Non-reference-based magnetic resonance image quality evaluation method and device and readable medium

Abstract

The invention discloses a non-reference-based magnetic resonance image quality evaluation method and device and a readable medium, and relates to an image processing module, and the method comprises the steps: constructing a non-reference-based magnetic resonance image quality evaluation model, and training the model; a to-be-evaluated magnetic resonance image is acquired and input into a trained magnetic resonance image quality evaluation model, and the magnetic resonance image is firstly input into an image estimation module and sequentially passes through a plurality of first convolutional layers to obtain a plurality of corresponding preliminary estimation images; inputting a preliminary estimation image output by the first convolutional layer into a multi-scale feature extraction module to obtain a multi-scale feature image; inputting the magnetic resonance image, the plurality of preliminary estimation images and the multi-scale feature image into a magnetic resonance curve estimation module to obtain a magnetic resonance curve estimation image; and inputting the magnetic resonance curve estimation image, the multi-scale feature image and the magnetic resonance image into a magnetic resonance multivariate Gaussian blind quality module to obtain a quality score. The invention solves the problem of lack of reference magnetic resonance images.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a non-reference based magnetic resonance image quality assessment method, apparatus and readable medium. Background Art

[0002] Due to the advantages of clearly showing soft tissue structures, no ionizing radiation hazards, and being able to perform tomographic examinations in any direction, magnetic resonance imaging (MRI) has been widely used in clinical diagnosis and computer-aided diagnosis. However, the imaging quality, acquisition process, etc. of MRI examination equipment may cause distortion, affecting the accuracy of subsequent image analysis (such as segmentation, detection, etc.). Therefore, image quality assessment is an important benchmark for comparing the quality of magnetic resonance (MR) images before image analysis and after image enhancement.

[0003] Currently, image quality assessment methods can be divided into full-reference image quality assessment, semi-reference image quality assessment, and non-reference image quality assessment. Full-reference image quality assessment and semi-reference image quality assessment require paired or unpaired high-quality and low-quality images for training. However, obtaining paired or unpaired magnetic resonance images is challenging, and it is difficult to conduct large-scale subjective evaluations to obtain the mean opinion score through human labels. Therefore, due to the lack of reference images to be compared, non-reference image quality assessment becomes important. Traditional non-reference image quality assessment methods and non-reference image quality assessment networks with shallow architectures mainly rely on low-level features and statistical features of images to evaluate image quality, and cannot deeply understand the semantic information and structure of images, and are easily affected by factors such as noise and artifacts. This is because low-level features are not sufficient to accurately represent the complex distortions encountered in real-world scenes. Non-reference image quality assessment methods based on deep learning mainly rely on deep features and the learning ability of neural network models to evaluate image quality, and can learn high-level features of images, such as texture, structure, semantic information, etc., through training data, and can more accurately simulate human visual perception.

[0004] In one of the methods proposed in the literature, a concept based on signal and noise is proposed as an objective metric for measuring the difference between two images. It is carried out by converting the evaluation of image quality into the ratio of the signal (the original image) to the noise (the distorted part). However, the global image quality evaluation method cannot accurately analyze the detailed features of the image and over-analyzes the noise in the background, thus affecting the result of image quality evaluation. Another method proposed in the literature takes one of the two images as the undistorted image and the other as the distorted image. The structural similarity between the two can be regarded as a metric for measuring the image quality of the distorted image. The structural similarity is more in line with the human eye's judgment of image quality in the measurement of image quality. However, such methods require a reference image for evaluation, and it is difficult to obtain the corresponding reference image, and the quality of the reference image will affect the training and evaluation effect of the model. Summary of the Invention

[0005] The purpose of this application is to propose a non-reference-based magnetic resonance image quality evaluation method, device, and readable medium for the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a non-reference-based magnetic resonance image quality evaluation method, including the following steps:

[0007] Construct a non-reference-based magnetic resonance image quality evaluation model and train it to obtain a trained magnetic resonance image quality evaluation model. The magnetic resonance image quality evaluation model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multi-variate Gaussian blind quality module; the image estimation module includes a number of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimation image;

[0008] Obtain the magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality evaluation model. The magnetic resonance image is first input into the image estimation module and passes through a number of first convolutional layers in sequence to obtain a number of corresponding preliminary estimation images; the preliminary estimation image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image; the magnetic resonance image, a number of preliminary estimation images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimation image; the magnetic resonance curve estimation image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multi-variate Gaussian blind quality module to obtain a corresponding quality score.

[0009] Preferably, in the magnetic resonance curve estimation module, local dynamic estimation is performed on the preliminary estimation image output by each first convolutional layer to obtain the corresponding magnetic resonance curve feature, as shown in the following formula:

[0010] ;

[0011] Among them, represents the magnetic resonance curve feature corresponding to the preliminary estimated image output by the first first convolutional layer, represents the magnetic resonance image; represents the trainable curve parameter of the first recursive operation, represents the preliminary estimated image output by the first first convolutional layer, represents the th pixel point;

[0012] The magnetic resonance curve feature corresponding to the preliminary estimated image output by the first first convolutional layer is weighted and fused with the multi-scale feature image to obtain the advanced estimated image of the first recursive operation, as shown in the following formula:

[0013] ;

[0014] Among them, represents the advanced estimated image of the first recursive operation, and represent the weight coefficients, represents the multi-scale feature image;

[0015] The magnetic resonance curve feature corresponding to the preliminary estimated image output by the i-th first convolutional layer is obtained by performing a recursive operation on the advanced estimated image of the (i - 1)-th recursive operation and the preliminary estimated image output by the i-th first convolutional layer, as shown in the following formula:

[0016] ;

[0017] Among them, represents the magnetic resonance curve feature corresponding to the preliminary estimated image output by the i-th first convolutional layer, represents the preliminary estimated image output by the i-th first convolutional layer, represents the trainable curve parameter of the i-th recursive operation, i = 2, 3, …, n, where n represents the total number of recursive operations, which is the same as the total number of the first convolutional layers, represents the advanced estimated image of the (i - 1)-th recursive operation, and its expression is as follows:

[0018] ;

[0019] After n recursive operations, the advanced estimated image of the n-th recursive operation is obtained. The advanced estimated image of the n-th recursive operation first passes through a ReLU activation function, then passes through the second convolutional layer and another ReLU activation function to obtain the magnetic resonance curve estimated image E.

