阅读说明：本技术 一种用于扩散加权磁共振成像的数据处理方法和系统 (Data processing method and system for diffusion weighted magnetic resonance imaging ) 是由 朱瑞星 张志遵 吕孟叶 于 2021-08-11 设计创作，主要内容包括：本发明提供一种用于扩散加权磁共振成像的数据处理方法和系统,使用扩散加权成像序列获取各方向的扩散加权的磁共振图像；将磁共振图像输入已经训练好的深度学习网络模型中运行,分别输出对应各方向的正参数和高分辨率图像；对于每个方向的所述高分辨率图像,采用对应方向的所述矫正参数进行矫正,以获得各个方向的具有高分辨率的扩散加权的所述磁共振图像。既得到每个方向图像的高分辨率估计,也提升了配准的准确性,有利于更准确的病情诊断和更深入的科学研究。也可以用于取得相似成像效果的前提下,缩短成像时间,提高机器利用率,减少检查成本,提升MRI普及性。(The invention provides a data processing method and a system for diffusion weighted magnetic resonance imaging, which uses a diffusion weighted imaging sequence to obtain diffusion weighted magnetic resonance images in all directions; inputting the magnetic resonance image into a trained deep learning network model for operation, and respectively outputting positive parameters and high-resolution images corresponding to all directions; and for the high-resolution image in each direction, correcting by adopting the correction parameters in the corresponding direction to obtain the magnetic resonance image with high-resolution diffusion weighting in each direction. The high-resolution estimation of each direction image is obtained, the registration accuracy is improved, and more accurate disease diagnosis and deeper scientific research are facilitated. The method can also be used for shortening the imaging time, improving the machine utilization rate, reducing the inspection cost and improving the MRI popularity on the premise of obtaining similar imaging effects.)

1. A data processing method for diffusion weighted magnetic resonance imaging, comprising the steps of:

step A1, acquiring diffusion weighted magnetic resonance images of each direction by using a diffusion weighted imaging sequence;

step A2, inputting the magnetic resonance image into a trained deep learning network model for operation, and respectively outputting correction parameters and high-resolution images corresponding to each direction;

step A3, for the high-resolution image in each direction, performing rectification by using the rectification parameters in the corresponding direction to obtain the magnetic resonance image with high-resolution diffusion weighting in each direction.

2. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 1, wherein the training process of the deep learning network model comprises the steps of:

a step B1 of acquiring a diffusion-weighted magnetic resonance image of each direction with a general resolution as a first image data set, and acquiring a diffusion-weighted magnetic resonance image with a high resolution corresponding to the first image data set as a second image data set, and a diffusion-weighted-free magnetic resonance image corresponding to the first image data set as a third image data set;

step B2, constructing the deep learning network model;

step B3, inputting the first image data set into the deep learning network model, where the deep learning network model takes first feature data as a first output and second feature data as a second output, the first feature data is a predicted correction parameter in each direction, and the second feature data is a predicted high-resolution image in each direction;

a step B3 of rectifying the first feature data based on the respective orientation for the images in the first image data set;

a step B4 of calculating a first loss value of the rectified first image data set and the rectified third image data set based on a first loss function, and calculating a second loss value of the second feature data and the second image data set based on a second loss function;

and step B5, optimizing parameters of the deep learning network model according to the first loss value and the second loss value until the first loss value and the second loss value reach a preset standard, and finishing training the deep learning network model.

3. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, wherein said step B1 further comprises: pre-registering the first image data set by taking a third image data set as a reference to acquire third characteristic data comprising pre-distortion parameters in all directions;

the step B3 further includes: using the first image data set and the third feature data as input of the deep learning network model.

4. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, wherein the deep learning network model is a convolutional neural network model comprising:

a trunk extraction module, which takes the first image data set as input and comprises a plurality of convolution layers connected in series in sequence;

the first branch module takes the output of the trunk extraction module as input and outputs the first characteristic data, and the first branch module comprises a plurality of serially connected convolution layers and an upper sampling layer;

and the second branch module takes the output of the trunk extraction module as input and outputs the second characteristic data, and the second characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer.

5. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, wherein the deep learning network model is a single Unet neural network model comprising:

the encoding module takes the first image data set as input and comprises a plurality of encoding layers which are connected in series, wherein the encoding layers comprise convolutional layers and downsampling layers;

the first decoding module takes the output of the coding module as input and outputs the first characteristic data, and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer;

and the second decoding module takes the output of the coding module as input and outputs the second characteristic data, and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer.

6. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, wherein the deep learning network model is a dual-Unet neural network model comprising:

the first Unet module is composed of an encoding module and a decoding module, the first image data set is used as the input of the encoding module, and the second characteristic data is used as the output of the decoding module;

and the second Unet module consists of an encoding module and a decoding module, the second characteristic data is used as the input of the encoding module, and the first characteristic data is used as the output of the decoding module.

7. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, characterized in that the high resolution diffusion weighted magnetic resonance image in the second image data set is acquired by a method employing multiple excitations and/or multiple averages.

8. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, characterized in that the first image data set is derived based on a high resolution diffusion weighted magnetic resonance image simulation process in the second image data set.

9. A data processing method for diffusion weighted magnetic resonance imaging as claimed in claim 2, wherein the first loss function is a mutual information function and the second loss function is an L1 norm function or an L2 norm function.

10. A data processing system for diffusion weighted magnetic resonance imaging, comprising a data processing method for diffusion weighted magnetic resonance imaging as claimed in any one of claims 1-9, comprising:

the image acquisition module is used for acquiring diffusion weighted magnetic resonance images in all directions by using a diffusion weighted imaging sequence;

the characteristic extraction module is connected with the image acquisition module and is used for inputting the magnetic resonance image into a trained deep learning network model for operation to respectively obtain correction parameters and high-resolution images corresponding to all directions;

and the image correction module is connected with the feature extraction module and is used for correcting the high-resolution image in each direction by adopting the correction parameters in the corresponding direction so as to obtain the diffusion-weighted magnetic resonance image with high resolution in each direction.

Technical Field

The invention relates to the technical field of medical image processing, in particular to a data processing method and a data processing system for diffusion weighted magnetic resonance imaging.

Background

Diffusion weighted magnetic resonance imaging (DWI) is a valuable imaging modality that can be used for clinical diagnostics and scientific research. The imaging means utilizes a diffusion gradient magnetic field to enable a magnetic resonance image to reflect the diffusion speed of water molecules, so that information such as bleeding, ischemia, edema, nerve fiber distribution and the like can be obtained. In diffusion weighted magnetic resonance imaging, in order to comprehensively obtain information in multiple diffusion directions, it is generally necessary to sequentially acquire data using diffusion gradient magnetic fields in multiple directions to obtain diffusion weighted magnetic resonance images in each direction, and then combine the images by an algorithm for further calculation. Generally speaking, the diffusion weighted magnetic resonance images in different directions have a certain geometrical spatial inconsistency, on one hand because the patient may have involuntary movements, and on the other hand, the more important factor is that the gradient magnetic fields in different directions used may generate eddy currents, so that the diffusion weighted magnetic resonance images in different directions naturally have different spatial deformations, such as stretching/compression, translation, and torsion. For this reason, the currently commonly used technique is classical image registration (registration), which generally uses an image without diffusion weighting as a reference, estimates deformation fields (deformation fields) of diffusion weighted maps in different directions, then aligns the diffusion weighted magnetic resonance images in different directions to the reference space, and finally performs an averaging operation to obtain an MD (Mean diffusion map) for diagnosis, and may further calculate FA (Fractional anisotropy), ADC (apparent diffusion coefficient) to aid diagnosis and analysis. However, the registration and averaging operations do not take full advantage of much of the unique information of differently directionally diffusion weighted magnetic resonance images, at the expense of resolution. Secondly, because the resolution of the original image is very low, the estimated deformation field is usually large in error, further resulting in poor image registration effect and image blurring.

