A physical intelligent MRI diffusion weighted image denoising and quantification method

Abstract

A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method relates to diffusion magnetic resonance weighted imaging. The method comprises the following steps: 1) obtaining diffusion weighted magnetic resonance data acquired by a planar echo imaging sequence in multiple diffusion directions; 2) designing a deep learning network model and loss function for physical intelligent magnetic resonance diffusion weighted image denoising and quantification; 3) using the training data set obtained in step 1), training a deep learning network model for physical intelligent magnetic resonance diffusion weighted image denoising and quantification; 4) inputting the diffusion weighted data to be denoised into the trained network to generate the final denoising and quantification results. Through the joint application of physical regularization loss and low b-value data, the low signal-to-noise ratio problem in magnetic resonance diffusion weighted imaging is effectively solved, image denoising is achieved, and accurate quantitative results are further provided. While reducing the number of acquisitions, the image quality and ADC quantitative accuracy can be maintained or even improved.

Description

Technical Field

[0001] The invention relates to diffusion magnetic resonance weighted imaging, in particular to a physical intelligent magnetic resonance diffusion weighted image denoising and quantitative method capable of improving signal-to-noise ratio. Background Art

[0002] Diffusion-weighted magnetic resonance imaging is a non-invasive, radiation-free imaging technique that can detect the microscopic movement of water molecules in living bodies, providing important information for clinical diagnosis and research. By acquiring at least two diffusion-weighted images with different b values ​​and performing parameter fitting, the Apparent Diffusion Coefficient (ADC) map can be calculated. However, at high b values, the signal-to-noise ratio (SNR) of DWI is low. This is because as the b value increases, the signal attenuation increases, while the noise level remains constant, resulting in a significant decrease in SNR. Low SNR not only affects the visual quality of the image, but also causes noise-induced deviations in the DWI amplitude signal, which in turn affects the accuracy and repeatability of ADC calculations.

[0003] In recent years, researchers have proposed a variety of denoising algorithms, such as non-local means (NLM) and sparse representation-based methods. In addition, reconstruction methods based on structured low-rank matrix recovery, such as PLRHM and PAIR, have improved image quality. These methods have improved the quality of magnetic resonance diffusion-weighted imaging to a certain extent, but are still slow in terms of calculation speed, especially when processing data under high b-value conditions, facing the challenges of reduced signal-to-noise ratio and increased computational complexity.

[0004] At the same time, deep learning has shown great potential in magnetic resonance imaging, especially convolutional neural networks (CNNs), which have been successfully applied to improve the quality of noisy magnetic resonance imaging. In addition, the combination of physical models and deep learning has shown potential in magnetic resonance diffusion weighted imaging reconstruction, providing new ideas for breaking through the bottleneck of training data.

[0005] However, how to effectively combine physical models with image denoising under high b-value conditions to improve the accuracy of parameter estimation remains to be solved. Existing DWI denoising and parameter estimation methods are difficult to achieve high-quality image denoising and accurate parameter estimation when processing high b-value and low SNR data. Low b-value images can provide higher signal-to-noise ratio and clear structural details in diffusion-weighted magnetic resonance imaging due to their lower diffusion sensitivity. Summary of the invention

[0006] The purpose of the present invention is to provide a physical intelligent magnetic resonance diffusion weighted image denoising and quantification method that not only reduces the number of acquisitions, but also achieves better denoising results and realizes higher quality ADC image quantification in order to address the above technical problems existing in the prior art.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0008] A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method comprises the following steps:

[0009] 1) Acquire diffusion-weighted MRI data acquired using a planar echo imaging sequence with multiple diffusion directions: acquire diffusion-weighted MRI frequency domain data with multiple diffusion directions and multiple times at low b-values ​​and high b-values;

[0010] 2) Design a deep learning network model and loss function for physical intelligent MRI diffusion weighted image denoising and quantification: adopt a U-shaped network architecture, use low b-value images and high b-value images as dual-path inputs of the network, make full use of the high signal-to-noise ratio characteristics of low b-value images to provide reliable structural information and contrast reference for high b-value image denoising, and design an overall network loss function that combines the regularization loss function based on the signal attenuation physical model with the data-driven loss;

[0011] 3) Using the training data set obtained in step 1), a deep learning network model for physical intelligent magnetic resonance diffusion weighted image denoising and quantification is trained. Low b-value images and noisy high b-value images are used as network inputs, and high-quality high b-value images are used as supervisory labels. The AdamW optimizer is used for iterative training to obtain the optimal network parameters.

