Medical image quality evaluation method and apparatus

Abstract

The application relates to a medical image quality evaluation method and device, which comprises the following steps: obtaining a sample set through a preset medical image quality evaluation data set, dividing the sample set into a training sample set and a test sample set; inputting the training sample set into a pre-constructed initial medical image quality evaluation network model, iteratively training the initial medical image quality evaluation network model according to a preset loss function, and obtaining a target medical image quality evaluation network model; and inputting the test sample set into the target medical image quality evaluation network model to obtain a medical image quality evaluation result. The initial medical image quality evaluation network model is trained by using the training sample set of the medical image, the target medical image quality evaluation network model obtained after training is used for evaluating the medical image, the visual perception quality of the image can be truly reflected, and the medical image quality evaluation is more in line with the characteristics of the human visual system.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and specifically relates to a method and apparatus for evaluating the quality of medical images. Background Technology

[0002] With the rapid development of modern large-scale medical imaging equipment, new imaging and image processing methods are constantly emerging, and medical images are becoming the basis for clinical medical research, diagnosis, and treatment. Developing medical image quality evaluation methods is of great significance for monitoring and adjusting medical image quality and for testing and optimizing medical image processing algorithms. The most reliable method for evaluating medical image quality is subjective evaluation. However, subjective evaluation methods suffer from drawbacks such as high cost, long processing time, poor real-time performance, susceptibility to subjective and objective factors, and inability to be integrated into practical medical image processing systems. Therefore, objective evaluation methods for medical image quality have gained significant attention.

[0003] In current medical image processing systems, Peak Signal-to-Noise Ratio (PSNR) is primarily used for objective evaluation of medical image quality. However, because PSNR does not consider the correlation between pixels and the perceptual characteristics of the human visual system, the evaluation results cannot truly reflect the visual perception quality of the image. Therefore, it is necessary to develop medical image quality evaluation methods that better conform to the characteristics of the human visual system. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to overcome the shortcomings of the prior art and provide a method and apparatus for evaluating the quality of medical images, so as to solve the problem that the evaluation results of objective evaluation of medical image quality in the prior art cannot truly reflect the visual perception quality of the image.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for evaluating the quality of medical images, comprising:

[0006] The medical image quality assessment dataset consists of test images with preset dimensions, reference images for each test image, and subjective quality rating labels for the test images. The sample set is then divided into a training sample set and a test sample set.

[0007] The training sample set is input into the pre-constructed initial medical image quality assessment network model G, and the initial medical image quality assessment network model G is iteratively trained according to the preset loss function L to obtain the target medical image quality assessment network model G. * ;

[0008] The test sample set is input into the target medical image quality assessment network model to obtain the medical image quality assessment results.

[0009] Furthermore, the initial medical image quality assessment network model includes:

[0010] The network consists of a feature extraction subnetwork D, a feature extraction subnetwork R, a multi-scale visual perception feature extraction subnetwork M, and a quality rating prediction subnetwork P. The feature extraction subnetwork D and the feature extraction subnetwork R are connected in parallel and then connected to the multi-scale visual perception feature extraction subnetwork M and the quality rating prediction subnetwork P in sequence through a feature subtraction operation.

[0011] The scale-based visual perception feature extraction subnetwork M includes parallel local information association modules S. l Module S associated with global information g and sequentially connected feature splicing layers; wherein, the local information association module includes a global pooling layer, a normalization layer, and local unit U. l The global information association module S g Including normalized layers and nonlocal units U g The local unit U l and the nonlocal unit U g Each of them includes multiple convolutional layers;

[0012] The quality rating prediction subnetwork P includes a global pooling layer and multiple fully connected layers.

[0013] The iterative training of the initial medical image quality assessment network model according to the preset loss function includes:

[0014] The initial iteration count k and the maximum iteration count K are preset; where K>20, the current initial medical image quality assessment network model G is... k G k The weight parameter is θ k Let k = 1, G k =G;

[0015] Multiple training samples will be randomly selected without replacement from the training sample set as the current initial medical image quality assessment network model G. k The feature extraction subnetwork D extracts features F from the test image in each training sample as input. o The feature extraction subnetwork R extracts the features F of the reference image of the test image in each training sample. r The multi-scale visual perception feature extraction subnetwork M obtains the features F of the image to be tested. o Features F of the reference image in each training sample r The difference is used to obtain the perceptual error feature F. d For the perceived error feature F d Multi-scale visual perception features F are obtained after local and global special processing. mThe multi-scale visual perception features F m The image quality rating is obtained by inputting it into the quality rating prediction subnetwork P.