[0020] Preferably, the multi-scale feature extraction module includes a plurality of multi-scale feature extraction layers. Each multi-scale feature extraction layer includes a third convolutional layer, an average pooling module, a max pooling module, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a feature fusion layer, and a seventh convolutional layer. The average pooling module includes an average pooling layer and a fully connected layer, and the max pooling module includes a max pooling layer and a fully connected layer. The ReLU activation function is used in all the fully connected layers. The input features of the current multi-scale feature extraction layer are input into the third convolutional layer and then pass through the average pooling layer and the max pooling layer respectively to obtain the average pooling features and the max pooling features. The average pooling features and the max pooling features are respectively input into the corresponding fully connected layers and then fused to obtain the first fusion feature, as shown in the following formula:

[0021] ;

[0022] where, represents the output features of the third convolutional layer, F avg and F max represent the average pooling and max pooling operations respectively, α0 and α1 represent the parameters of the fully connected layer, σ represents the sigmoid activation function, ReLU represents the ReLU activation function, represents the first fusion feature;

[0023] The first fusion feature is weighted and fused with the input features of the current multi-scale feature extraction layer and then respectively input into the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer of different scales to obtain the first-scale feature, the second-scale feature, and the third-scale feature. The first-scale feature, the second-scale feature, and the third-scale feature are input into the feature fusion layer to obtain the second fusion feature, as shown in the following formula:

[0024] ;

[0025] where, Z represents the second fusion feature, F j represents the j-th scale feature, j = 1, 2,.., m, m represents the total number of scales, Concat represents the feature fusion operation, w j represents the weight of the j-th scale feature, and its expression is as follows:

[0026] ;

[0027] where, [] represents the concatenation operation, represents the Softmax activation function;

[0028] After the second fusion feature passes through the Softmax activation function, it is then residually connected to the input feature of the current multi-scale feature extraction layer to obtain the third fusion feature. The third fusion feature passes through the seventh convolutional layer and is then weighted and fused with the second fusion feature to obtain the output feature of the current multi-scale feature extraction layer. The output feature of the current multi-scale feature extraction layer is then input into the next multi-scale feature extraction layer. The output feature of the last multi-scale feature extraction layer is the multi-scale feature image f b 。

[0029] Preferably, the number of multi-scale feature extraction layers in the multi-scale feature extraction module is 6. The input feature of the first multi-scale feature extraction layer is the feature obtained by passing the preliminary estimated image output by the first first convolutional layer through the ReLU activation function. The input features of the 2nd - 5th multi-scale feature extraction layers are the features obtained by concatenating the output feature of the previous multi-scale feature extraction layer and the preliminary estimated image output by the first first convolutional layer and then passing through the ReLU activation function. The input feature of the 6th multi-scale feature extraction layer is the feature obtained by concatenating the output feature of the previous multi-scale feature extraction layer and the preliminary estimated image output by the first first convolutional layer and then passing through the Tanh activation function. The convolutional kernel sizes of the third convolutional layer and the fourth convolutional layer are 3×3, and the convolutional kernel sizes of the second convolutional layer, the fifth convolutional layer, and the seventh convolutional layer are 5×5. The convolutional kernel size of the sixth convolutional layer is 7×7.

[0030] Preferably, in the magnetic resonance multi-Gaussian blind quality module, the magnetic resonance curve estimation image and the multi-scale feature image are first weighted and fused to obtain the final fusion feature, as shown in the following formula:

[0031] ;

[0032] where, f b represents the multi-scale feature image, E represents the magnetic resonance curve estimation image, represents the final fusion feature, and t1 and t2 represent the fusion parameters;

[0033] The Gaussian distributions of the mean, standard deviation, and deviation between the magnetic resonance image and its corresponding magnetic resonance curve estimation image are calculated using the following formulas respectively, as shown below:

[0034] ;

[0035] ;

[0036] ;

[0037] where, p m 、p d and p sRepresents the Gaussian distribution of the mean, standard deviation, and skewness difference between a magnetic resonance image and its corresponding estimated magnetic resonance curve image, σ m 、σ d and σ s respectively represent the average of the averages of the gray values calculated for several local regions in the magnetic resonance image, the average of the standard deviations of the gray values calculated for several regions, and the average of the skewnesses of the gray values calculated for several regions. μ m 、μ d and μ s respectively represent the standard deviation of the averages of the gray values calculated for several regions in the magnetic resonance image, the standard deviation of the standard deviations of the gray values calculated for several regions, and the standard deviation of the skewnesses of the gray values calculated for several regions;

[0038] By using the 3-sigma rule to calculate the normal distribution and analyzing the mean, standard deviation, and skewness between the magnetic resonance image and its corresponding estimated magnetic resonance curve image, the quality score of the magnetic resonance image is obtained as shown in the following formula:

[0039] ;

[0040] where, represents the quality score of the magnetic resonance image, and T represents the transpose.