[0012] 4) Image denoising and ADC quantification: The low b-value image and high b-value noisy image of the sample to be processed are input into the trained network to obtain the denoised high b-value image, and the apparent diffusion coefficient ADC is calculated in combination with the low b-value image.

[0013] In step 1), diffusion-weighted magnetic resonance data acquired by a planar echo imaging sequence with multiple diffusion directions are obtained, including obtaining low b-value, multi-diffusion direction, and multi-order planar echo diffusion-weighted magnetic resonance frequency domain data. and high b-value, multi-diffusion direction, multi-order echo-planar diffusion-weighted MRI frequency domain data ,in, represents a set of complex numbers, FE represents the frequency of the frequency domain data, PE represents the length of the phase encoding dimension, H represents the number of channels, KS represents the number of acquisitions of low b-value frequency domain data, K represents the number of acquisitions of high b-value frequency domain data, Q represents the number of diffusion directions, and N represents the number of layers; ,for , N represents the number of layers, It represents the frequency domain data of the nth layer with low b value collected KS times in Q diffusion directions of H channels; ,for , N represents the number of layers, It represents the frequency domain data of the nth layer with high b value acquired in K times of H channels in Q diffusion directions;

[0014] The frequency domain data for each acquisition number and layer used for training and testing needs to be re-gridded; specifically, the non-uniformly sampled data is adjusted using the trapezoidal re-gridding method; then, the weight coefficients of the coils in each layer are calculated and applied to the re-gridded frequency domain data to reconstruct the complete image in the case of undersampling; then, the frequency domain data is converted to the image domain by applying a two-dimensional inverse Fourier transform;

[0015] For low b-value and high b-value data, coil sensitivity modulation and weighted summation of coil weight coefficients are performed on the image data of each channel to achieve channel data merging; all collected data are selected, and the amplitude images of multiple diffusion directions are processed using the geometric mean method to obtain low b-value image data and high b-value image data ,in, represents a set of real numbers, FE represents the frequency of frequency domain data, PE represents the length of the phase encoding dimension, and N represents the number of layers; ,for , N represents the number of layers, Represents the image data of the nth layer with low b value; ,for , N represents the number of layers, Represents the image data of the nth layer with high b value;

[0016] In order to generate a specific noise image, two random selections are made from the data collected multiple times with high b values, and the images in multiple diffusion directions are processed using the geometric mean method to obtain the noisy image data with high b values. ; For high b-value noisy image data: ,for , N represents the number of layers, Represents the image data of the nth layer with high b value noise; in supervised training, the input is the image data with low b value , noisy image data with high b values , the label is the image data with high b value , together constitute the training data set.

[0017] In step 2), the physical intelligent magnetic resonance diffusion weighted image denoising and quantification deep learning network model adopts a U-shaped network architecture; the architecture consists of four encoder-decoder modules, integrating residual connections and nonlinear activation modules, and enhancing feature extraction and reconstruction capabilities through layer-by-layer downsampling and upsampling operations;

[0018] The data-driven loss function is defined as follows:

[0019] (1)

[0020] Among them, I represents the total number of samples, is the image data with high b value obtained by denoising the i-th sample network. represents the image data of the high b value of the i-th sample reference, Represents a high b-value image data set The maximum possible value of , Σ represents the summation operation;

[0021] The physical regularization loss function is designed according to the signal attenuation model, which is as follows:

[0022] (2)

[0023] Where S is the signal strength after diffusion weighting, is the unweighted signal strength, e is the base of the natural logarithm, is the gyromagnetic ratio, G is the gradient field strength, is the duration of the gradient pulse, is the time interval between gradient pulses, D is the diffusion coefficient, which reflects the diffusion rate of water molecules in the tissue;

[0024] Define the b value: , the diffusion single exponential model is obtained as follows:

[0025] (3)

[0026] Where S is the signal strength after diffusion weighting, is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and D is the diffusion coefficient;

[0027] In actual diffusion MRI, the apparent diffusion coefficient ADC is measured instead of directly measuring the physical diffusion coefficient D; therefore, the diffusion single exponential model is reformulated as:

[0028] (4)

[0029] Where S is the signal strength after diffusion weighting, is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and ADC is the apparent diffusion coefficient;

[0030] ADC-based calculations:

[0031] (5)