[0016] The error between the predicted quality rating and the subjective quality rating label corresponding to each training sample is calculated using the loss function L, and this error is used as the current initial medical image quality assessment network model G. k loss value d k Then through the loss value d k Calculate the loss value G k The weight parameter θ k partial derivative g k Then, the stochastic gradient descent method is used, through the partial derivative g k For G k The weight parameter θ k Update;

[0017] Determine whether k = K holds true. If so, obtain the target medical image quality assessment network model G. * Otherwise, let k = k + 1 and repeat the training steps.

[0018] Furthermore, the perceptual error features are processed locally and globally to obtain multi-scale visual perceptual features, including:

[0019] Global Information Association Module S g For the perceived error feature F d Pixel-by-pixel sampling along the third channel dimension yields multiple feature vectors. and all feature vectors spliced ​​into nonlocal feature X g Local information association module S l For the perceived error feature F d Global pooling yields multiple feature vectors. and all feature vectors spliced ​​into local feature X l ;

[0020] Global Information Association Module S g The normalization layer affects nonlocal features X g Normalization is performed to obtain the normalized nonlocal feature X' g Local information association module S l The normalization layer affects local features X l Normalization is performed to obtain the normalized nonlocal feature X' l ;

[0021] Global Information Association Module S g Central African local unit U gFor the normalized nonlocal features X' g Extending along the third dimension of the feature extraction sub-network D yields the nonlocal extended feature X”. g and along the third dimension for X” g The decomposition yields three weight matrices Q of the same size. g K g V g Then for Q g K g V g Perform convolution operations separately, then pass Q... g K g Convolution results Calculate the nonlocal weight matrix A g Finally, through the nonlocal weight matrix A g For nonlocal features X' g We perform weighted summation to obtain the global perception feature F. g ;

[0022] Local information association module S l Central African local unit U l For the normalized nonlocal features X' l Extending along the third dimension of the feature extraction sub-network D yields the nonlocal extended feature X. l ", and along the third dimension for X l "Decomposition yields three weight matrices Q of the same size." l K l V l Then for Q l K l V l Perform convolution operations separately, then pass Q... l K l Convolution results Calculate the nonlocal weight matrix A l Finally, through A l For nonlocal features X' l We perform weighted summation to obtain the local perceptual features F. l ;

[0023] Feature concatenation layer for globally perceived features F g Perceptual error characteristics F d and local quality perception features F l By stitching along the third dimension, we obtain the multi-scale visual perception features F. m .

[0024] Furthermore, the feature extraction subnetwork D and the feature extraction subnetwork R have the same structure, and both feature extraction subnetwork D and feature extraction subnetwork R include: a first convolutional layer, a max pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer and a seventh convolutional layer;

[0025] The input of the second convolutional layer is connected to the output residual of the fourth convolutional layer, and the input of the fifth convolutional layer is connected to the output residual of the seventh convolutional layer. The first convolutional layer has a kernel size of 7×7, a number of kernels of 64, and a stride of 2. The max pooling layer has a pooling window size of 3×3 and a stride of 2. The second, fourth, fifth, and seventh convolutional layers have kernel sizes of 1×1, with a number of kernels of 64, 256, 128, and 512, respectively, and a stride of 1. The third and sixth convolutional layers have kernel sizes of 3×3, with a number of kernels of 128 and 256, respectively, and a stride of 2.

[0026] Furthermore, the expression for the loss function L of the initial medical image quality assessment network model G is as follows:

[0027]

[0028] Among them, s i , Let represent the subjective quality rating and the predicted quality rating of the i-th training sample in a plurality of training samples, respectively; ||·||1 represents the first-order norm regularization, Σ represents the summation symbol, and δ represents the threshold.

[0029] Furthermore, using the stochastic gradient descent method, the partial derivative g is employed in the following manner. k For G k The weight parameter θ k Update

[0030]

[0031] in, G represents the current initial medical image quality assessment network model. k The updated weight parameters; α represents the learning rate.

[0032] Furthermore, the feature concatenation layer uses the following calculation method to process the globally perceived features F. g Perceptual error characteristics F d and local quality perception features F l By stitching along the third dimension, we obtain the multi-scale visual perception features F. m ,

[0033]

[0034] in, This represents the matrix dot product operation.

