Unsupervised canonical learning method and system for diffusion magnetic resonance image quality assessment

Abstract

The application designs an unsupervised canonical learning method and system for diffusion magnetic resonance image quality evaluation. Firstly, the natural logarithm method is used to normalize and unify the multi-b value images, wherein b is a diffusion sensitive gradient parameter. Then, a vector quantization variational autoencoder (VQ-VAE) is used to train a model on a standard (high-quality artifact-free) data set. Next, the image containing artifacts and the artifact-free image are input into the previously trained model, the structural similarity (SSIM) index before and after reconstruction is compared, and finally the quality of the data is evaluated to determine whether it belongs to the standard data, and the probability of the detection sample belonging to the artifact data and the standard data is given to assist the judgment of the experts. The application applies the unsupervised generative model and the anomaly detection method to the recognition of diffusion magnetic resonance artifacts, realizes the unlabeled training of the data, and solves the disadvantage of the lack of a large number of labeled artifacts in the training data.

Description

Technical Field

[0001] This invention belongs to the fields of magnetic resonance imaging, medical image processing and auxiliary detection technology, and specifically relates to an unsupervised normative learning method and system for quality assessment of diffusion magnetic resonance images. Background Technology

[0002] Diffusion magnetic resonance imaging (dMRI) is a magnetic resonance imaging technique used to observe the movement of water molecules in biological tissues. dMRI scans are susceptible to various types of artifacts, posing significant risks to clinical diagnosis and treatment. Artifacts can distort or blur image information, interfering with the physician's accurate diagnosis of patient lesions or abnormalities. This can lead to missed important lesions, misdiagnosis, or delayed diagnosis, posing serious health risks to patients. Because artifacts can degrade image quality or cause missing information, physicians may be unable to fully assess the patient's condition when developing a treatment plan, affecting the selection and implementation of treatment options. Furthermore, artifacts increase unnecessary examinations and costs for patients. When artifacts are present, physicians may need to perform additional examinations to confirm the diagnosis, increasing the waste of medical resources, patient inconvenience, and potentially unnecessary radiation exposure. Therefore, artifacts in dMRI pose a serious threat to clinical diagnosis and treatment, potentially leading to severe consequences. Reducing the impact of artifacts is crucial for accurate physician judgment and effective patient treatment.

[0003] Prior to subsequent analysis, quality assessment of dMRI data is crucial. This allows for improved data quality through appropriate preprocessing, ensuring that the presence of artifacts does not negatively impact the results of later analyses. Manual quality assessment of data is subjective, prone to error, and impractical, especially given the increasing number of similar collaborative studies, highlighting the necessity for automating this process. Currently, dMRI data quality control and artifact identification methods rely heavily on manual inspection—a labor-intensive and subjective process requiring significant effort and human resources from experts meticulously examining large datasets or individual slices.

[0004] Recently, a deep learning-based automated pipeline has been developed to accurately identify artifacts in diffusion scans using a 3D architecture. This architecture allows for labeling and training within the same timeframe as 2D architectures, but requires three times the amount of data (Ahmad, A., et al.: 3D-QCNet – A pipeline for automated artifact detection indiffusion MRI images. Comput. Med. Imaging Graph. 103, 102151 (2023)); however, it can only identify problematic volumes and cannot effectively locate slices containing specific artifacts. Furthermore, most existing methods rely on labeled medical image data, and the challenge of acquiring large amounts of medical image data, coupled with privacy and sharing restrictions, remains a persistent issue. To address the aforementioned issues, the problem of insufficient training samples can be solved by simulating motion and eddy current artifacts, generating high-quality training data without manual labeling (Graham, M., S., Drobnjak, I., & Zhang, H.: A supervised learning approach for diffusion MRI quality control with minimal training data. NeuroImage 178, 668–676 (2018)). However, this method primarily targets specific artifacts, leaving room for other types of artifacts. Various types of artifacts can occur when acquiring dMRI data. Most techniques rely on expensive and time-consuming supervised learning to identify these artifacts, but due to the vast amount of dMRI data and the limited number of labeled samples, a large amount of labeled data containing various artifact types is required. Therefore, how to reduce or discard data labeling while still being able to identify artifacts not seen during training for each slice is a significant and urgent challenge that needs to be addressed. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention proposes an unsupervised canonical learning method and system for quality assessment of diffusion magnetic resonance imaging. This method applies unsupervised generative models and anomaly detection methods to the identification of diffusion magnetic resonance artifacts, realizes the unlabeling of training data, solves the problem of lacking a large number of labeled artifacts in the training data, and thus reduces the human resources required for data collection and labeling.