Disclosure of Invention

Based on the prior art, the invention provides a data processing method and a data processing system for diffusion weighted magnetic resonance imaging, and aims to solve the technical problems of resolution loss, large deformation field error, image blurring and the like caused by image registration in the prior art.

A data processing method for diffusion weighted magnetic resonance imaging, comprising the steps of:

step A1, acquiring diffusion weighted magnetic resonance images of each direction by using a diffusion weighted imaging sequence;

step A2, inputting the magnetic resonance image into the trained deep learning network model for operation, and respectively outputting positive parameters and high-resolution images corresponding to all directions;

step A3, for the high-resolution image in each direction, performing rectification by using the rectification parameters in the corresponding direction to obtain the magnetic resonance image with high-resolution diffusion weighting in each direction.

Further, the training process of the deep learning network model comprises the following steps:

a step B1 of acquiring a diffusion-weighted magnetic resonance image of each direction with a general resolution as a first image data set, and acquiring a diffusion-weighted magnetic resonance image with a high resolution corresponding to the first image data set as a second image data set, and acquiring a diffusion-free weighted magnetic resonance image corresponding to the first image data set as a third image data set;

step B2, constructing a deep learning network model;

step B3, inputting the first image data set into a deep learning network model, wherein the deep learning network model takes the first characteristic data as a first output and takes the second characteristic data as a second output, the first characteristic data is the predicted correction parameters of each direction, and the second characteristic data is the predicted high-resolution image of each direction;

a step B3 of normalizing the first feature data based on the respective orientation for the images in the first image data set;

step B4, calculating a first loss value of the rectified first image data set and the rectified third image data set based on the first loss function, and calculating a second loss value of the second feature data and the second image data set based on the second loss function;

and step B5, optimizing parameters of the deep learning network model according to the first loss value and the second loss value until the first loss value and the second loss value reach preset standards, and finishing training the deep learning network model.

Further, step B1 further includes: pre-registering the first image data set by taking a third image data set as a reference to acquire third characteristic data comprising pre-distortion parameters in all directions;

step B3 further includes: and taking the first image data set and the third feature data as input of the deep learning network model.

Further, the deep learning network model is a convolutional neural network model, and includes:

a trunk extraction module, which takes the first image data set as input and comprises a plurality of convolution layers connected in series in sequence;

the first branch module takes the output of the trunk extraction module as input and outputs first characteristic data, and the first characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer;

and the second branch module takes the output of the trunk extraction module as input and outputs second characteristic data, and the second characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer.

Further, the deep learning network model is a single Unet neural network model, including:

the encoding module takes a first image data set as input and comprises a plurality of encoding layers which are connected in series, wherein the encoding layers comprise a convolution layer and a down-sampling layer;

the first decoding module takes the output of the coding module as input and outputs first characteristic data, and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer;

and the second decoding module takes the output of the coding module as input and outputs second characteristic data and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer.

Further, the deep learning network model is a dual-Unet neural network model, including:

the first Unet module consists of an encoding module and a decoding module, takes a first image data set as the input of the encoding module, and takes second characteristic data as the output of the decoding module;

and the second Unet module consists of an encoding module and a decoding module, the second characteristic data is used as the input of the encoding module, and the first characteristic data is used as the output of the decoding module.

Further, a high resolution diffusion weighted magnetic resonance image in the second image data set is acquired by using a multi-shot and/or multi-average method.

Further, the first image data set is derived based on a high resolution diffusion weighted magnetic resonance image simulation process in the second image data set.

Further, the first loss function is a mutual information function, and the second loss function is a norm function of L1 or a norm function of L2.