[0041] Preferably, the convolution kernel size of the first convolution layer is 3×3, and the number of the first convolution layers is 7. The output features of the first 3 first convolution layers are the corresponding preliminary estimated images. The output feature of the 3rd first convolution layer is skip-connected with the output feature of the 4th first convolution layer to obtain the preliminary estimated image output by the 4th first convolution layer; the output feature of the 2nd first convolution layer is skip-connected with the output feature of the 5th first convolution layer to obtain the preliminary estimated image output by the 5th first convolution layer; the output feature of the 1st first convolution layer is skip-connected with the output feature of the 6th first convolution layer to obtain the preliminary estimated image output by the 6th first convolution layer. The preliminary estimated images output by the first 5 first convolution layers need to pass through the ReLU activation function before being input to the next first convolution layer, and the preliminary estimated image output by the 6th first convolution layer needs to pass through the Tanh activation function before being input to the 7th first convolution layer.

[0042] In a second aspect, the present invention provides a non-reference-based magnetic resonance image quality assessment device, including:

[0043] A model construction module, configured to construct and train a non-reference-based magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multivariate Gaussian blind quality module. The image estimation module includes a number of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimation image.

[0044] An evaluation module, configured to obtain a magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality assessment model. The magnetic resonance image is first input into the image estimation module and sequentially passes through a number of first convolutional layers to obtain a number of corresponding preliminary estimation images. The preliminary estimation image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image. The magnetic resonance image, a number of preliminary estimation images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimation image. The magnetic resonance curve estimation image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multivariate Gaussian blind quality module to obtain a corresponding quality score.

[0045] In a third aspect, the present invention provides an electronic device, including one or more processors; a storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the first aspect.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the method described in any implementation manner of the first aspect.

[0047] In a fifth aspect, the present invention provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the method described in any implementation manner of the first aspect.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The non-reference-based magnetic resonance image quality assessment method proposed by the present invention proposes an image estimation module and a magnetic resonance curve estimation module, and estimates the original brightness and detail features through non-linear curve mapping to solve the problem of lack of reference magnetic resonance images.

[0050] (2) The non-reference-based magnetic resonance image quality assessment method proposed by the present invention proposes a multi-scale feature extraction module, and estimates lesion and tissue features by decomposing magnetic resonance images into multi-scale features to fully extract context feature information and solve the problems of variability of tissue features and randomness of lesion regions in magnetic resonance images.

[0051] (3) The non-reference based magnetic resonance image quality assessment method proposed by the present invention proposes a magnetic resonance multi-Gaussian blind quality module, which compares and evaluates various features between the magnetic resonance image and the magnetic resonance curve estimation image to obtain the quality score of the magnetic resonance image. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0053] Figure 1 It is a schematic flowchart of the non-reference based magnetic resonance image quality assessment method according to the embodiment of the present application;

[0054] Figure 2 It is a schematic diagram of the magnetic resonance image quality assessment model of the non-reference based magnetic resonance image quality assessment method according to the embodiment of the present application;

[0055] Figure 3 It is a schematic diagram of the non-reference based magnetic resonance image quality assessment device according to the embodiment of the present application;

[0056] Figure 4 It is a schematic diagram of the hardware structure of the electronic device provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0058] Figure 1 A non-reference based magnetic resonance image quality assessment method provided by the embodiment of the present application is shown, including the following steps:

[0059] S1, constructing and training a non-reference based magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multi-Gaussian blind quality module; the image estimation module includes a plurality of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimation image.

[0060] Specifically, refer toFigure 2 , first, magnetic resonance images are acquired in a magnetic resonance imaging system to form a magnetic resonance image set; the magnetic resonance image set is divided in a certain proportion to obtain a training set, a validation set, and a test set. A non-reference based magnetic resonance image quality assessment model is constructed. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multi-variate Gaussian blind quality module. The magnetic resonance image set is used to train the magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. During the training process of the magnetic resonance image quality assessment model, the calculated corresponding magnetic resonance curve features are used as the loss function. By minimizing the magnetic resonance curve features, the optimal trainable parameters are obtained and saved. Subsequently, the trained magnetic resonance image quality assessment model can be used to analyze parts such as the brightness, lesion features, and edge features of the magnetic resonance image to obtain the corresponding quality score of the magnetic resonance image.

[0061] S2. Obtain the magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality assessment model. The magnetic resonance image is first input into the image estimation module and passes through a number of first convolutional layers in sequence to obtain a number of corresponding preliminary estimation images; the preliminary estimation image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image; the magnetic resonance image, the number of preliminary estimation images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimation image; the magnetic resonance curve estimation image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multi-variate Gaussian blind quality module to obtain the corresponding quality score.

[0062] In a specific embodiment, the convolution kernel size of the first convolutional layer is 3×3, and the number of the first convolutional layers is 7. Among them, the output features of the first 3 first convolutional layers are the corresponding preliminary estimation images. The output feature of the 3rd first convolutional layer is skip-connected with the output feature of the 4th first convolutional layer to obtain the preliminary estimation image output by the 4th first convolutional layer; the output feature of the 2nd first convolutional layer is skip-connected with the output feature of the 5th first convolutional layer to obtain the preliminary estimation image output by the 5th first convolutional layer; the output feature of the 1st first convolutional layer is skip-connected with the output feature of the 6th first convolutional layer to obtain the preliminary estimation image output by the 6th first convolutional layer. The preliminary estimation images output by the first 5 first convolutional layers need to pass through the ReLU activation function before being input into the next first convolutional layer, and the preliminary estimation image output by the 6th first convolutional layer needs to pass through the Tanh activation function before being input into the 7th first convolutional layer.

[0063] Specifically, the image estimation module and the magnetic resonance curve estimation module in the magnetic resonance image quality assessment model proposed in the embodiments of the present application mainly design curves based on the magnetic resonance imaging principle to estimate the original brightness and detail features of magnetic resonance images. According to the general imaging formula of MRI, it aims to dynamically adjust the gray range of the input magnetic resonance image through the curve, estimate the gray values of the magnetic resonance image at the optimal contrast and brightness, rather than mapping from image to image.

[0064] Reference Figure 2 , the magnetic resonance image quality assessment model learns the mapping between the magnetic resonance image and its best estimated magnetic resonance curve through an image estimation module including 7 first convolutional layers. Each first convolutional layer is composed of a convolutional layer with a convolutional kernel size of 3×3. The adjacent two first convolutional layers are connected in sequence, and there are skip connections. There is no upsampling and downsampling. After the input magnetic resonance image is input into the magnetic resonance image quality assessment model, after each first convolutional layer, a corresponding preliminary estimated image will be obtained, thus obtaining a set of preliminary estimated images. As the number of network layers increases, more features can be learned at different scales to gradually adapt to the true features of the magnetic resonance image.

[0065] In a specific embodiment, the multi-scale feature extraction module includes several multi-scale feature extraction layers. The multi-scale feature extraction layer includes a third convolutional layer, an average pooling module, a max pooling module, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a feature fusion layer, and a seventh convolutional layer. The average pooling module includes an average pooling layer and a fully connected layer. The max pooling module includes a max pooling layer and a fully connected layer. The ReLU activation function is used in all fully connected layers; the input features of the current multi-scale feature extraction layer are input into the third convolutional layer and then pass through the average pooling layer and the max pooling layer respectively to obtain the average pooling feature and the max pooling feature. The average pooling feature and the max pooling feature are respectively input into the corresponding fully connected layers and then fused to obtain the first fusion feature, as shown in the following formula:

[0066] ;

[0067] Where represents the output features of the third convolutional layer, F avg and F max respectively represent the average pooling and max pooling operations, α0 and α1 represent the parameters of the fully connected layers, σ represents the sigmoid activation function, ReLU represents the ReLU activation function, represents the first fusion feature;

[0068] The first fusion feature is weighted and fused with the input feature of the current multi-scale feature extraction layer, and then respectively input into the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer at different scales to obtain the first-scale feature, the second-scale feature, and the third-scale feature. The first-scale feature, the second-scale feature, and the third-scale feature are input into the feature fusion layer to obtain the second fusion feature, as shown in the following formula:

[0069] ;

[0070] where Z represents the second fusion feature, F j represents the j-th scale feature, j = 1, 2,.., m, m represents the total number of scales, Concat represents the feature fusion operation, w j represents the weight of the j-th scale feature, and its expression is as follows:

[0071] ;

[0072] where [] represents the concatenation operation, represents the Softmax activation function;

[0073] The second fusion feature passes through the Softmax activation function and then performs a residual connection with the input feature of the current multi-scale feature extraction layer to obtain the third fusion feature. The third fusion feature passes through the seventh convolutional layer and then is weighted and fused with the second fusion feature to obtain the output feature of the current multi-scale feature extraction layer. The output feature of the current multi-scale feature extraction layer is then input into the next multi-scale feature extraction layer, and the output feature of the last multi-scale feature extraction layer is the multi-scale feature image f b .

[0074] In a specific embodiment, the number of multi-scale feature extraction layers in the multi-scale feature extraction module is 6. The input feature of the first multi-scale feature extraction layer is the feature obtained by passing the preliminary estimated image output by the first first convolutional layer through the ReLU activation function. The input features of the 2nd - 5th multi-scale feature extraction layers are the features obtained by concatenating the output feature of the previous multi-scale feature extraction layer with the preliminary estimated image output by the first first convolutional layer and then passing through the ReLU activation function. The input feature of the 6th multi-scale feature extraction layer is the feature obtained by concatenating the output feature of the previous multi-scale feature extraction layer with the preliminary estimated image output by the first first convolutional layer and then passing through the Tanh activation function. The convolutional kernel sizes of the third convolutional layer and the fourth convolutional layer are 3×3, the convolutional kernel sizes of the second convolutional layer, the fifth convolutional layer, and the seventh convolutional layer are 5×5, and the convolutional kernel size of the sixth convolutional layer is 7×7.

[0075] Specifically, in clinical diagnosis and image analysis, local features in magnetic resonance images, such as tumors, tissue edges, and detailed features, are mainly concerned. However, the variability of tissue features and the randomness of damaged areas make it difficult for deep learning networks to learn their features. Therefore, the embodiments of the present application propose a multi-scale feature extraction module. The preliminary estimated image y1 output by the first first convolutional layer is input. After passing through 6 multi-scale feature extraction layers, a multi-scale feature image is obtained. Specifically, after the input features of the current multi-scale feature extraction layer are input into the current multi-scale feature extraction layer, they first pass through a third convolutional layer with a convolutional kernel size of 3×3, an average pooling module, and a max pooling module to obtain a first fused feature. The first fused feature is fused with the estimated image y1, and convolutional operations with convolutional kernel sizes of 3×3, 5×5, and 7×7 are respectively performed to obtain features of different scales. Through feature fusion operations, a second fused feature is obtained. After passing through the Softmax activation function, a residual operation is performed with the input features of the current multi-scale feature extraction layer to obtain a third fused feature. After passing through a seventh convolutional layer with a convolutional kernel size of 5×5, a weighted fusion operation is performed with the third fused feature to obtain the output features of the current multi-scale feature extraction layer; the input features of the first multi-scale feature extraction layer are the first preliminary estimated image y1. The input features of the 2nd - 5th multi-scale feature extraction layers are the features obtained after skip connection of the output features of the previous multi-scale feature extraction layer and the first preliminary estimated image y1 output by the first first convolutional layer and then passing through the ReLU activation function. The input features of the 6th multi-scale feature extraction layer are the features obtained after skip connection of the output features of the previous multi-scale feature extraction layer and the first preliminary estimated image y1 output by the first first convolutional layer and then passing through the Tanh activation function.