[0006] An unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment includes the following steps:

[0007] Step 1: Collect artifact-free diffusion magnetic resonance images and extract cross-sectional midbrain slice images from them, i.e., multi-b value midbrain slice images.

[0008] Step 2: Based on the attenuation curve of the magnetic resonance signal in the diffusion-sensitive gradient, normalized data preprocessing is performed on the multi-b-value midbrain slice images. The normalized data preprocessing of the collected multi-b-value midbrain slice images is performed to minimize the influence of the difference in midbrain image signal intensity under different b-values, and the normalized preprocessed b-value midbrain slice images are obtained as the training set.

[0009] Step 3: Input the normalized preprocessed b-value midbrain slice image into the VQ-VAE network for reconstruction to obtain the trained VQ-VAE model;

[0010] Step 4: Input the sample to be tested into the VQ-VAE model obtained in Step 3 to obtain the reconstructed b-value midbrain slice image. Compare the structural similarity index of each multi-b-value midbrain slice image before and after reconstruction, and then determine the threshold. Images above the threshold are images without artifacts and of good quality; images below the threshold are images containing artifacts and of poor quality.

[0011] Step 5: Based on the threshold in Step 4, compare the distance of the structural similarity index of each slice to the threshold, and then perform a quality assessment to determine the probability that it belongs to the standard image.

[0012] Furthermore, the diffuse magnetic resonance images were obtained using the SUDMEX dataset.

[0013] Furthermore, step 2 specifically includes:

[0014] The attenuation law of magnetic resonance signals in the presence of diffusion-sensitive gradients is as follows:

[0015] S(b)=S0e -bD

[0016] In the formula, b represents the sensitive diffusion gradient parameter, S represents the signal intensity of the voxel under different b values, S0 represents the signal intensity when the b value is 0, and D represents the diffusion characteristics of fluid molecules in a region.

[0017] For different signal intensities corresponding to different b values ​​in the above formula, natural logarithmic normalization preprocessing is performed:

[0018]

[0019] Natural logarithmic normalization preprocessing was performed on each voxel to obtain normalized preprocessed b-valued midbrain slice images.

[0020] Furthermore, VQ-VAE, or Vector Quantization Variational Autoencoder, introduces a discrete latent space, simplifying the model's understanding of the data and producing more detailed images. The purpose of using VQ-VAE in this invention is to learn standard features from dMRI data and map each image patch to a discrete latent variable Z. q To evaluate dMRI image quality. For the reconstructed portion of VQ-VAE, let X represent the original image. This represents the reconstructed image. The elements in the code are called embedding vectors, denoted as e. VQ-VAE maps K discrete codes to a D-dimensional embedding space Z. e In (x), where Z e (x)∈R D , representing the encoder's output vector. The loss function of VQ-VAE consists of two error components:

[0021] The first part consists of the reconstruction error between the original image and the reconstructed image:

[0022] The second part consists of the error generated during the vector quantization process;

[0023] The VQ-VAE network uses the following loss function:

[0024]

[0025] In the above equation, the quantization loss is calculated using a minimum distance quantizer, which maps each latent variable encoding to the embedding vector with the nearest Euclidean distance. X represents the original image. The reconstructed image is represented by the element called the embedding vector, denoted as e, and sg[·] represents the gradient stopping operator. Since the backpropagation error cannot be directly propagated, VQ-VAE uses a pass-through estimator to approximate the gradient, thus facilitating the implementation of backpropagation.

[0026] Furthermore, the μ-2σ principle is used to determine the discrimination threshold for data that conforms to a normal distribution, while Chebyshev's inequality is used to determine the threshold for data that does not conform to a normal distribution.