A data processing system for diffusion weighted magnetic resonance imaging, comprising the aforementioned data processing method for diffusion weighted magnetic resonance imaging, comprising:

the image acquisition module is used for acquiring diffusion weighted magnetic resonance images in all directions by using a diffusion weighted imaging sequence;

the characteristic extraction module is connected with the image acquisition module and is used for inputting the magnetic resonance image into the trained deep learning network model for operation and respectively obtaining correction parameters and high-resolution images corresponding to all directions;

and the image correction module is connected with the feature extraction module and is used for correcting the high-resolution image in each direction by adopting the correction parameters in the corresponding direction so as to obtain the diffusion weighted magnetic resonance image with high resolution in each direction.

The beneficial technical effects of the invention are as follows: the deep learning network model is used for simultaneously predicting the correction parameters (namely the deformation field) and the magnetic resonance image with high resolution diffusion weighting, complementary information of the magnetic resonance images with diffusion weighting in different directions is fully utilized, super-resolution calculation is carried out while registration is carried out, high resolution estimation of the image in each direction is obtained, registration accuracy is improved, and more accurate disease diagnosis and deeper scientific research are facilitated. The method can also be used for shortening the imaging time, improving the machine utilization rate, reducing the inspection cost and improving the MRI popularity on the premise of obtaining similar imaging effects.

Drawings

FIG. 1 is a flow chart of the steps of a data processing method for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 2 is a flow chart of model training steps of a data processing method for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 3 is a block diagram of one embodiment of a deep learning network model of a data processing method and system for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 4 is a block diagram of another embodiment of a deep learning network model of a data processing method and system for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 5 is a block diagram of another embodiment of a deep learning network model of a data processing method and system for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 6 is a block diagram of a data processing system for diffusion weighted magnetic resonance imaging in accordance with the present invention;

FIG. 7 is a block diagram of a data processing system for diffusion weighted magnetic resonance imaging network model training in accordance with the present invention;

figure 8 is a block diagram of a preferred embodiment of the network model training of a data processing system for diffusion weighted magnetic resonance imaging in accordance with the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.

The invention is further described with reference to the following drawings and specific examples, which are not intended to be limiting.

Referring to fig. 1, the present invention provides a data processing method for diffusion weighted magnetic resonance imaging, comprising the steps of:

step A1, acquiring diffusion weighted magnetic resonance images of each direction by using a diffusion weighted imaging sequence;

step A2, inputting the magnetic resonance image into the trained deep learning network model for operation, and respectively outputting correction parameters and high-resolution images corresponding to each direction;

step A3, for the high-resolution image in each direction, performing rectification by using the rectification parameters in the corresponding direction to obtain the magnetic resonance image with high-resolution diffusion weighting in each direction.

In the human body, water molecules cannot freely move randomly due to the influence of various cell structures, but can only move in a limited environment and range. The movement of water molecules may be more mobile in one direction and more restricted in the other direction. The DWI indirectly reflects the change of the microstructure by detecting the information such as the limited direction and degree of the diffusion movement of water molecules in human tissues.

In particular, frequently used DWI sequences are, for example, GRE sequences, SE sequences, FSE sequences, T1WI, T2WI or T2 wii sequences.

Referring to fig. 2, further, the training process of the deep learning network model includes the following steps:

a step B1 of acquiring a diffusion-weighted magnetic resonance image of each direction with a general resolution as a first image data set, and acquiring a diffusion-weighted magnetic resonance image with a high resolution corresponding to the first image data set as a second image data set, and acquiring a diffusion-free weighted magnetic resonance image corresponding to the first image data set as a third image data set;

step B2, constructing a deep learning network model;

step B3, inputting the first image data set into a deep learning network model, wherein the deep learning network model takes the first characteristic data as a first output and takes the second characteristic data as a second output, the first characteristic data is the predicted correction parameters of each direction, and the second characteristic data is the predicted high-resolution image of each direction;

a step B3 of rectifying the images in the first image dataset based on the first feature data of the respective orientation;

step B4, calculating a first loss value of the rectified first image data set and the rectified third image data set based on the first loss function, and calculating a second loss value of the second feature data and the second image data set based on the second loss function;

and step B5, optimizing parameters of the deep learning network model according to the first loss value and the second loss value until the first loss value and the second loss value reach preset standards, and finishing training the deep learning network model.