[0076] In a specific embodiment, in the magnetic resonance curve estimation module, local dynamic estimation is performed on the preliminary estimated image output by each first convolutional layer to obtain the corresponding magnetic resonance curve features, as shown in the following formula:

[0077] ;

[0078] Wherein, represents the magnetic resonance curve features corresponding to the preliminary estimated image output by the first first convolutional layer, represents the magnetic resonance image; represents the trainable curve parameters of the first recursive operation, represents the preliminary estimated image output by the first first convolutional layer, represents the th pixel point;

[0079] The magnetic resonance curve features corresponding to the preliminary estimated image output by the first first convolutional layer are weighted and fused with the multi-scale feature image to obtain the advanced estimated image of the first recursive operation, as shown in the following formula:

[0080] ;

[0081] Among them, represents the advanced estimated image of the first recursive operation, and represent the weight coefficients, represents the multi-scale feature image;

[0082] The magnetic resonance curve features corresponding to the preliminary estimated image output by the i-th first convolutional layer are obtained through a recursive operation of the advanced estimated image of the (i - 1)-th recursive operation and the preliminary estimated image output by the i-th first convolutional layer, as shown in the following formula:

[0083] ;

[0084] Among them, represents the magnetic resonance curve features corresponding to the preliminary estimated image output by the i-th first convolutional layer, represents the preliminary estimated image output by the i-th first convolutional layer, represents the trainable curve parameters of the i-th recursive operation, i = 2, 3,..., n, where n represents the total number of recursive operations, which is the same as the total number of first convolutional layers, represents the advanced estimated image of the (i - 1)-th recursive operation, and its expression is as follows:

[0085] ;

[0086] After n recursive operations, the advanced estimated image of the n-th recursive operation is obtained. The advanced estimated image of the n-th recursive operation first passes through a ReLU activation function, then through a second convolutional layer and another ReLU activation function to obtain the magnetic resonance curve estimated image E.

[0087] Specifically, in order to adapt to the feature information of each layer, each preliminary estimated image in the preliminary estimated image set needs to be input into the magnetic resonance curve estimation module, which can not only estimate the magnetic resonance curve features, but also iteratively use the magnetic resonance curve features for local dynamic estimation. In the magnetic resonance curve estimation module, the multi-scale feature images extracted by the multi-scale feature extraction module need to be recursively operated on the preliminary estimated image set. After the input magnetic resonance image is input into the magnetic resonance image quality assessment model, through the convolution operation of the first convolutional layer with a convolutional kernel size of 3×3 for n times, a preliminary estimated image is obtained. The number of iterations is n, which is the same as the number of the first convolutional layers. Each first convolutional layer of the magnetic resonance image has a corresponding magnetic resonance curve to estimate the best fitting features. During the iteration process, ∈[0,1], and its initial value can be calculated as:

[0088] ;

[0089] where, I max and I min represent the maximum and minimum values of the gray values in the magnetic resonance image. By setting different parameters β, the low-frequency and high-frequency features of magnetic resonance images with different gray levels can be estimated, and the magnetic resonance curve features of each gray level can be accurately estimated. Through iterative estimation, the parameter gradually approaches the distribution of the image features, making the estimated magnetic resonance curve features closer to the real features without distortion. After n recursive operations, an advanced estimated image for each recursive operation is obtained, and the magnetic resonance curve estimation image E is predicted through the advanced estimated image of the nth recursive operation.

[0090] In a specific embodiment, in the magnetic resonance multi-Gaussian blind quality module, the magnetic resonance curve estimation image and the multi-scale feature image are first weighted and fused to obtain the final fusion feature, as shown in the following formula:

[0091] ;

[0092] where, f b represents the multi-scale feature image, E represents the magnetic resonance curve estimation image, represents the final fusion feature, and t1 and t2 represent the fusion parameters;

[0093] The Gaussian distributions of the mean, standard deviation, and deviation between the magnetic resonance image and its corresponding magnetic resonance curve estimation image are calculated respectively using the following formulas, as shown in the following formulas:

[0094] ;

[0095] ;

[0096] ;

[0097] Among them, p m , p d and p s represent the Gaussian distribution of the mean, standard deviation, and skewness difference between the magnetic resonance image and its corresponding estimated magnetic resonance curve image. σ m , σ d and σ s respectively represent the mean of the means of the gray values calculated for several local parts in the magnetic resonance image, the mean of the standard deviations of the gray values calculated for several regions, and the mean of the skewnesses of the gray values calculated for several regions. μ m , μ d and μ s respectively represent the standard deviation of the means of the gray values calculated for several regions in the magnetic resonance image, the standard deviation of the standard deviations of the gray values calculated for several regions, and the standard deviation of the skewnesses of the gray values calculated for several regions;

[0098] By using the 3-sigma rule to calculate the normal distribution and analyzing the mean, standard deviation, and skewness between the magnetic resonance image and its corresponding estimated magnetic resonance curve image, the quality score of the magnetic resonance image is obtained as shown in the following formula:

[0099] ;

[0100] Among them, represents the quality score of the magnetic resonance image, and T represents the transpose.