[0027] Furthermore, for data that conforms to a normal distribution, the threshold is determined using the μ-2σ principle, i.e.:

[0028] P(|x-μ|<2σ)≈0.954

[0029] In the formula, μ and σ represent the mean and variance of the batch of samples to be tested, respectively. This formula indicates that outlier data has a very high probability of falling outside μ-2σ.

[0030] Furthermore, for data that do not conform to a normal distribution, the threshold is determined using Chebyshev's inequality, i.e.:

[0031]

[0032] In the formula, λ is any real number greater than 1. In this invention, λ = 2 is taken, meaning that data that does not conform to a normal distribution are also determined using the μ-2σ principle, i.e., μ-2σ is used as the threshold T.

[0033] Furthermore, the probability of conforming to the standard image is:

[0034]

[0035] Where k is the scaling factor, d represents the distance of the structural similarity index of each slice to the threshold, d∈[-1,1], the distance is positive above T and negative below T.

[0036] Furthermore, k = 20.

[0037] An unsupervised canonical learning system for quality assessment of diffusion magnetic resonance imaging (DMRI) images is characterized by comprising a data acquisition and image processing unit, a data preprocessing unit, a model training unit, a threshold determination unit, and an image evaluation unit, wherein:

[0038] The image acquisition unit collects artifact-free diffusion magnetic resonance images and extracts cross-sectional midbrain slice images from them, i.e., multi-b-value midbrain slice images.

[0039] The data preprocessing unit performs normalized data preprocessing on multi-b-value midbrain slice images based on the attenuation curve of the magnetic resonance signal in the diffusion-sensitive gradient, and obtains normalized preprocessed b-value midbrain slice images as the training set.

[0040] The model training unit inputs the normalized preprocessed b-value midbrain slice image into the VQ-VAE network for reconstruction, thereby obtaining the trained VQ-VAE model;

[0041] The threshold determination unit inputs the sample to be detected into the VQ-VAE model to obtain the reconstructed b-value midbrain slice image, compares the structural similarity index of each multi-b-value midbrain slice image before and after reconstruction, and then determines the threshold. Images above the threshold are images without artifacts and of acceptable quality; images below the threshold are images containing artifacts and of unacceptable quality.

[0042] The image evaluation unit compares the distance of the structural similarity index of each slice to the threshold, and then performs a quality assessment to determine the probability that it belongs to a standard image.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] (1) An unsupervised training method is adopted to avoid the labeling work of a large amount of training data. Most of the current quality assessment methods for diffusion magnetic resonance imaging are carried out in a supervised manner. Due to the special nature of medical image data, it is extremely difficult to obtain a large number of labeled diffusion magnetic resonance images containing artifacts as training data. This invention studies the above problems from the perspective of unsupervised anomaly detection. It adopts an unsupervised generative model vector quantization variational autoencoder VQ-VAE, which only requires artifact-free dMRI data for training, thus avoiding the labor-intensive work of collecting and labeling artifacts.

[0045] (2) A standardized data processing method was adopted to improve the generalization of the model. The characteristic of dMRI data is that the signal value decreases exponentially with the increase of diffusion gradient (b value). Due to different instruments and actual needs, dMRI images with different b values ​​will be scanned, which is a major challenge to the generalization and robustness of the model. Based on the signal attenuation law curve, this invention proposes a standardized preprocessing method for image signals under different b values, so that the image to be detected is no longer affected by the b value, thereby enabling the model to robustly identify images with various b values.

[0046] (3) Based on statistical principles, it has strong theoretical support. It can not only provide the results of data quality assessment, but also give the probability of it containing artifacts. Since a small number of data containing artifacts can be regarded as outliers compared with a large number of normal data, this invention selects μ-2σ as the dividing line T between outliers and normal values ​​based on the statistical principle of outlier removal. The farther the test sample is above T, the lower the probability that the sample contains artifacts. The farther it is below T, the higher the probability that the sample contains artifacts. Samples at the threshold are indistinguishable samples (may only contain a small number of artifacts). By adding them to the improved Sigmoid function, the final probability of containing artifacts can be obtained. However, whether it is 3D-QCNet or most other methods, they can only give whether it is a positive or negative sample, without discussing the probability of which category it belongs to. This advantage of this invention can more meticulously assist image experts in making judgments.

