE), the commonly used b-value for DTI acquisition is 800s / mm. 2 , 1000s / mm 2 Or 1500s / mm 2 Different b values ​​result in different diffusion weighted signals of the same tissue, and the b value used in the training data is 1000s / mm 2 ,In order to ensure that the trained network can be applied to dMRI data with other b values, ,we use formula (1) to perform data transformation on dMRI data.

[0144] The processing module 3 further includes:

[0145] Calculation submodule for complex conjugate Y based on diffusion attenuation image and spherical harmonics l m* (g n ), using the generalized Fourier transform, the spherical harmonic coefficients corresponding to each voxel are calculated voxel by voxel. The calculation formula of the spherical harmonic coefficients is as follows:

[0146]

[0147] Where, are the spherical harmonic coefficients, Y l m* (r) is the spherical harmonic function Y l m The complex conjugate of (r), r is the coordinate on the unit sphere, r∈ 3 , s(r) is a spherical signal, indicating different diffusion gradient directions (r = n ) corresponds to the diffusion signal attenuation value, s(r)∈R.

[0148] A deep learning network is pre-built and trained using pre-set training data. The deep learning network inputs the spherical harmonic coefficient map, and the deep learning network outputs the DTI diffusion tensor field. Therefore, input module 4 inputs the calculated spherical harmonic coefficient map into the pre-built deep learning network, and the DTI diffusion tensor field is obtained through the deep learning network. The deep learning network can be a U-net network, which is a U-shaped network.

[0149] Furthermore, the DTI quantitative parameter map includes at least an FA map and an MD map; the calculation module 5 includes:

[0150] The decomposition submodule is used to perform eigenvalue decomposition on the diffusion tensor D of the DTI diffusion tensor field to obtain eigenvalues ​​λ1, λ2, and λ3, and calculate the FA map according to formula (3) and the MD map according to formula (4), where formula (3) is as follows:

[0151]

[0152] Formula (4) is as follows:

[0153]

[0154] DTI diffusion tensor field D field ∈R M×N×L×3×3 The DTI diffusion tensor D∈R corresponding to each voxel in the image 3×3 The field map is constructed, where M×N×L is the size of the dMRI data. The diffusion tensor D is a symmetric positive definite matrix composed of six unknown parameters; the submodule converts the six unknown parameters of any diffusion tensor in the DTI diffusion tensor field {D xx ,D xy ,D xz ,D yy ,D yz ,D zz The positive definite symmetric matrix composed of} is expressed as:

[0155]

[0156] Therefore, the decomposition submodule performs eigendecomposition on the positive definite symmetric matrix of the diffusion tensor D to obtain the eigenvalues ​​λ1, λ2, and λ3, and combines the eigenvalues ​​λ1, λ2, and λ3 to calculate the FA map and MD map, thereby obtaining the DTI quantization parameter map.

[0157] Please refer to Figure 7 , Figure 7 The electronic device of the present invention is a block diagram showing the structure of an embodiment of the electronic device. The electronic device 1001 of the present invention is provided in an embodiment of the present invention, including a memory 1003 and a processor 1002. The memory 1003 stores a computer program 1004. When the processor 1002 executes the computer program 1004, the steps of any of the above-mentioned diffusion tensor imaging parameter quantification methods are implemented, including: S1, obtaining dMRI data {S0, S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ; S2, according to the diffusion coding direction g n , find the spherical harmonic function Y l m (g n ); S3, the dMRI data is transformed to obtain a diffusion attenuation image, and the complex conjugate Y of the diffusion attenuation image and the spherical harmonic function is obtained. l m* (g n ) calculate the spherical harmonic coefficient map; S4, input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field; S5, calculate the DTI quantization parameter map based on the DTI diffusion tensor field.