[0101] Specifically, different from natural images, magnetic resonance images mainly use different gray values to represent different tissue types or different water contents. The node significance score (NSS) based on the image intensity distribution cannot directly capture the characteristics of the structural details in the magnetic resonance image, and the detailed characteristics contained in the magnetic resonance images of different organs are also different. Therefore, after inputting the estimated magnetic resonance curve image and the multi-scale feature image, by combining the gray histogram distribution of the image, a fitting curve conforming to the magnetic resonance image distribution is constructed.

[0102] Then, since the distributions of different features in the magnetic resonance image follow the normal distribution, by using the 3-sigma rule to calculate the normal distribution and analyzing the mean, standard deviation, and skewness between the magnetic resonance image and its corresponding estimated magnetic resonance curve image, the quality score of the magnetic resonance image is obtained.

[0103] The following uses specific experiments to illustrate the technical effects of the non-reference-based magnetic resonance image quality assessment method proposed in the embodiments of the present application.

[0104] The experimental parameter settings for the embodiments of this application are as follows: The batch size is set to 16, the model is trained for 200 epochs, and the initial learning rate is 0.005. All experiments use PyTorch version 1.3.1 and are conducted on two NVIDIA RTX 3090 Ti GPUs, each with 24 GB of memory.

[0105] To further verify the effectiveness of each module of the non-reference based magnetic resonance image quality model proposed in the embodiments of this application, and considering that the image estimation module, magnetic resonance curve estimation module, and magnetic resonance scene statistical blind quality model are indispensable components, a step-by-step performance evaluation was conducted for the multi-scale feature extraction module. The results of the ablation study are shown in Table 1. According to the ablation results of the non-reference based magnetic resonance image quality model proposed in the embodiments of this application without the multi-scale feature extraction module, the average values of PLCC and SROCC are 0.756 and 0.791 respectively. When the multi-scale feature extraction module is added to the magnetic resonance image quality model, the mean values of PLCC and SROCC increase by 12.1% respectively. This indicates that the multi-scale feature extraction module can effectively extract context information and detail features from magnetic resonance images, and can more accurately estimate background and detail features when fused with the estimated magnetic resonance images.

[0106] Table 1 Results of the ablation experiment:

[0107]

[0108] To evaluate the image quality assessment performance of the non-reference based magnetic resonance image quality assessment method proposed in the embodiments of the present application, four advanced RR-IQA methods were adopted in the present application for comparative experiments, including the Blind Referenceless Image Spatial Quality Evaluator (BRISQUE), the Wavelet Domain Image Quality Assessment Model (WaDIOaM), the Non-Reference Quality Metric (NROM), and the Neural Image Assessment (NIMA) model. In addition, three advanced NR-IQA methods were also adopted: Discrete Entropy (DE), Gray Mean Gradient (GMG), and Integrated Local Natural Image Quality Evaluator (IL-NIQE) for comparison.

[0109] According to the PLCC and SROCC evaluation metrics of each image quality assessment method listed in Table 2, the values obtained by the non-reference based magnetic resonance image quality assessment method proposed in the embodiments of the present application are all optimal. Specifically, the PLCC and SROCC values are 0.848 and 0.864 respectively, which indicates that the non-reference based magnetic resonance image quality assessment method proposed in the embodiments of the present application has excellent performance compared to other advanced image quality assessment methods.

[0110] Table 2 Comparative experiment results:

[0111]

[0112] Further referring to Figure 3 As an implementation of the methods shown in the above figures, the present application provides an embodiment of a non-reference based magnetic resonance image quality assessment device, and this device embodiment corresponds to the Figure 1 method embodiment shown, and this device can be specifically applied to various electronic devices.

[0113] The embodiments of the present application provide a non-reference based magnetic resonance image quality assessment device, including:

[0114] The model construction module 1 is configured to construct and train a non-reference-based magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multivariate Gaussian blind quality module. The image estimation module includes a plurality of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimation image.

[0115] The evaluation module 2 is configured to obtain a magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality assessment model. The magnetic resonance image is first input into the image estimation module and sequentially passes through a plurality of first convolutional layers to obtain a plurality of corresponding preliminary estimation images. The preliminary estimation image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image. The magnetic resonance image, the plurality of preliminary estimation images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimation image. The magnetic resonance curve estimation image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multivariate Gaussian blind quality module to obtain a corresponding quality score.

[0116] Figure 4 It is a schematic hardware structure diagram of the electronic device provided by the embodiment of the present invention. As Figure 4 shown, the electronic device of this embodiment includes: a processor 401 and a memory 402. Among them, the memory 402 is used to store computer execution instructions. The processor 401 is used to execute the computer execution instructions stored in the memory to implement each step executed by the electronic device in the above embodiment. Specifically, reference can be made to the relevant descriptions in the foregoing method embodiments.

[0117] Optionally, the memory 402 can be either independent or integrated with the processor 401.

[0118] When the memory 402 is independently provided, the electronic device further includes a bus 403 for connecting the memory 402 and the processor 401.

[0119] The embodiment of the present invention also provides a computer storage medium. Computer execution instructions are stored in the computer storage medium. When the processor 401 executes the computer execution instructions, the above method is implemented.

[0120] The embodiment of the present invention also provides a computer program product, including a computer program. When the computer program is executed by the processor 401, the above method is implemented.

[0121] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections between each other can be indirect couplings or communication connections through some interfaces, devices or modules, and can be in electrical, mechanical or other forms.

[0122] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to implement the solution of this embodiment.

[0123] In addition, each functional module in various embodiments of the present invention can be integrated in a processing unit, or each module can exist physically alone, or two or more modules can be integrated in one unit. The units formed by the above modules can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0124] The above integrated modules implemented in the form of software functional modules can be stored in a computer-readable storage medium. The above software functional modules are stored in a storage medium, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor 401 to execute some steps of the methods in various embodiments of the present application.