[0047] (4) A slice-level analysis strategy was adopted, which enabled quality assessment of each midbrain slice of the input dMRI data, which is more detailed than 3D-QCNet, which can only analyze volume-level data. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention.

[0049] Figure 2 This is a flowchart of the unsupervised normative learning method for quality assessment of diffusion magnetic resonance imaging according to the present invention.

[0050] Figure 3 The diagram shows the effect of the method according to the present invention on real human brain data in clinical datasets of different properties.

[0051] Figure 4 The diagram shows the effect of the method according to the present invention on real human brain data with different acquisition parameters. Detailed Implementation

[0052] To enable those skilled in the art to better understand the technical solution of this application, the technical solution provided by this application will be described in further detail and completely below with reference to the accompanying drawings and specific embodiments:

[0053] Combination Figure 1 An unsupervised normative learning method for quality assessment of diffusion magnetic resonance imaging includes the following steps:

[0054] Step 1: Collect high-quality, artifact-free diffusion-weighted magnetic resonance (DWI) images of the subjects, and extract axial plane midbrain slices from them. These slices are all from midbrain slices that meet quality control standards in the SUDMEX dataset. The dMRI data size of each subject in the SUDMEX dataset is 112×112×50×136, and the slice data size along the z-axis is 112×112. This invention selects data with third-dimensional index numbers 24, 25, and 26 to characterize midbrain information; high-quality, artifact-free multi-b-value diffusion-weighted images (DWI) are selected, and other DWI images containing artifacts are not required. The entire training process is unsupervised.

[0055] Step 2: Based on the attenuation curve of the magnetic resonance signal in the diffusion-sensitive gradient, normalize the original multi-b-value images to minimize the difference in image brightness at different b-values. The attenuation formula is as follows:

[0056]

[0057] In the above formula, the value of b represents the sensitive diffusion gradient parameter, S ij (b) represents the signal strength of voxel (i,j) at this value of b, S ij (0) represents the signal strength of voxel (i,j) when b is 0, D ij It is a constant representing the dispersion characteristics of fluid molecules within a region.

[0058] Since the D-value is approximately the same for the same brain region, based on this theory, we perform the following normalization preprocessing on images with different b-values:

[0059]

[0060] In this embodiment, preferably, as follows: Figure 2 As shown in the data preprocessing section, in the SUDMEX dataset, the data with b=1000 and 3000 are all within the same brightness range after natural logarithmic preprocessing. This embodiment only shows the preprocessing operation of the SUDMEX dataset. The above normalization measures are also adopted for other datasets with different b values, such as the BTC tumor dataset, which will not be described separately here.

[0061] Based on the attenuation principle of magnetic resonance signals, the influence of image signal differences under different b-values ​​is reduced, thereby unifying the input network, better learning image features, and improving model stability. Even when testing images with different b-values, the influence of signal strength can be adaptively eliminated, achieving end-to-end input and output of images with multiple b-values.

[0062] Step 4: Input the dMRI image obtained in Step 3 into the VQ-VAE network for reconstruction to obtain the trained generative model. This embodiment is based on the VQ-VAE generative model. The basic principle of VQ-VAE is introduced below.

[0063] VQ-VAE, or Vector Quantization Variational Autoencoder, introduces a discrete latent space, simplifying the model's understanding of the data and producing more detailed images. This invention uses VQ-VAE to learn standard features from dMRI data and maps each image patch to a discrete latent variable Z. q To evaluate dMRI image quality.

[0064] In this embodiment, preferably, as follows: Figure 2 The VQ-VAE reconstruction section is shown in the image, where X represents the original image. This represents the reconstructed image. The elements in the code are called embedding vectors, denoted as e. VQ-VAE maps K discrete codes to a D-dimensional embedding space Z. e In (x), where Z e (x)∈R D , representing the encoder's output vector. The loss function of VQ-VAE consists of two error components:

[0065] 1) The first part consists of the reconstruction error between the original image and the reconstructed image;

[0066] 2) The second part consists of the error generated during the vector quantization process;

[0067] The specific loss function is as follows:

[0068]

[0069] In the above equation, the quantization loss is calculated using a minimum distance quantizer, which maps each latent variable encoding to the embedding vector with the nearest Euclidean distance. `sg[·]` represents the gradient stopping operator. Since the backpropagation error cannot be directly propagated, VQ-VAE uses a pass-through estimator to approximate the gradient, thus facilitating the implementation of backpropagation.