Specifically, the high resolution image refers to an image with a resolution greater than a preset resolution, and the general resolution refers to an image with a resolution lower than the preset resolution.

Specifically, using the high-resolution diffusion-weighted magnetic resonance image in each direction obtained in step a3, a high-resolution MD map (Mean dispersion map), an FA map (anisotropic fraction map), and an ADC map (apparent diffusion coefficient map) corresponding thereto are further calculated.

The invention simultaneously predicts the correction parameters, namely the deformation field and the high-resolution diffusion weighted magnetic resonance image by using the artificial neural network (called as a deep learning network model), fully utilizes the complementary information of the diffusion weighted images in different directions, carries out super-resolution calculation while registering, obtains the high-resolution estimation of the image in each direction, improves the registration accuracy, and improves the DWI imaging effect, thereby obtaining clearer MD, FA and ADC images and being beneficial to more accurate disease diagnosis and deeper scientific research. The method can also be used for shortening the imaging time, improving the machine utilization rate, reducing the inspection cost and improving the MRI popularity on the premise of obtaining similar imaging effects.

Specifically, the training data set obtained from the real mr scanner, including the acquisition of the first image data set and the second image data set, may be acquired from human volunteers, or from some experimental animal or non-biological body according to different data of the target application scenario. A general resolution diffusion weighted magnetic resonance image is acquired using a conventional diffusion weighted imaging sequence, and a high resolution diffusion weighted magnetic resonance image is acquired using one or a combination of multiple excitations, readout direction multiple excitations, multiple averages, etc., which are typically time consuming. The general resolution image and the corresponding high resolution image are taken as a set of training data.

In particular, the training data may also be obtained by a simulation algorithm, such as a high resolution diffusion weighted magnetic resonance image simulation process based on the second image dataset resulting in the first image dataset.

Specifically, the weights of the deep learning network model are initialized randomly, and the weights are trained by using prepared training data such as first image data, namely, the weights are optimized. Adam or SGD may be used as the optimizer.

In particular, a production immunity network (GAN) can be additionally used as a part of the second loss function, so that the high-resolution image output by the deep learning network model is more vivid.

Further, step B1 further includes: pre-registering the first image data set by taking a third image data set as a reference to acquire third characteristic data comprising pre-distortion parameters in all directions;

step B3 further includes: and taking the first image data set and the third feature data as input of the deep learning network model.

Referring to fig. 3, further, the deep learning network model is a convolutional neural network model (CNN model), including:

a trunk extraction module, which takes the first image data set as input and comprises a plurality of convolution layers connected in series in sequence;

the first branch module takes the output of the trunk extraction module as input and outputs first characteristic data, and the first characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer;

and the second branch module takes the output of the trunk extraction module as input and outputs second characteristic data, and the second characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer.

Specifically, after a trunk extraction module formed by a plurality of convolution layers connected in series in a one-way mode in sequence, the trunk extraction module is divided into two branches to output, wherein one branch outputs a deformation field, and the other branch outputs a high-resolution diffusion weighted magnetic resonance image.

Specifically, the stem extraction module may also include a residual connection layer.

As a specific embodiment, the skeleton extraction module includes 5 convolutional layers, each convolutional layer has 64 convolutional kernels, the size of the convolutional kernel is 3 × 3, and using LeakyReLu as an activation function.