[0035] Furthermore, through the nonlocal weight matrix A g The following method is used to process nonlocal features X' g We perform weighted summation to obtain the global perception feature F. g ,

[0036]

[0037]

[0038] Through the nonlocal weight matrix A l The following method is used to process the nonlocal feature X' l We perform weighted summation to obtain the local perceptual features F. l ,

[0039]

[0040]

[0041] Where soft max(·) denotes the normalization exponential function, express The transpose operation. express The transpose operation. This represents the matrix dot product operation.

[0042] This application provides a medical image quality assessment device, including:

[0043] The acquisition module is used to acquire a sample set through a preset medical image quality evaluation dataset, and divide the sample set into a training sample set and a test sample set;

[0044] The training module is used to input the training sample set into a pre-constructed initial medical image quality assessment network model G, and iteratively train the initial medical image quality assessment network model G according to a preset loss function L to obtain the target medical image quality assessment network model G. * ;

[0045] The evaluation module is used to input the test sample set into the target medical image quality evaluation network model to obtain the medical image quality evaluation results.

[0046] The beneficial effects that can be achieved by adopting the above technical solution in this invention include:

[0047] This invention provides a method and apparatus for medical image quality assessment. This application uses a training sample set of medical images to train an initial medical image quality assessment network model, and uses the target medical image quality assessment network model obtained after training to evaluate the medical images. This method can truly reflect the visual perception quality of the images and is a medical image quality assessment that is more in line with the characteristics of the human visual system. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a schematic diagram illustrating the steps of the medical image quality evaluation method of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of the initial medical image quality assessment network model provided by the present invention;

[0051] Figure 3 This is a schematic diagram of the medical image quality evaluation device of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0053] The following describes a specific medical image quality evaluation method and apparatus provided in an embodiment of this application, with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the medical image quality assessment method provided in this application embodiment includes:

[0055] S101, the medical image quality evaluation dataset is divided into a training sample set and a test sample set, consisting of the test images with preset dimensions, the reference images of each test image, and the subjective quality rating labels of the test images.

[0056] The medical image quality assessment dataset can be medical images stored in the data in advance. In this application, the medical image quality assessment dataset is composed of test images with preset dimensions, reference images of each test image, and subjective quality rating labels of the test images. The sample set is divided into a training sample set and a test sample set.

[0057] Specifically, obtain L test images of dimension W×H from the medical image quality assessment dataset, along with a reference image for each test image, and form a sample set A = {(I... l ,R l ,s l )|1≤l≤L}, and form a training sample set T={(I m ,R m ,s m Given the N samples, we can form a test sample set E = {(I | 1 ≤ m ≤ M}, where I ≤ m ≤ M}. n ,s n Let W = |1≤n≤N}, where W and H represent the width and height of the image to be measured, respectively, W≥224, H≥224, L≥1000. M+N=L,I l R l s l These represent the l-th image to be tested, its reference image, and the corresponding subjective quality rating label, respectively.

[0058] S102, input the training sample set into the pre-constructed initial medical image quality assessment network model G, and iteratively train the initial medical image quality assessment network model G according to the preset loss function L to obtain the target medical image quality assessment network model G. * ;

[0059] In some embodiments, such as Figure 2 As shown, the initial medical image quality assessment network model includes:

[0060] The network consists of a feature extraction subnetwork D, a feature extraction subnetwork R, a multi-scale visual perception feature extraction subnetwork M, and a quality rating prediction subnetwork P. The feature extraction subnetwork D and the feature extraction subnetwork R are connected in parallel and then connected to the multi-scale visual perception feature extraction subnetwork M and the quality rating prediction subnetwork P in sequence through a feature subtraction operation.

[0061] The feature extraction subnetwork D and feature extraction subnetwork R have the same structure, and both feature extraction subnetwork D and feature extraction subnetwork R include: a first convolutional layer, a max pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, a sixth convolutional layer and a seventh convolutional layer;

[0062] The input of the second convolutional layer is connected to the output residual of the fourth convolutional layer, and the input of the fifth convolutional layer is connected to the output residual of the seventh convolutional layer. The first convolutional layer has a kernel size of 7×7, a number of kernels of 64, and a stride of 2. The max pooling layer has a pooling window size of 3×3 and a stride of 2. The second, fourth, fifth, and seventh convolutional layers have kernel sizes of 1×1, with a number of kernels of 64, 256, 128, and 512, respectively, and a stride of 1. The third and sixth convolutional layers have kernel sizes of 3×3, with a number of kernels of 128 and 256, respectively, and a stride of 2.