[0070] All artifact-free, standard-compliant data obtained in step 3 are input into the VQ-VAE network as two-dimensional slices for reconstruction. This process does not require label information or labeling of the original data.

[0071] Step 5: Collect standard images and artifact images for different b values ​​from various datasets. After the preprocessing operations in Step 3, input all test samples into the model obtained in Step 4 and compare the structural similarity (SSIM) index before and after reconstruction. Since this method trains the VQ-VAE model with standard data, the slices obtained by inputting standard data will be as similar as possible to the slices before reconstruction, resulting in a higher SSIM index. If the input contains slices with artifacts, the reconstructed slices will have an effect close to that of the denoised slices, and the result will be as close as possible to the slices without artifacts, thus the similarity between the two will be low, and the SSIM index will be low. Figure 2 As shown in the artifact anomaly detection section below, the two types of slices tested exhibit two different distributions: slices that meet the standards generally have higher SSIM values, while slices containing artifacts generally have lower SSIM values.

[0072] The small number of artifact-containing slice samples input in step 5 are outliers compared to the standard data. If these samples follow a normal distribution, the model uses the mean-standard deviation method to remove outliers. The probability that a sample conforming to the standard falls within a given range can be given by the following formula:

[0073] P(|x-μ|<σ)≈0.683; P(|x-μ|<2σ)≈0.954; P(|x-μ|<3σ)≈0.997#(9)

[0074] In the formula, x represents a sample that meets the standard, and μ and σ represent the mean and standard deviation of this batch of samples, respectively. As can be seen from the above formula, outliers will fall outside the mean-standard deviation with a very high probability. If the data to be detected does not conform to a normal distribution, this invention is also supported by statistical theory, namely, according to Chebyshev's inequality, the probability of any sample falling within a given range can be given by the following formula:

[0075]

[0076] In the formula, λ represents a real number greater than 1. This formula shows that even if the overall sample population does not conform to a normal distribution, outliers will fall outside μ-2σ and μ-3σ with a relatively high probability. To ensure the robustness of the classification threshold, this method uses μ-2σ as the threshold T for class division. Samples below T are considered non-standard artifact data; samples above T are considered standard data. Figure 2 As shown, a threshold is automatically selected for each batch of samples, ensuring the adaptability and flexibility of the threshold selection.

[0077] To better assist experts and doctors in assessing the quality of dMRI, this invention can not only determine whether the sample to be tested belongs to standard data or artifact data, but also provide the probability of belonging to each category.

[0078] Step 6: For the threshold T selected in Step 5, compare the directed distance d from the SSIM index of each slice to the threshold T, and then perform a quality assessment to determine the probability that it belongs to the standard image or does not meet the standard image. The probability is described by the following formula:

[0079]

[0080] In the formula, f(d) represents the probability that a sample at a distance d from the threshold T belongs to the standard data. The formula shows that the farther a sample is above the threshold T, the greater its probability of belonging to the standard data; conversely, the farther a sample is below the threshold T, the greater its probability of belonging to the artifact data. Here, k is chosen as 20, and the resulting probability P... standard As the probability of belonging to data that conforms to the standard, 1-P standard The probability P that does not conform to the standard data standard anomaly .

[0081] To verify the feasibility and effectiveness of the proposed technical solution, this method was tested for quality assessment on datasets with different clinical pathological properties, datasets with different acquisition parameters, and datasets that had not undergone preprocessing.