[0125] It should be understood that the above processor 401 can be a central processing unit (Central Processing Unit, abbreviated as CPU), and can also be other general-purpose processors, digital signal processors (Digital Signal Processor, abbreviated as DSP), application-specific integrated circuits (Application Specific Integrated Circuit, abbreviated as ASIC), etc. The general-purpose processor can be a microprocessor or the processor 401 can also be any conventional processor 401, etc. The steps of the method disclosed in combination with the invention can be directly embodied as being executed by the hardware processor 401, or executed by a combination of hardware and software modules in the processor 401.

[0126] The memory 402 may include high-speed RAM memory and may also include non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a portable hard drive, a read-only memory, a magnetic disk, or an optical disc, etc.

[0127] The bus 403 may be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus 403 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience in representation, the bus 403 in the drawings of this application is not limited to only one bus 403 or one type of bus 403.

[0128] The above-mentioned storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc. The storage medium may be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0129] An exemplary storage medium is coupled to the processor 401, enabling the processor 401 to read information from the storage medium and write information to the storage medium. Of course, the storage medium may also be a component of the processor 401. The processor 401 and the storage medium may be located in an Application Specific Integrated Circuits (ASIC). Of course, the processor 401 and the storage medium may also exist as discrete components in an electronic device or a master control device.

[0130] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including the above method embodiments; and the foregoing storage medium includes: ROM, RAM, magnetic disks, or optical discs and other media that can store program codes.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A non-reference based magnetic resonance image quality assessment method, characterized in that, The method includes the following steps: Construct and train a non-reference based magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multivariate Gaussian blind quality module. The image estimation module includes a number of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimated image. Obtain a magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality assessment model. The magnetic resonance image is first input into the image estimation module and sequentially passes through a number of first convolutional layers to obtain a number of corresponding preliminary estimated images. The preliminary estimated image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image. The magnetic resonance image, the number of preliminary estimated images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimated image. The magnetic resonance curve estimated image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multivariate Gaussian blind quality module to obtain a corresponding quality score.

2. The non-reference based magnetic resonance image quality assessment method according to claim 1, wherein In the magnetic resonance curve estimation module, local dynamic estimation is performed on the preliminary estimated image output by each first convolutional layer to obtain a corresponding magnetic resonance curve feature, as shown in the following formula: ； Among them, represents the magnetic resonance curve feature corresponding to the preliminary estimated image output by the first first convolutional layer, represents the magnetic resonance image; represents the trainable curve parameter of the first recursive operation, represents the preliminary estimated image output by the first first convolutional layer, represents the th pixel point; The magnetic resonance curve feature corresponding to the preliminary estimated image output by the first first convolutional layer is weighted and fused with the multi-scale feature image to obtain an advanced estimated image for the first recursive operation, as shown in the following formula: ； Among them, represents the advanced estimated image of the first recursive operation, and represents the weight coefficient, represents the multi-scale feature image; The advanced estimated image from the (i - 1)-th recursive operation is recursively operated with the preliminary estimated image output by the i-th first convolutional layer to obtain a magnetic resonance curve feature corresponding to the preliminary estimated image output by the i-th first convolutional layer, as shown in the following formula: ； Among them, represents the magnetic resonance curve feature corresponding to the preliminary estimated image output by the i-th first convolutional layer, represents the preliminary estimated image output by the i-th first convolutional layer, represents the trainable curve parameter for the i-th recursive operation, where i = 2, 3, …, n, and n represents the total number of recursive operations, which is the same as the total number of the first convolutional layers, represents the advanced estimated image for the (i - 1)-th recursive operation, and its expression is as follows: ； After n recursive operations, an advanced estimated image for the n-th recursive operation is obtained. The advanced estimated image for the n-th recursive operation first passes through a ReLU activation function, then passes through a second convolutional layer and another ReLU activation function to obtain a magnetic resonance curve estimated image E.

3. The method for evaluating the quality of a magnetic resonance image based on non-reference according to claim 2, wherein The multi-scale feature extraction module includes a number of multi-scale feature extraction layers. The multi-scale feature extraction layer includes a third convolutional layer, an average pooling module, a max pooling module, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer, a feature fusion layer, and a seventh convolutional layer. The average pooling module includes an average pooling layer and a fully connected layer. The max pooling module includes a max pooling layer and a fully connected layer. The ReLU activation function is used in all fully connected layers. The input feature of the current multi-scale feature extraction layer is input into the third convolutional layer and then passes through the average pooling layer and the max pooling layer respectively to obtain an average pooling feature and a max pooling feature. The average pooling feature and the max pooling feature are respectively input into the corresponding fully connected layers and then fused to obtain a first fusion feature, as shown in the following formula: ； Among them, represents the output feature of the third convolutional layer, F avg and F max respectively represent average pooling and max pooling operations, α0 and α1 represent the parameters of the fully connected layer, σ represents the sigmoid activation function, and ReLU represents the ReLU activation function, represents the first fused feature; The first fusion feature and the input feature of the current multi-scale feature extraction layer are weighted and fused and then respectively input into the fourth convolutional layer, the fifth convolutional layer, and the sixth convolutional layer of different scales to obtain the first-scale feature, the second-scale feature, and the third-scale feature. The first-scale feature, the second-scale feature, and the third-scale feature are input into the feature fusion layer to obtain the second fusion feature, as shown in the following formula: ； Among them, Z represents the second fusion feature, F j represents the feature at the j-th scale, j = 1, 2,.., m, where m represents the total number of scales, Concat represents the feature fusion operation, and w j represents the weight of the feature at the j-th scale, and its expression is as follows: ； Among them, [] represents the concatenation operation, represents the Softmax activation function; The second fused feature is passed through the Softmax activation function and then residual-connected with the input feature of the current multi-scale feature extraction layer to obtain a third fused feature. The third fused feature is weighted and fused with the second fused feature after passing through the seventh convolutional layer to obtain the output feature of the current multi-scale feature extraction layer. The output feature of the current multi-scale feature extraction layer is then input into the next multi-scale feature extraction layer. The output feature of the last multi-scale feature extraction layer is the multi-scale feature image f b .