[0063] The scale-based visual perception feature extraction subnetwork M includes parallel local information association modules S. l Module S associated with global information g and sequentially connected feature splicing layers; wherein, the local information association module includes a global pooling layer, a normalization layer, and local unit U. l The global information association module S g Including normalized layers and nonlocal units U g The local unit U l and the nonlocal unit U g Each of them includes multiple convolutional layers;

[0064] The quality rating prediction subnetwork P includes a global pooling layer and multiple fully connected layers.

[0065] It should be noted that the initial medical image quality assessment network model G has a pre-defined loss function L, the expression of which is:

[0066]

[0067] Among them, s i , Let represent the subjective quality rating and the predicted quality rating of the i-th training sample in a plurality of training samples, respectively; ||·||1 represents the first-order norm regularization, Σ represents the summation symbol, and δ represents the threshold.

[0068] Then, the initial medical image quality assessment network model G is iteratively trained according to the preset loss function L to obtain the target medical image quality assessment network model G. * .

[0069] The specific process of iterative training is as follows:

[0070] Step (1): Initialize the number of iterations to k, the maximum number of iterations to K, K>20, and the current no-reference image quality assessment network model is G. k G k The weight parameter is θ k And let k = 1, Gk =G;

[0071] Step (2) involves randomly selecting B training samples without replacement from the training sample set T as the medical image quality assessment network model G. k The input is used to extract features from the test image in each training sample by the feature extraction subnetwork D, which obtains F. o The feature extraction subnetwork R extracts features from the reference image in each training sample to obtain F. r The multi-scale visual perception feature extraction subnetwork M first calculates F. o With F r The difference is used to obtain the perceptual error feature F. d This is used as input, and then subjected to local and global special processing to obtain multi-scale visual perception features F. m The quality rating prediction subnetwork P will F m As input, the quality rating of the image to be tested is obtained; among them, the multi-scale visual perception feature extraction subnetwork M is used to evaluate the quality of the image. d The steps for implementing local and global processing are as follows:

[0072] Global Information Association Module S g For F d Pixel-by-pixel sampling along the third channel dimension yields multiple feature vectors. and all feature vectors spliced ​​into nonlocal feature X g Local information association module S g For F d Global pooling yields multiple feature vectors. and all feature vectors spliced ​​into local feature X l ;

[0073] In the global information association module, the normalization layer affects non-local features X. g Normalization is performed to obtain the normalized nonlocal feature X' g In the local information association module, the normalization layer affects local features X. l Normalization is performed to obtain the normalized nonlocal feature X' l ;

[0074] Nonlocal unit U in the global information association module g For the normalized nonlocal features X' g Extending along the third dimension of D yields the nonlocal extended feature X”. g And along the third dimension, the nonlocal extended feature X” g The decomposition yields three weight matrices Q of the same size. g K g Vg Then for Q g K g V g Perform convolution operations separately, then pass Q... g K g Convolution results Calculate the nonlocal weight matrix A g Finally, through A g For nonlocal features X' g We perform weighted summation to obtain the global perception feature F. g :

[0075]

[0076]

[0077] Where soft max(·) denotes the normalization exponential function, express The transpose operation;

[0078] Feature concatenation layer for globally perceived features F g Perceptual error characteristics F d and local quality perception features F l By stitching along the third dimension, we obtain the multi-scale visual perception features F. m :

[0079]

[0080] in, This represents the matrix dot product operation.

[0081] Step (3) uses the loss function L and calculates the error between the predicted quality rating and the subjective quality rating label corresponding to each training sample as G. k loss value d k Then through d k Calculate G k The weight parameter θ k partial derivative g k Then, the stochastic gradient descent method is used, through the partial derivative g k For G k The weight parameter θ k Update;

[0082] In some embodiments, stochastic gradient descent is employed, using the partial derivative g in the following manner. k For G k The weight parameter θ k Update

[0083]

[0084] in, G represents the current initial medical image quality assessment network model. k The updated weight parameters; α represents the learning rate.

[0085] Step (4): Determine whether k = K holds true. If so, obtain the trained no-reference image quality assessment network model G based on curriculum learning. * Otherwise, let k = k + 1 and execute step (2).

[0086] S103, input the test sample set into the target medical image quality evaluation network model to obtain the medical image quality evaluation result.