[0082] like Figure 3 As shown, this invention simulates four types of artifacts on datasets of different clinical natures for testing, including the most common ghosting artifact during scanning, and less common spike artifacts, noise artifacts, and exchange artifacts. The first row shows the identification results of the four artifacts in the cocaine subjects (b=1000, 3000s / mm). 2 The second row shows the identification results of four artifacts in randomly selected tumor patients from the BTC dataset (b = 700, 1200, 2800 s / mm). 2 The third row shows the identification results of four artifacts in healthy subjects in the BTC dataset (b = 700, 1200, 2800 s / mm). 2The fourth row shows the identification results of four types of artifacts in the embryo subject data of the dHCP dataset (b = 400, 1000, 2600 s / mm). 2 Blue dots represent slices that meet the standards, while yellow dots represent slices that do not. The results show that this invention can robustly identify most artifacts not only on trained data but also robustly assess image quality on clinical data across age groups and diseases.

[0083] like Figure 4 As shown, this invention simulates four types of artifacts on data with different scanning parameters and without preprocessing, including the most common ghosting artifact and less common spike artifacts, noise artifacts, and exchange artifacts. The first row shows the echo time TE = 75ms after preprocessing, the second row shows the echo time TE = 95ms after preprocessing, and the third row shows the echo time TE = 115ms after preprocessing; the fourth row shows the echo time TE = 75ms before preprocessing, the fifth row shows the echo time TE = 95ms before preprocessing, and the sixth row shows the echo time TE = 115ms before preprocessing. Blue dots represent standard-compliant slices, and yellow dots represent non-standard-compliant slices. The results shown demonstrate the invention's artifact recognition performance under the above six data processing methods and different scanning parameters. The results show that this invention can accurately identify ghosting artifacts, spike artifacts, and noise artifacts with different scanning parameters, and it has robust recognition capabilities even on unprocessed first-hand data.

[0084] This invention, through the standardized preprocessing method in step two, enables the model to adaptively handle diffusion-weighted magnetic resonance (DMR) images under different b-values, enhancing the model's generalization ability. Step four allows the model to leverage the advantages of unsupervised learning, eliminating the labor-intensive data labeling required in most existing supervised learning methods. This invention does not require training with data containing artifacts, thus theoretically enabling unsupervised anomaly detection for any type of artifact. Steps five and six apply the method of statistical outlier removal to artifact detection for the first time; by calculating the distance from the sample point to the threshold, the probability of the test sample being an artifact sample or a quality-compliant sample can be obtained. In summary, this invention proposes a novel framework for DMR image quality assessment from the perspective of unsupervised learning. This method can significantly reduce the labor-intensive process of training supervised models with numerous labels, assisting radiologists and experts in determining whether the acquired DMR image quality meets clinical standards.

[0085] This method effectively addresses various artifacts across datasets with different scanning parameters and provides the probability of belonging to positive or negative samples, offering a novel auxiliary reference method for practical clinical applications. It explores how to effectively assess and control the quality of diffusion-weighted magnetic resonance imaging (DMRI) images without labeled data, contributing to the advancement of medical image processing technology towards greater intelligence and efficiency, and possessing significant research value. Furthermore, this method provides physicians with more reliable medical image data, thereby improving the accuracy and reliability of medical diagnoses; it can also further refine automated medical image processing workflows, helping to optimize the allocation of medical resources and improve the efficiency of medical services, thus possessing significant social value. Finally, this method will reduce the labor costs of medical image processing, improve processing efficiency, and is expected to significantly enhance the efficiency and cost-effectiveness of workflows in clinical applications; it also has the potential to promote the widespread adoption and use of medical imaging equipment, possessing significant application value.

[0086] The above description is merely an embodiment of this application, and those skilled in the art can make appropriate changes and modifications to the above embodiments. This invention is not limited to the specific embodiments described above, and modifications, equivalent substitutions, and improvements to this invention should also be included within the scope of protection of the claims. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on this invention.