[0158] Please refer to Figure 8 , Figure 8 This is a schematic block diagram of the structure of an embodiment of a computer-readable storage medium of the present invention. An embodiment of the present invention further provides a computer-readable storage medium 2001, on which a computer program 1004 is stored. When the computer program 1004 is executed by the processor 1002, the steps of any of the above-mentioned diffusion tensor imaging parameter quantification methods are implemented, including: S1, obtaining dMRI data {S0, S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ; S2, according to the diffusion coding direction g n , find the spherical harmonic function Y l m (g n); S3, the dMRI data is transformed to obtain a diffusion attenuation image, and the complex conjugate Y of the diffusion attenuation image and the spherical harmonic function is obtained. l m* (g n ) calculate the spherical harmonic coefficient map; S4, input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field; S5, calculate the DTI quantization parameter map based on the DTI diffusion tensor field.

[0159] Compared with the existing technology, the beneficial effects of the present invention are: using deep learning networks and spherical harmonic functions to quantify diffusion tensor imaging parameters is faster than traditional methods and can obtain more accurate DTI quantification parameters; and compared with other diffusion tensor parameter quantization methods based on deep learning networks, it is not restricted to a specific acquisition scheme and can accurately estimate the DTI diffusion tensor field of dMRI data with different acquisition schemes, with strong generalization ability.

[0160] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAM bus dynamic RAM (RDRAM), etc.

[0161] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0162] The present invention is not limited to the above-mentioned embodiments. If various changes or modifications of the present invention do not depart from the spirit and scope of the present invention, and if these changes and modifications fall within the scope of the claims of the present invention and equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A diffusion tensor imaging parameter quantification method, characterized in that: The following steps are involved: Acquire dMRI data {S0,S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ; According to the diffusion coding direction g n , find the spherical harmonics The dMRI data is transformed to obtain the diffusion attenuation image, and the diffusion attenuation image and the complex conjugate of the spherical harmonic function are used to calculate the diffusion attenuation image. Calculate spherical harmonic coefficients plot; The spherical harmonic coefficient map is input into the constructed deep learning network to obtain the DTI diffusion tensor field; According to the DTI diffusion tensor field, the DTI quantitative parameter map is calculated; The step of performing data transformation on the dMRI data includes: The dMRI image is transformed using formula (1), which is: Where, {S(g n ), n=1,2,…,N} are diffusion attenuation images independent of b value, S0 and S b (g n ) are the non-diffusion-weighted image and diffusion-weighted image in dMRI data, b is the diffusion-weighted image S b (g n ) corresponding to the b value, b train is the b value of the training data of the deep learning network; The complex conjugate of the diffusion attenuation image and the spherical harmonic function The steps to calculate the spherical harmonic coefficients plot include: Complex conjugate based on diffusion attenuation image and spherical harmonics The generalized Fourier transform is used to calculate the spherical harmonic coefficients corresponding to each voxel. The calculation formula of the spherical harmonic coefficients is as follows: Where, are the spherical harmonic coefficients, is the spherical harmonic function The complex conjugate of , r is the coordinate on the unit sphere, r∈R 3 , s(r) is the spherical signal, indicating the direction of different diffusion gradients g n The corresponding diffusion signal attenuation value, s(r)∈R.

2. The diffusion tensor imaging parameter quantification method according to claim 1, characterized in that: The deep learning network includes: An encoder is used to extract features of the spherical harmonic coefficient graph to obtain a feature image; The decoder is used to decode the feature image extracted by the encoder to obtain a decoded feature map; The skip connection layer connects the feature image output by the encoder with the decoded feature image output by the decoder.

3. The diffusion tensor imaging parameter quantification method according to claim 1, characterized in that: The DTI quantification parameter map includes at least an FA map and an MD map; The step of calculating the DTI quantization parameter map according to the DTI diffusion tensor field includes: Perform eigenvalue decomposition on the diffusion tensor D of the DTI diffusion tensor field to obtain the eigenvalues ​​λ1, λ2, and λ3, and calculate the FA map according to formula (3), and the MD map according to formula (4), where formula (3) is as follows: Formula (4) is as follows:

4. The diffusion tensor imaging parameter quantification method according to claim 1, characterized in that: The deep learning network training process includes the following steps: (1) Training data preparation: obtain multiple dMRI data and process each dMRI data. The data processing process is as follows: (1-1) selecting a diffusion-weighted image with a preset b value and a corresponding non-diffusion-weighted image, and performing preprocessing, wherein the preprocessing process includes image denoising, Gibbs artifact removal, and bias field correction; (1-2) For the preprocessed diffusion-weighted image and the corresponding non-diffusion-weighted image, the weighted linear least squares method is used to perform pixel-by-pixel fitting to obtain the DTI diffusion tensor field, and the obtained DTI diffusion tensor field is used as the real diffusion tensor field of the dMRI data. And the true value of the network output when the dMRI data is used as the training dataset; (1-3) Apply the DTI model to the DTI diffusion tensor field and reconstruct the dMRI data using the preset diffusion gradient table to obtain a diffusion-weighted image. The DTI model formula is: Where s b is the diffusion weighted signal, s0 is the non-diffusion weighted signal, s0∈R, g is the diffusion gradient direction, g∈R 3×1 , D is the DTI diffusion tensor, D∈R 3×3 , is a positive definite symmetric matrix; (1-4) Add Rice noise to the reconstructed dMRI data: Where I is the real noise-free dMRI data after reconstruction, N1 and N2 are independent and identically distributed Gaussian noises; (1-5) For dMRI data I with Rice noise added noise , use formula (1) and formula (2) to obtain the corresponding spherical harmonic coefficient map, and use the obtained spherical harmonic coefficient map as the input of the network when dMRI data is used as the training data set; (2) Deep learning network training; (2-1) Randomly select n cases from multiple dMRI data as training sets, and the remaining cases as validation sets; (2-2) Use the training set to train the deep learning network until the network converges and the validation set loss reaches a stable level.

5. The diffusion tensor imaging parameter quantification method according to claim 4, characterized in that: The loss function of the deep learning network is: Where, They are the true diffusion tensor field corresponding to the training data and the diffusion tensor field predicted by the deep learning network, respectively.

6. A diffusion tensor imaging parameter quantification system, characterized in that: include: Acquisition module, used to acquire dMRI data {S0,S b (g n ), n=1,2,…,N}, and obtain the b value and diffusion encoding direction g corresponding to the dMRI data n ; Obtaining module, used to encode the direction g according to the diffusion n , find the spherical harmonics The processing module is used to transform the dMRI data to obtain the diffusion attenuation image and the complex conjugate of the diffusion attenuation image and the spherical harmonic function. Calculate spherical harmonic coefficients plot; The input module is used to input the spherical harmonic coefficient map into the constructed deep learning network to obtain the DTI diffusion tensor field; A calculation module, used for calculating a DTI quantitative parameter map based on the DTI diffusion tensor field; The processing module includes: The processing submodule is used to perform data transformation on the dMRI image using formula (1), wherein the formula (1) is: Where, {S(g n ), n=1,2,…,N} are diffusion attenuation images independent of b value, S0 and S b (g n ) are the non-diffusion-weighted image and diffusion-weighted image in dMRI data, b is the diffusion-weighted image S b (g n ) corresponding to the b value, b train is the b value of the training data of the deep learning network; Computational submodule for complex conjugate based on diffusion attenuation image and spherical harmonics The generalized Fourier transform is used to calculate the spherical harmonic coefficients corresponding to each voxel. The calculation formula of the spherical harmonic coefficients is as follows: Where, are the spherical harmonic coefficients, is the spherical harmonic function The complex conjugate of , r is the coordinate on the unit sphere, r∈R 3 , s(r) is the spherical signal, indicating the direction of different diffusion gradients g n The corresponding diffusion signal attenuation value, s(r)∈R.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. 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 method according to any one of claims 1 to 5 are implemented.

Images