[0032] Where ADC is the apparent diffusion coefficient, b is the diffusion sensitivity factor, ln represents the natural logarithm, S is the signal intensity after diffusion weighting, is the unweighted signal strength;

[0033] The corresponding physical regularization loss function is:

[0034] (6)

[0035] in, For high b values, is a low b value, I represents the total number of samples, Σ represents the summation operation, ln represents the natural logarithm, is the image data with high b value obtained by denoising the i-th sample network. represents the image data of the high b value of the i-th sample reference, Represents the image data of the low b value of the i-th sample;

[0036] The overall network loss function is defined as:

[0037] (7)

[0038] in, Represents the internal parameter set of the entire network, and the trained network parameters are represented as the optimal parameter set , , is the weight coefficient, represents the data-driven loss function, represents the physical regularization loss function.

[0039] In step 3), the training data set generated in step 1) is input into the deep learning network of physical intelligent magnetic resonance diffusion weighted image denoising and quantification, and the loss function is minimized through iterative training using the AdamW optimizer. , to obtain the optimal parameter set The optimal network model under .

[0040] In step 4), the acquired echo-planar diffusion-weighted magnetic resonance data are used to , I represents the total number of samples, and the high b-value noisy image and low b-value image of the i-th sample are taken respectively, and these two images are input into the trained physical intelligent magnetic resonance diffusion weighted image denoising and quantitative deep learning network to obtain the denoised high b-value image data of the i-th sample ; Then use the denoised high b value image data of the i-th sample and the image data of the low b value of the i-th sample The apparent diffusion coefficient ADC of the i-th sample is calculated by formula (5):i .

[0041] For the present invention, the high b value can be defined as ≥1000 s / mm 2 , the low b value definition range can be <1000 s / mm 2 .

[0042] Compared with the prior art, the present invention has the following outstanding technical effects and advantages: the present invention innovatively combines physical model constraints with deep learning, designs a physical intelligent network and a dual loss function mechanism, and introduces low b-value image information as network input to achieve a better denoising effect. To ensure image quality, the loss function is also physically regularized. Incorporating the physical properties of diffusion-weighted imaging makes the calculation results of the apparent diffusion coefficient ADC more accurate. Compared with the traditional pure data-driven method, this physically guided denoising method can better maintain the detailed information and true diffusion characteristics of the image, while ensuring the physical rationality of the ADC quantitative calculation. Experiments show that the present invention can reduce the number of acquisitions to 1 / 6 of the original while still achieving the denoising effect and ADC quantitative calculation accuracy comparable to that of high b-value reference images. The method of the present invention can significantly shorten the examination time, improve patient comfort, and reduce the examination cost, and effectively solve the problem of low signal-to-noise ratio in magnetic resonance diffusion-weighted imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The diffusion-weighted MRI images and generated ADC images of the three diffusion directions of the prostate under different b values. (a) is a diffusion-weighted MRI image with a b value of 50 s / mm² in the three diffusion directions of the prostate, (b) is a diffusion-weighted MRI noisy image with a b value of 1000 s / mm² in the three diffusion directions of the prostate, and (c) is an ADC image generated from (a) and (b).

[0044] Figure 2 The diffusion-weighted magnetic resonance images of the prostate in three diffusion directions after denoising and the generated high-quality ADC images. (a) is a diffusion-weighted magnetic resonance image with a b value of 50 s / mm² in three diffusion directions of the prostate, (b) is the result of denoising a diffusion-weighted magnetic resonance noisy image with a b value of 1000 s / mm² in three diffusion directions of the prostate by the method described in the embodiment, and (c) is a high-quality ADC image generated by (a) and (b).

[0045] Figure 3The diffusion-weighted MRI images of the prostate in three diffusion directions acquired multiple times and the generated reference ADC images. (a) is a diffusion-weighted MRI image with a b value of 50 s / mm² in three diffusion directions of the prostate, (b) is a diffusion-weighted MRI reference image with a b value of 1000 s / mm² in three diffusion directions of the prostate, and (c) is a reference ADC image generated from (a) and (b).

[0046] Figure 4 Deep learning network architecture for physical intelligence MRI diffusion weighted image denoising and quantification methods. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in the following embodiments in conjunction with the accompanying drawings.

[0048] Example: De-noising and quantification of diffusion-weighted imaging data in three diffusion directions of the prostate area.