[0087] Specifically, after obtaining the target medical image quality assessment network model, the test sample set E is used as the trained medical image quality assessment network model G based on multi-scale visual perception mechanism. * The input is used for forward inference to obtain the quality prediction score for each test sample.

[0088] like Figure 3 As shown, a medical image quality assessment device includes:

[0089] The acquisition module 201 is used to acquire a sample set through a preset medical image quality evaluation dataset, and divide the sample set into a training sample set and a test sample set;

[0090] Training module 202 is used to input the training sample set into a pre-constructed initial medical image quality assessment network model G, and iteratively train the initial medical image quality assessment network model G according to a preset loss function L to obtain the target medical image quality assessment network model G. * ;

[0091] Evaluation module 203 is used to input the test sample set into the target medical image quality evaluation network model to obtain medical image quality evaluation results.

[0092] The working principle of the medical image quality assessment device provided in this application is as follows: the acquisition module 201 acquires a sample set through a preset medical image quality assessment dataset, and divides the sample set into a training sample set and a test sample set; the training module 202 inputs the training sample set into a pre-constructed initial medical image quality assessment network model G, and iteratively trains the initial medical image quality assessment network model G according to a preset loss function L to obtain the target medical image quality assessment network model G. * The evaluation module 203 inputs the test sample set into the target medical image quality evaluation network model to obtain the medical image quality evaluation results.

[0093] In summary, this invention provides a method and apparatus for medical image quality assessment, comprising: obtaining a sample set through a preset medical image quality assessment dataset; dividing the sample set into a training sample set and a test sample set; inputting the training sample set into a pre-constructed initial medical image quality assessment network model; iteratively training the initial medical image quality assessment network model according to a preset loss function L to obtain a target medical image quality assessment network model; and inputting the test sample set into the target medical image quality assessment network model to obtain the medical image quality assessment result. This invention uses a training sample set of medical images to train the initial medical image quality assessment network model, and uses the trained target medical image quality assessment network model to evaluate the medical images, which can realistically reflect the visual perception quality of the images and is a medical image quality assessment method that is more consistent with the characteristics of the human visual system.

[0094] It is understood that the method embodiments provided above correspond to the device embodiments described above, and the specific details can be referred to each other, which will not be repeated here.

[0095] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0096] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0098] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for evaluating the quality of medical images, characterized in that, include: The medical image quality assessment dataset consists of test images with preset dimensions, reference images for each test image, and subjective quality rating labels for the test images. The sample set is then divided into a training sample set and a test sample set. Input the training sample set into the pre-constructed initial medical image quality assessment network model In the middle, according to the preset loss function L The initial medical image quality assessment network model Iterative training was performed to obtain the target medical image quality assessment network model. ; The test sample set is input into the target medical image quality assessment network model to obtain the medical image quality assessment results; The initial medical image quality assessment network model includes: Feature extraction subnetwork Feature extraction subnetwork Multi-scale visual perception feature extraction subnetwork and quality rating prediction subnetwork The feature extraction subnetwork and the feature extraction subnetwork After parallel connection, the multi-scale visual perception feature extraction subnetwork is sequentially connected through feature subtraction operation. and quality rating prediction subnetwork ; Among them, the scale visual perception feature extraction subnetwork Includes parallel local information association modules Module associated with global information and a sequentially connected feature splicing layer; wherein, the local information association module includes a global pooling layer, a normalization layer, and local units. The global information association module Including normalized layers and nonlocal units The local unit and the nonlocal unit Each of them includes multiple convolutional layers; The quality rating prediction subnetwork Includes a global pooling layer and multiple fully connected layers; The iterative training of the initial medical image quality assessment network model according to the preset loss function includes: Preset initial iteration count and maximum number of iterations ;in, Current initial medical image quality assessment network model , The weight parameters are ,make , ; Multiple training samples will be randomly selected without replacement from the training sample set as the current initial medical image quality assessment network model. The input of the feature extraction subnetwork Extract features from the test image in each training sample The feature extraction subnetwork Extract features from the reference image of the test image in each training sample. The multi-scale visual perception feature extraction subnetwork Determine the features of the image to be tested Features of the reference image in each training sample The difference is used to obtain the perceptual error characteristics. Regarding the perceived error features Multi-scale visual perception features are obtained after local and global special processing. The multi-scale visual perception features Input to the quality rating prediction subnetwork To obtain the quality rating of the image under test; Through the loss function The error between the predicted quality rating for each training sample and the subjective quality rating label for that training sample is used as the current initial medical image quality assessment network model. loss value Then through the loss value Calculate the loss value weight parameters partial derivatives Then, the stochastic gradient descent method is used, through partial derivatives... right weight parameters Update; judge If the condition is met, then the target medical image quality assessment network model is obtained. Otherwise, let Repeat the training steps; The multi-scale visual perception features obtained by processing the perception error features through local and global special processing include: Global Information Association Module For the perceived error features Pixel-by-pixel sampling along the third channel dimension yields multiple feature vectors. and all feature vectors splicing into nonlocal features Local information association module For the perceived error features Global pooling yields multiple feature vectors. and all feature vectors splicing into local features ; Global Information Association Module The normalization layer for nonlocal features Normalization is performed to obtain the normalized nonlocal features. Local information association module The normalization layer for local features Normalization is performed to obtain the normalized nonlocal features. ; Global Information Association Module Central African local unit For normalized nonlocal features Extending along the third dimension of the feature extraction sub-network D yields nonlocal extended features. and along the third dimension The decomposition yields three weight matrices of the same size. , , Then to , , Perform convolution operations separately, and then... , Convolution results , Calculate the nonlocal weight matrix Finally, through the nonlocal weight matrix For nonlocal features Weighting is performed to obtain global perception features. ; Local information association module Central African local unit For normalized nonlocal features Extending along the third dimension of the feature extraction sub-network D yields nonlocal extended features. and along the third dimension The decomposition yields three weight matrices of the same size. , , Then to , , Perform convolution operations separately, and then... , Convolution results , Calculate the nonlocal weight matrix Finally passed For nonlocal features Weighting is performed to obtain local perceptual features. ; Feature splicing layer for globally perceptive features Perceptual error characteristics and local quality perception features By stitching along the third dimension, multi-scale visual perception features are obtained. .