4. The method for evaluating the quality of a magnetic resonance image based on non-reference according to claim 3, wherein The number of multi-scale feature extraction layers in the multi-scale feature extraction module is 6. The input feature of the first multi-scale feature extraction layer is the feature obtained by passing the preliminary estimated image output by the first convolutional layer through the ReLU activation function. The input features of the second to fifth multi-scale feature extraction layers are the features obtained by splicing the output feature of the previous multi-scale feature extraction layer and the preliminary estimated image output by the first convolutional layer and then passing through the ReLU activation function. The input feature of the sixth multi-scale feature extraction layer is the feature obtained by splicing the output feature of the previous multi-scale feature extraction layer and the preliminary estimated image output by the first convolutional layer and then passing through the Tanh activation function. The convolution kernels of the third convolutional layer and the fourth convolutional layer are 3×3, and the convolution kernels of the second convolutional layer, the fifth convolutional layer, and the seventh convolutional layer are 5×5. The convolution kernel of the sixth convolutional layer is 7×7.

5. The non-reference based magnetic resonance image quality assessment method according to claim 1, wherein In the magnetic resonance multi-Gaussian blind quality module, the magnetic resonance curve estimation image and the multi-scale feature image are first weighted and fused to obtain the final fusion feature, as shown in the following formula: ； Among them, f b represents the multi-scale feature image, E represents the magnetic resonance curve estimation image, represents the final fusion feature, and t1 and t2 represent the fusion parameters; The Gaussian distributions of the mean, standard deviation, and deviation between the magnetic resonance image and its corresponding magnetic resonance curve estimation image are calculated respectively using the following formula, as shown in the following formula: ； ； ； where p m , p d and p s represent the Gaussian distribution of the mean, standard deviation, and skewness difference between the magnetic resonance image and its corresponding estimated magnetic resonance curve image, and σ m , σ d and σ s respectively represent the average of the averages of the gray values calculated for several local areas in the magnetic resonance image, the average of the standard deviations of the gray values calculated for several regions, and the average of the skewnesses of the gray values calculated for several regions. μ m , μ d and μ s respectively represent the standard deviation of the averages of the gray values calculated for several regions in the magnetic resonance image, the standard deviation of the standard deviations of the gray values calculated for several regions, and the standard deviation of the skewnesses of the gray values calculated for several regions; By calculating the normal distribution using the 3-sigma rule and analyzing the mean, standard deviation, and skewness between the magnetic resonance image and its corresponding magnetic resonance curve estimation image, the quality score of the magnetic resonance image is obtained, as shown in the following formula: ； Among them, represents the quality fraction of the magnetic resonance image, and T represents the transpose.

6. The non-reference-based magnetic resonance image quality assessment method according to claim 1, wherein The convolution kernel of the first convolutional layer is 3×3, and the number of the first convolutional layer is 7. The output features of the first 3 first convolutional layers are the corresponding preliminary estimated images. The output feature of the third first convolutional layer and the output feature of the fourth first convolutional layer are skip-connected to obtain the preliminary estimated image output by the fourth first convolutional layer. The output feature of the second first convolutional layer and the output feature of the fifth first convolutional layer are skip-connected to obtain the preliminary estimated image output by the fifth first convolutional layer. The output feature of the first first convolutional layer and the output feature of the sixth first convolutional layer are skip-connected to obtain the preliminary estimated image output by the sixth first convolutional layer. The preliminary estimated images output by the first 5 first convolutional layers need to pass through the ReLU activation function before being input into the next first convolutional layer, and the preliminary estimated image output by the sixth first convolutional layer needs to pass through the Tanh activation function before being input into the seventh first convolutional layer.

7. A non-reference based magnetic resonance image quality assessment device, characterized in that, Including: A model construction module, configured to construct and train a non-reference-based magnetic resonance image quality assessment model to obtain a trained magnetic resonance image quality assessment model. The magnetic resonance image quality assessment model includes an image estimation module, a multi-scale feature extraction module, a magnetic resonance curve estimation module, and a magnetic resonance multivariate Gaussian blind quality module. The image estimation module includes a number of first convolutional layers, and each first convolutional layer outputs a corresponding preliminary estimated image. An evaluation module, configured to obtain a magnetic resonance image to be evaluated and input it into the trained magnetic resonance image quality assessment model. The magnetic resonance image is first input into the image estimation module, and passes through a number of first convolutional layers in sequence to obtain a number of corresponding preliminary estimated images. The preliminary estimated image output by the first first convolutional layer is input into the multi-scale feature extraction module to extract a multi-scale feature image. The magnetic resonance image, a number of preliminary estimated images, and the multi-scale feature image are input into the magnetic resonance curve estimation module to predict a magnetic resonance curve estimation image. The magnetic resonance curve estimation image, the multi-scale feature image, and the magnetic resonance image are input into the magnetic resonance multivariate Gaussian blind quality module to obtain a corresponding quality score.

8. An electronic device, comprising: One or more processors; A storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method according to any one of claims 1-6.

Images