[0049] 1) Diffusion-weighted MRI data of 90 subjects were collected using a 3.0 Tesla MRI scanner. The frequency domain data of the prostate area were collected using 4 acquisitions of echoplanar diffusion MRI weighted in 3 diffusion directions with a b value of 50 s / mm² (low b value). ,in, represents a complex number set, FE represents the frequency of the frequency domain data, PE represents the length of the phase encoding dimension, FE=100, PE=100, H is the number of channels, , N represents the number of layers, ; ,for , N represents the number of layers, The frequency domain data of the nth layer with low b value collected in Q diffusion directions by KS acquisitions of H channels are specifically represented in this embodiment as the frequency domain data of the nth layer with b value of 50 s / mm² collected in 3 diffusion directions by 4 acquisitions of H channels; the frequency domain data of the nth layer with low b value of 1000 s / mm² (high b value) collected in 3 diffusion directions and 12 acquisitions of echo plane diffusion magnetic resonance weighted frequency domain data are represented as , ,for , N represents the number of layers, It represents the frequency domain data of the nth layer with high b value in Q diffusion directions acquired by K times of H channels. In this embodiment, it is specifically represented as the frequency domain data of the nth layer with b value of 1000 s / mm² in 3 diffusion directions acquired by 12 times of H channels.

[0050] The frequency domain data for each acquisition number and layer used for training and testing needs to be re-gridded; specifically, the non-uniformly sampled data is adjusted using the trapezoidal re-gridding method; the weight coefficients of the coils in each layer are calculated and applied to the re-gridded frequency domain data to reconstruct the complete image in the case of undersampling; then, the frequency domain data is converted to the image domain by applying a two-dimensional inverse Fourier transform;

[0051] For the b value of 50 s / mm² (low b value) and 1000 s / mm² (high b value) data, the coil sensitivity modulation and the weighted sum of the coil weight coefficient are performed on the image data of each channel to achieve the merging of channel data; all the collected data are selected, and the amplitude images of multiple diffusion directions are processed by the geometric average method, and finally the image data with a low b value of 50 s / mm² is obtained. and high b-value image data with b-value of 1000 s / mm² ,in, represents the set of real numbers; ,for , represents the image data of the nth layer with a low b value. In this embodiment, the low b value adopts a b value of 50 s / mm²; ,for , represents image data of the nth layer with a high b value, and in this embodiment represents image data of the nth layer with a b value of 1000 s / mm².

[0052] In order to generate a specific noise image, 2 data with a b value of 1000 s / mm² are randomly selected from the 12 data collected, and the amplitude images in the three diffusion directions are processed using the geometric mean method. Finally, the image data with a high b value of 1000 s / mm² and noise can be obtained. ; For noisy image data with a b value of 1000 s / mm²: ,for , represents the image data of the nth layer with high b value and noise, and in this embodiment specifically represents the image data of the nth layer with b value of 1000 s / mm² and noise; in supervised training, the input is the image data with low b value , noisy image data with high b values , the label is the image data with high b value , together constitute the training data set.

[0053] 2) The physical intelligent magnetic resonance diffusion weighted image denoising and quantitative deep learning network model adopts a U-shaped network architecture; the architecture is as follows Figure 4As shown in the figure, it consists of four encoder-decoder modules, integrating residual connection and nonlinear activation modules, and enhancing feature extraction and reconstruction capabilities through layer-by-layer downsampling and upsampling operations. The network uses 2D convolution as the input and output interface. In the encoding path, it gradually compresses the feature dimensions through continuous downsampling to achieve multi-scale feature extraction, and the decoding path reconstructs the feature space dimensions through corresponding upsampling operations; the residual connection mechanism between each level realizes the cross-layer transmission of features, effectively retaining the image detail information, and the nonlinear activation module set in the deepest layer of the network further enhances the feature expression ability of the model;

[0054] The input image is preliminarily feature extracted through a 2D convolutional layer with a convolution kernel of 3×3 and an output channel number of 60. The encoder consists of 4 layers of encoding blocks, and the number of encoding blocks is 2, 2, 4, and 6 from shallow to deep. The encoding block uses CascadedGazeBlock (Ghasemabadi, Amirhosein, et al., "Cascadedgaze: Efficiency in global context extraction for imagerestoration", published in "arXiv" (2401.15235, 2024)). Each layer is followed by a downsampling layer; the intermediate nonlinear activation module consists of 10 NAFBlocks (Zamir SW, Arora A., et al., "Restormer: Efficient transformer for high-resolution image restoration", published in "Proceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition" (5728-5739, 2022)). The decoder performs the reverse operation, restoring the spatial resolution of the image through an upsampling layer and combining it with the encoder output through a skip connection. The decoder consists of 4 layers of decoding blocks, with 2 decoders in each layer. Finally, a 2D convolutional layer with a convolution kernel of 3×3 is used to output the denoised image.

[0055] The data-driven loss function is defined as follows:

[0056] (1)

[0057] Among them, I represents the total number of samples, that is, the total number of levels of all subjects, I=2744, The image data with a b value of 1000 s / mm² obtained by denoising the i-th sample network, represents the image data of the i-th sample reference with a b value of 1000 s / mm², Indicates a data set of images with a b value of 1000 s / mm² The maximum possible value of , Σ represents the summation operation;

[0058] The physical regularization loss function is designed according to the signal attenuation model, which is as follows:

[0059] (2)

[0060] Where S is the signal strength after diffusion weighting, is the unweighted signal strength, e is the base of the natural logarithm, is the gyromagnetic ratio, G is the gradient field strength, is the duration of the gradient pulse, is the time interval between gradient pulses, D is the diffusion coefficient, which reflects the diffusion rate of water molecules in the tissue;

[0061] Define the b value: , the diffusion single exponential model is obtained as follows:

[0062] (3)

[0063] Where S is the signal strength after diffusion weighting, is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and D is the diffusion coefficient;

[0064] In actual diffusion MRI, the apparent diffusion coefficient ADC is measured instead of directly measuring the physical diffusion coefficient D; therefore, the diffusion single exponential model is reformulated as:

[0065] (4)

[0066] Where S is the signal strength after diffusion weighting, is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and ADC is the apparent diffusion coefficient;

[0067] ADC-based calculations:

[0068] (5)

[0069] Where ADC is the apparent diffusion coefficient, b is the diffusion sensitivity factor, ln represents the natural logarithm, S is the signal intensity after diffusion weighting, is the unweighted signal strength;

[0070] The corresponding physical regularization loss function is:

[0071] (6)

[0072] in, Indicates a high b value, which in this embodiment is 1000 s / mm²; represents the low b value, which is 50 s / mm² in this embodiment; I represents the total number of samples, Σ represents the summation operation, ln represents the natural logarithm, is the image data with high b value obtained by denoising the i-th sample network. represents the image data of the high b value of the i-th sample reference, Represents the image data of the low b value of the i-th sample;

[0073] The overall network loss function is defined as:

[0074] (7)

[0075] in, Represents the internal parameter set of the entire network, and the trained network parameters are represented as the optimal parameter set , , is the weight coefficient. In this embodiment , , represents the data-driven loss function, represents the physical regularization loss function.

[0076] 3) Input the training data set generated in step 1) into the deep learning network of physical intelligent magnetic resonance diffusion weighted image denoising and quantification, and use the AdamW optimizer to minimize the loss function through iterative training , to obtain the optimal parameter set The optimal network model under .

[0077] 4) The acquired echo-planar diffusion-weighted MRI data are used for , I represents the total number of samples, and the image of the i-th sample with b=50 s / mm² is as follows Figure 1 The image with noise is shown in (a) and b=1000 s / mm². Figure 1 As shown in (b), these two images are input into the trained physical intelligent magnetic resonance diffusion weighted image denoising and quantitative deep learning network to obtain the denoised high b value image data of the i-th sample with b=1000 s / mm². like Figure 2As shown in (b) in the figure, the denoised high b value image data of the i-th sample is then used. and the image data of the low b value of the i-th sample The apparent diffusion coefficient ADC of the i-th sample is calculated by formula (5): i ,like Figure 2 As shown in (c) in the figure. Figure 1 As can be seen from the figure, the ADC image calculated from the prostate diffusion-weighted MRI images acquired twice is as follows: Figure 1 As shown in (c), the noise seriously affects the calculation accuracy of ADC. After being processed by the method of the present invention, the calculated ADC image is as follows Figure 2 As shown in (c), it shows excellent image quality and quantitative accuracy. The 12 reference images collected as controls are shown in Figure 3 As shown, b=50 s / mm² image is as follows Figure 3 As shown in (a), b=1000 s / mm² reference image Figure 3 As shown in (b) in the figure, the reference ADC image calculated thereby is shown in Figure 3 As shown in (c) in the figure. Figure 2 and 3 It can be found that the method of the present invention can obtain image quality and quantitative accuracy comparable to the reference image with only two acquisitions.