2. The method according to claim 1, characterized in that, The feature extraction subnetwork Feature extraction subnetwork The structure is the same as that of the feature extraction subnetwork. Feature extraction subnetwork All of them include: the first convolutional layer, the max pooling layer, the second convolutional layer, the third convolutional layer, the fourth convolutional layer, the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer; The input of the second convolutional layer is connected to the output residual of the fourth convolutional layer, and the input of the fifth convolutional layer is connected to the output residual of the seventh convolutional layer. The kernel size of the first convolutional layer is... The number of pools is 64, the step size is 2; the maximum pooling layer pooling window size is... The stride is 2; the kernel sizes of the second, fourth, fifth, and seventh convolutional layers are... The numbers are 64, 256, 128, and 512 respectively, with a stride of 1; the kernel sizes of the third and sixth convolutional layers are... The quantities are 128 and 256 respectively, with a step size of 2.

3. The method according to claim 1, characterized in that, The initial medical image quality assessment network model loss function The expression is, in, , They represent the nth training sample among multiple training samples. Subjective quality rating and predicted quality rating corresponding to each training sample; This represents first-order norm regularization. The summation symbol is used to represent the summation symbol. This represents the threshold.

4. The method according to claim 2, characterized in that, Using stochastic gradient descent, partial derivatives are employed in the following manner. right weight parameters Update in, This represents the current initial medical image quality assessment network model. Updated weight parameters; This represents the learning rate.

5. The method according to claim 1, characterized in that, The feature concatenation layer uses the following calculation method to process globally perceived features. Perceptual error characteristics and local quality perception features By stitching along the third dimension, multi-scale visual perception features are obtained. , in, This represents the matrix dot product operation.

6. The method according to claim 1, characterized in that, Through the nonlocal weight matrix The following methods are used for nonlocal features Weighting is performed to obtain global perception features. , Through the nonlocal weight matrix The following method is used for nonlocal features Weighting is performed to obtain local perceptual features. , in, Represents the normalized exponential function, express The transpose operation. express The transpose operation. This represents the matrix dot product operation.

7. A medical image quality evaluation device, characterized in that, The medical image quality assessment device, applied to the medical image quality assessment method as described in any one of claims 1 to 6, comprises: The acquisition module is used to acquire a sample set through a preset medical image quality evaluation dataset, and divide the sample set into a training sample set and a test sample set; The training module is used to input the training sample set into the pre-built initial medical image quality assessment network model. In the middle, according to the preset loss function L The initial medical image quality assessment network model Iterative training was performed to obtain the target medical image quality assessment network model. ; The evaluation module is used to input the test sample set into the target medical image quality evaluation network model to obtain the medical image quality evaluation results.

Images