As a specific embodiment, the second branch module sequentially comprises:

1 convolution layer, including 64 convolution kernels, the size of the convolution kernel is 3 x 3, and LeakyReLu is used as an activation function;

1 2 times of upper sampling layer;

2 convolution kernels comprising 64 convolution kernels, wherein the size of the convolution kernels is 3 multiplied by 3, and LeakyReLu is used as an activation function;

1 2 times of upper sampling layer;

1 convolution layer, comprising 64 convolution kernels, the convolution kernel size being 3 x 3, using a linear activation function.

As a specific embodiment, the first branch module sequentially comprises:

1 convolution layer, including 64 convolution kernels, the size of the convolution kernel is 3 x 3, and LeakyReLu is used as an activation function;

1 2 times of upper sampling layer;

2 convolution kernels comprising 64 convolution kernels, wherein the size of the convolution kernels is 3 multiplied by 3, and LeakyReLu is used as an activation function;

1 2 times of upper sampling layer;

2 convolutional layers, comprising 64 convolutional kernels, with a convolutional kernel size of 3 × 3, using a linear activation function.

As a specific embodiment, for example, the input size of the convolutional neural network model (CNN model) may be 64 × 64 × 16. The first branching module output size is 256 × 256 × 2. The second branching module output size is 256 × 256 × 16.

Referring to fig. 4, further, the deep learning network model is a single Unet neural network model, including:

the encoding module takes a first image data set as input and comprises a plurality of encoding layers which are connected in series, wherein the encoding layers comprise a convolution layer and a down-sampling layer;

the first decoding module takes the output of the coding module as input and outputs first characteristic data, and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer;

and the second decoding module takes the output of the coding module as input and outputs second characteristic data and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer.

Specifically, in a single Unet neural network model, a coding module performs convolution and downsampling on input for multiple times to obtain embedding with a lower dimensionality, the embedding is divided into two branches on the basis, convolution and upsampling are performed for multiple times respectively, one branch outputs a deformation field, and the other branch outputs a high-resolution diffusion weighted graph.

Specifically, each upsampling layer is 2 times, each downsampling layer is 1/2 times, and the last upsampling layer of each of the first decoding module and the second decoding module is 4 times.

Specifically, all convolutional layers use 64 convolutional kernels, have a size of 3 × 3, and use LeakyReLu as an activation function. However, the number of convolution kernels in the last layer should be matched with the number of output channels, and the activation function is linear activation.

As a specific embodiment, the input size of, for example, a single Unet neural network model may be 96 × 96 × 16. The first branching module output size is 192 × 192 × 16. The second branching module output size is 192 × 192 × 16.

Referring to fig. 5, further, the deep learning network model is a dual Unet neural network model, including:

the first Unet module consists of an encoding module and a decoding module, takes a first image data set as the input of the encoding module, and takes second characteristic data as the output of the decoding module;

and the second Unet module consists of an encoding module and a decoding module, the second characteristic data and the first image data set are used as the input of the encoding module, and the first characteristic data is used as the output of the decoding module.

Specifically, a UNet is used to obtain a high-resolution image, and then the high-resolution image is spliced with an original general-resolution image and input into a second UNet to obtain a deformation field.

Further, a high resolution diffusion weighted magnetic resonance image in the second image data set is acquired by using a multi-shot and/or multi-average method.

Further, the first image data set is derived based on a high resolution diffusion weighted magnetic resonance image simulation process in the second image data set.

Further, the first loss function is a mutual information function, and the second loss function is a norm function of L1 or a norm function of L2.

Referring to figure 6, the present invention also provides a data processing system for diffusion weighted magnetic resonance imaging, comprising a data processing method for diffusion weighted magnetic resonance imaging using the foregoing, comprising:

an image acquisition module (1) for acquiring diffusion weighted magnetic resonance images of various directions using a diffusion weighted imaging sequence;

the characteristic extraction module (2) is connected with the image acquisition module (1) and is used for inputting the magnetic resonance image into the trained deep learning network model for operation and respectively obtaining correction parameters and high-resolution images corresponding to all directions;

and the image correction module (3) is connected with the feature extraction module (2) and is used for correcting the high-resolution image in each direction by adopting the correction parameters in the corresponding direction so as to obtain the magnetic resonance image with high-resolution diffusion weighting in each direction. .