[0078] In order to verify the effectiveness of the physical intelligent MRI diffusion weighted image denoising and quantification method proposed in the present invention, the method of the present invention greatly reduces the number of acquisitions from the traditional 12 times to 2 times. Under the condition of 2 acquisitions, the image with b=1000 s / mm² originally has serious noise, which leads to reduced ADC calculation accuracy. However, the results after being processed by the method of the present invention are as follows: Figure 2 As shown, Figure 2 It can be seen that the present invention uses physical regularization loss and combines it with low b value b=50 s / mm² data to denoise the high b value image and significantly improve the image signal-to-noise ratio. Figure 3 The results of the reference method requiring 12 acquisitions are shown. By comparison, it can be seen that the image processed under the condition of 2 acquisitions by the method of the present invention reaches a level similar to that of the reference method requiring 12 acquisitions in terms of image quality and ADC quantitative accuracy. The experiment shows that the present invention can maintain or even improve image quality and ADC quantitative accuracy while reducing the number of acquisitions. The present invention greatly shortens the examination time, improves patient comfort, and reduces the cost of examination. The present invention effectively solves the problem of low signal-to-noise ratio in magnetic resonance diffusion weighted imaging, which fully proves that the method of the present invention has the technical advantage of significantly reducing the number of acquisitions while ensuring image quality.

[0079] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method, characterized in that: The following steps are involved: 1) Acquiring diffusion-weighted magnetic resonance data acquired using a planar echo imaging sequence with multiple diffusion directions: acquiring planar echo diffusion-weighted magnetic resonance frequency domain data with multiple diffusion directions and multiple times at low b-values ​​and high b-values; 2) Design a deep learning network model and loss function for physical intelligent MRI diffusion weighted image denoising and quantification: adopt a U-shaped network architecture, use low b-value images and high b-value images as dual-path inputs of the network, make full use of the high signal-to-noise ratio characteristics of low b-value images to provide reliable structural information and contrast reference for high b-value image denoising, and design an overall network loss function that combines the regularization loss function based on the signal attenuation physical model with the data-driven loss; 3) Using the training data set obtained in step 1), a deep learning network model for physical intelligent magnetic resonance diffusion weighted image denoising and quantification is trained, low b-value images and noisy high b-value images are used as network inputs, high-quality high b-value images are used as supervision labels, and the AdamW optimizer is used for iterative training to obtain the optimal network parameters; 4) Image denoising and ADC quantification: The low b-value image and high b-value noisy image of the sample to be processed are input into the trained network to obtain the denoised high b-value image, and the apparent diffusion coefficient ADC is calculated in combination with the low b-value image.

2. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 1, characterized in that: In step 1), the acquisition of diffusion-weighted magnetic resonance data acquired by a planar echo imaging sequence with multiple diffusion directions includes acquiring low b-value, multi-diffusion direction, and multi-order planar echo diffusion-weighted magnetic resonance frequency domain data. and high b-value, multi-diffusion direction, multi-order echo-planar diffusion-weighted MRI frequency domain data in, represents a set of complex numbers, FE represents the frequency of the frequency domain data, PE represents the length of the phase encoding dimension, H represents the number of channels, KS represents the number of acquisitions of low b-value frequency domain data, K represents the number of acquisitions of high b-value frequency domain data, Q represents the number of diffusion directions, and N represents the number of layers; For n∈[1,N], N represents the number of layers, It represents the frequency domain data of the nth layer with low b value collected KS times in Q diffusion directions of H channels; For n∈[1,N], N represents the number of layers, It represents the frequency domain data of the nth layer with high b value acquired by K times of H channels in Q diffusion directions.

3. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 1, characterized in that: In step 2), the physical intelligent magnetic resonance diffusion weighted image denoising and quantitative deep learning network model is used for training and testing the frequency domain data of each acquisition number and layer, which needs to be re-gridded; specifically, the trapezoidal re-gridding method is used to adjust the non-uniformly sampled data; Calculate the weight coefficients of the coils in each slice and apply these weight coefficients to the re-gridded frequency domain data to reconstruct the complete image under undersampling; apply a two-dimensional inverse Fourier transform to the frequency domain data to convert it to the image domain; For low b-value and high b-value data, coil sensitivity modulation and weighted summation of coil weight coefficient are performed on the image data of each channel to achieve channel data merging; Select all collected data, use geometric mean method to process amplitude images of multiple diffusion directions, and obtain image data with low b value. and high b-value image data in, represents a set of real numbers, FE represents the frequency of frequency domain data, PE represents the length of the phase encoding dimension, and N represents the number of layers; For n∈[1,N], N represents the number of layers, Represents the image data of the nth layer with low b value; For n∈[1,N], N represents the number of layers, Represents the image data of the nth layer with high b value; In order to generate a specific noise image, two random selections are made from the data collected multiple times with high b values, and the images in multiple diffusion directions are processed using the geometric mean method to obtain the noisy image data with high b values. For noisy image data with high b values: For n∈[1,N], N represents the number of layers, Represents the image data of the nth layer with high b value noise.

4. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 1, characterized in that In step 2), the physical intelligent magnetic resonance diffusion weighted image denoising and quantitative deep learning network model is to input the image data M with low b value in supervised training. b_low , noisy image data M with high b value b_high_noisy , the label is the image data M with high b value b_high .

5. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 1, characterized in that: In step 2), the physical intelligent magnetic resonance diffusion weighted image denoising and quantification deep learning network model adopts a U-shaped network architecture; the architecture consists of four encoder-decoder modules, integrating residual connections and nonlinear activation modules, and enhancing feature extraction and reconstruction capabilities through layer-by-layer downsampling and upsampling operations; The data-driven loss function is defined as follows: Among them, I represents the total number of samples, is the image data with high b value obtained by denoising the i-th sample network. represents the image data of the high b value of the i-th sample reference, Represents the high b-value image data set M b_high The maximum possible value of , Σ represents the summation operation; The physical regularization loss function is designed according to the signal attenuation model, which is as follows: Where S is the signal intensity after diffusion weighting, S0 is the unweighted signal intensity, e is the base of the natural logarithm, γ is the gyromagnetic ratio, G is the gradient field strength, δ is the duration of the gradient pulse, Δ is the time interval between gradient pulses, and D is the diffusion coefficient, which reflects the diffusion rate of water molecules in tissues; Define the b value: The diffusion single exponential model is obtained as follows: S=S0·e -b·D (3) Where S is the diffusion-weighted signal intensity, S0 is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and D is the diffusion coefficient; In actual diffusion MRI, the apparent diffusion coefficient ADC is measured instead of directly measuring the physical diffusion coefficient D; therefore, the diffusion single exponential model is reformulated as: S=S0·e -b·ADC (4) Where S is the diffusion-weighted signal intensity, S0 is the unweighted signal intensity, e is the base of the natural logarithm, b is the diffusion sensitivity factor, and ADC is the apparent diffusion coefficient; ADC-based calculations: Where ADC is the apparent diffusion coefficient, b is the diffusion sensitivity factor, ln represents the natural logarithm, S is the signal intensity after diffusion weighting, and S0 is the unweighted signal intensity; The corresponding physical regularization loss function is: Among them, b high For high b values, b low is a low b value, I represents the total number of samples, Σ represents the summation operation, ln represents the natural logarithm, is the image data with high b value obtained by denoising the i-th sample network. represents the image data of the high b value of the i-th sample reference, Represents the image data of the low b value of the i-th sample; The overall network loss function is defined as: Among them, Θ represents the internal parameter set of the entire network, and the trained network parameters are represented as the optimal parameter set α and β are weight coefficients, represents the data-driven loss function, represents the physical regularization loss function.

6. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 1, characterized in that: In step 3), the training data set obtained in step 1) is input into the deep learning network of physical intelligent magnetic resonance diffusion weighted image denoising and quantification, and the loss function is minimized through iterative training using the AdamW optimizer. To obtain the optimal parameter set The optimal network model under .

7. A physical intelligent magnetic resonance diffusion weighted image denoising and quantification method as claimed in claim 5, characterized in that: In step 4), the specific steps of image denoising and ADC quantification are as follows: for i∈[1,I], where I represents the total number of samples, the high b-value noisy image and low b-value image of the i-th sample are respectively taken from the acquired echo-planar diffusion-weighted magnetic resonance data, and the two images are input into the trained physical intelligent magnetic resonance diffusion-weighted image denoising and quantification deep learning network to obtain the denoised high b-value image data of the i-th sample. Then use the denoised high b value image data of the i-th sample and the image data of the low b value of the i-th sample By formula The apparent diffusion coefficient ADCi of the i-th sample is calculated.

Images