Claims

1. An unsupervised normative learning method for quality assessment of diffusion magnetic resonance imaging, characterized in that, Including the following steps: Step 1: Collect artifact-free diffusion magnetic resonance images and extract cross-sectional midbrain slice images from them, i.e., multi-b value midbrain slice images. Step 2: Based on the attenuation curve of the magnetic resonance signal in the diffusion-sensitive gradient, normalized data preprocessing is performed on the multi-b-value midbrain slice images to obtain normalized preprocessed b-value midbrain slice images, which serve as the training set; Step 2 specifically includes: The attenuation law of magnetic resonance signals in the presence of diffusion-sensitive gradients is as follows: In the formula, Indicates the sensitive diffusion gradient parameter, Indicates differences The signal intensity of the voxel is below the value. This represents the signal strength when b is 0. It indicates the dispersion characteristics of fluid molecules within a region; For the different in the above formula The different signal intensities corresponding to the values ​​are preprocessed using natural logarithm normalization: Natural logarithmic normalization preprocessing was performed on each voxel to obtain normalized preprocessed b-valued midbrain slice images; Step 3: Input the normalized preprocessed b-value midbrain slice image into the VQ-VAE network for reconstruction to obtain the trained VQ-VAE model; Step 4: Input the sample to be tested into the VQ-VAE model obtained in Step 3 to obtain the reconstructed b-value midbrain slice image. Compare the structural similarity index of each multi-b-value midbrain slice image before and after reconstruction, and then determine the threshold. Images above the threshold are images without artifacts and of good quality; images below the threshold are images containing artifacts and of poor quality. For data that conforms to a normal distribution, use The principle is to determine the discrimination threshold; for data that does not conform to a normal distribution, Chebyshev's inequality is used to determine the threshold. Step 5: Based on the threshold in Step 4, compare the distance of the structural similarity index of each slice to the threshold, and then perform a quality assessment to determine the probability that it belongs to the standard image; the probability of conforming to the standard image is: in, This is the scaling factor. This represents the distance of the structural similarity index of each slice to the threshold. ∈[-1,1], in The distance above is positive, in The distance below is negative.

2. The unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment according to claim 1, characterized in that, The diffuse magnetic resonance images were obtained using the SUDMEX dataset.

3. The unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment according to claim 1, characterized in that, The VQ-VAE network uses the following loss function: In the above equation, the quantization loss is calculated using a minimum distance quantizer, which maps each latent variable encoding to the embedding vector with the nearest Euclidean distance. Represents the original image. In representing a reconstructed image, the elements in the encoding are called embedding vectors, denoted as . This represents the gradient stopping operator.

4. The unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment according to claim 1, characterized in that, For data that conforms to a normal distribution, use The principle determines the threshold, that is: In the formula, and Let these represent the mean and variance of the batch of samples to be tested, respectively. This formula indicates that outlier data has a very high probability of falling within the range of... outside.

5. The unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment according to claim 1, characterized in that, For data that does not conform to a normal distribution, the threshold is determined using Chebyshev's inequality, i.e.: In the formula, It is arbitrary The real number.

6. The unsupervised normative learning method for diffusion magnetic resonance imaging quality assessment according to claim 1, characterized in that, The .

7. An unsupervised normative learning system for diffusion magnetic resonance imaging quality assessment that implements the unsupervised normative learning method of any one of claims 1-6, characterized in that, It includes a data acquisition and image processing unit, a data preprocessing unit, a model training unit, a threshold determination unit, and an image evaluation unit, wherein: The image acquisition unit collects artifact-free diffusion magnetic resonance images and extracts cross-sectional midbrain slice images from them, i.e., multi-b-value midbrain slice images. The data preprocessing unit performs normalized data preprocessing on multi-b-value midbrain slice images based on the attenuation curve of the magnetic resonance signal in the diffusion-sensitive gradient, and obtains normalized preprocessed b-value midbrain slice images as the training set. The model training unit inputs the normalized preprocessed b-value midbrain slice image into the VQ-VAE network for reconstruction, thereby obtaining the trained VQ-VAE model; The threshold determination unit inputs the sample to be detected into the VQ-VAE model to obtain the reconstructed b-value midbrain slice image, compares the structural similarity index of each multi-b-value midbrain slice image before and after reconstruction, and then determines the threshold. Images above the threshold are images without artifacts and of acceptable quality; images below the threshold are images containing artifacts and of unacceptable quality. The image evaluation unit compares the distance of the structural similarity index of each slice to the threshold, and then performs a quality assessment to determine the probability that it belongs to a standard image.

Images