Referring to fig. 7, the system further includes training the deep learning network model using:

an image acquisition module (4) for acquiring diffusion-weighted magnetic resonance images of various directions with a general resolution as an image set, acquiring a plurality of image sets as a first image dataset, acquiring a diffusion-weighted magnetic resonance image with a high resolution corresponding to the first image dataset as a second image dataset, and acquiring a diffusion-weightless magnetic resonance image corresponding to the first image dataset as a third image dataset;

the model construction module (5) is used for constructing a deep learning network model;

the characteristic learning module (6) is respectively connected with the image collection module (4) and the model construction module (5) and is used for inputting the first image data set into the deep learning network model, the deep learning network model takes the first characteristic data as first output and takes the second characteristic data as second output, the first characteristic data is predicted correction parameters in all directions, and the second characteristic data is predicted high-resolution images in all directions;

the correction learning module (7) is respectively connected with the feature learning module (6) and the image collection module (4) and is used for correcting the first feature data of the images in the first image data set based on the corresponding directions;

the loss calculation module (8) is respectively connected with the image collection module (4), the correction learning module (7) and the feature learning module (6), calculates first loss values of the corrected first image data set and the corrected third image data set based on a first loss function, and calculates second loss values of the second feature data and the second image data set based on a second loss function;

and the parameter adjusting module (9) is respectively connected with the loss calculating module (8) and the feature learning module (6) to optimize parameters of the deep learning network model according to the first loss value and the second loss value until the first loss value and the second loss value reach preset standards, so that the deep learning network model is trained.

Referring to fig. 8, further, the system further includes:

the pre-registration module (10) is connected with the image collection module (4) and is used for pre-registering the image group on the basis of at least one group of image group in the third image data set to acquire a plurality of groups of third characteristic data comprising pre-distortion parameters in all directions;

the feature learning module (6) is further connected with the pre-registration module (10) and is used for taking the first image data set and the third feature data as input of the deep learning network model.

Referring to fig. 8, further, the deep learning network model is a convolutional neural network model, which includes:

a trunk extraction module, which takes the first image data set as input and comprises a plurality of convolution layers connected in series in sequence;

the first branch module takes the output of the trunk extraction module as input and outputs first characteristic data, and the first characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer;

and the second branch module takes the output of the trunk extraction module as input and outputs second characteristic data, and the second characteristic data comprises a plurality of serially connected convolution layers and an upper sampling layer.

Further, the deep learning network model is a single Unet neural network model, including:

the encoding module takes a first image data set as input and comprises a plurality of encoding layers which are connected in series, wherein the encoding layers comprise a convolution layer and a down-sampling layer;

the first decoding module takes the output of the coding module as input and outputs first characteristic data, and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer;

and the second decoding module takes the output of the coding module as input and outputs second characteristic data and comprises a plurality of decoding layers which are connected in series, wherein each decoding layer consists of a convolutional layer and an upper sampling layer.

Further, the deep learning network model is a dual-Unet neural network model, including:

the first Unet module consists of an encoding module and a decoding module, takes a first image data set as the input of the encoding module, and takes second characteristic data as the output of the decoding module;

and the second Unet module consists of an encoding module and a decoding module, the second characteristic data is used as the input of the encoding module, and the first characteristic data is used as the output of the decoding module.

Further, a high resolution diffusion weighted magnetic resonance image in the second image data set is acquired by using a multi-shot and/or multi-average method.

Further, the first image data set is derived based on a high resolution diffusion weighted magnetic resonance image simulation process in the second image data set.

Further, the first loss function is a mutual information function, and the second loss function is a norm function of L1 or a norm function of L2.

While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention.