Method and device for correcting geometric distortion of MRI image

Abstract

The invention discloses a method and device for correcting geometric distortion of an MRI (Magnetic Resonance Imaging) image, and relates to the field of medical image processing, and the method comprises the steps: obtaining plane echo imaging diffusion-weighted imaging and fast spin echo diffusion-weighted imaging from magnetic resonance imaging equipment; and preprocessing the plane echo diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image. According to the main technical scheme and effects, as geometric distortion can directly influence accurate presentation of an anatomical structure and a lesion area in an MRI image, more real and accurate image data can be provided by effectively correcting geometric distortion in a plane echo imaging diffusion weighted imaging image, so that the diagnosis precision of a doctor is improved, and the diagnosis efficiency is improved. And a doctor is helped to better judge the position, the size and the form of the focus, so that a more reasonable treatment scheme is formulated.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a method and a device for correcting geometric distortion of MRI images. Background Art

[0002] Magnetic resonance imaging (MRI) technology is one of the most widely used imaging technologies in modern medical imaging. MRI technology uses strong magnetic fields and radio frequency pulses to generate detailed images of tissues in the body, and has become an indispensable tool for diagnosing a variety of diseases (especially brain, spine, tumors, etc.). Because MRI images can provide high-resolution tissue details and do not use harmful X-rays, they are widely used in clinical diagnosis.

[0003] However, MRI images, especially echo planar imaging diffusion-weighted imaging (EPI-DWI) images, often encounter geometric distortion problems. This distortion is caused by a variety of factors, including magnetic field inhomogeneity, patient movement, and limitations of the imaging device itself. Geometric distortion can cause the shape, position, and size of anatomical structures in the image to be distorted, which in turn affects the doctor's accurate judgment of the location, size, and morphology of the lesion, reducing the accuracy of diagnosis.

[0004] Current methods based on image processing algorithms, such as model-based correction algorithms, usually assume that image distortion has a certain regularity. However, in reality, the geometric distortion of echo planar imaging and diffusion-weighted imaging images is affected by a combination of factors, and these algorithms are difficult to accurately model and correct all types of distortion. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to solve the technical problem in the prior art that MRI images, especially planar echo imaging diffusion-weighted imaging images, often encounter geometric distortion problems, and to provide a method, device, equipment and medium for correcting geometric distortion of MRI images.

[0006] One or more embodiments of the present specification also relate to a device for correcting geometric distortion of MRI images, an electronic device, a computer-readable storage medium, and a computer program product to solve the technical defects existing in the prior art.

[0007] Technical solution:

[0008] In a first aspect, the present application proposes a method for correcting geometric distortion of MRI images, comprising the steps of:

[0009] Acquire echo-planar imaging diffusion-weighted images and fast spin-echo diffusion-weighted images from magnetic resonance imaging equipment;

[0010] Preprocessing the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method;

[0011] Input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters;

[0012] The initially acquired planar echo imaging diffusion weighted imaging is calibrated by using distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging.

[0013] Preferably, obtaining echo planar imaging diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device includes:

[0014] Diffusion-weighted imaging of echo planar imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion-weighted imaging through an echo planar imaging sequence;

[0015] Fast spin echo diffusion weighted imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion weighted imaging through a fast spin echo sequence.

[0016] Preferably, the training preprocessing method includes image range normalization, random image contrast transformation, random image rotation, random image flipping, image effective area cutting, image resolution resampling, and random image fixed size block cutting;

[0017] Inference preprocessing methods include image range normalization, image effective area cutting, and image resolution resampling.

[0018] Preferably, the preprocessed image is input into a trained distortion parameter prediction model, and the distortion parameter prediction model is trained before obtaining the distortion parameters, including:

[0019] Acquiring historical magnetic resonance imaging data, including echo-planar imaging diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging data;

[0020] Screening and preprocessing of historical MRI data;

[0021] The sorted data is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used to adjust the model parameters and evaluate the model performance during the training process, and the test set is used to finally evaluate the generalization ability and correction effect of the model.

[0022] During the training process, the validation set data is used to evaluate the model regularly, and the model's hyperparameters are adjusted based on the evaluation results;

[0023] When the performance of the model on the validation set reaches the preset threshold or no longer improves, training is stopped and the model parameters at this time are saved.

[0024] Preferably, the preprocessed image is input into a trained distortion parameter prediction model, and the distortion parameter prediction model is trained before obtaining the distortion parameters, including:

[0025] The loss function of model training includes three loss functions: structural similarity index, mutual information, and normalized cross-correlation. After weighting, the final loss function formula is:

[0026] ;

[0027] W1, W2, and W3 are the weight coefficients of structural similarity index, mutual information, and normalized cross-correlation, respectively. SSIM is the structural similarity index loss function, MI is the mutual information loss function, and NCC is the normalized cross-correlation loss function.

[0028] The optimizer for model training uses a combination of adaptive momentum estimation with weight decay optimizer (AdamW) and forward-looking optimization (Lookahead).

[0029] Preferably, the preprocessed image is input into a trained distortion parameter prediction model to obtain the distortion parameters, including:

[0030] Inputting planar echo imaging diffusion weighted imaging and fast spin echo diffusion weighted imaging for preprocessing to obtain preprocessed images;

[0031] Preprocessing the planar echo imaging diffusion weighted imaging to obtain a preprocessed image and preprocessing the fast spin echo diffusion weighted imaging to obtain a preprocessed image for stitching operation;

[0032] The three-scale encoding features are obtained through three sliding window transform modules (Swin Transformer Block) and pooling layers in sequence;

[0033] The encoded features of the three scales are then passed through a decoding module consisting of a residual convolution, a rectified linear unit activation layer (Relu activation layer) and an upsampling layer to obtain the decoded features;

[0034] The three decoded features are concatenated, decoded through a convolutional layer, and upsampled 4 times to obtain the distortion parameters.

[0035] Preferably, the performance test of the trained model also includes an inference phase, which includes:

[0036] The initial echo-planar imaging diffusion-weighted imaging needs to be annotated;

[0037] The trained model is tested using the annotated echo planar imaging diffusion weighted imaging to obtain the corrected echo planar imaging diffusion weighted imaging;

[0038] The error index between the corrected echo planar imaging diffusion weighted imaging and the true echo planar imaging diffusion weighted imaging was calculated.

[0039] Preferably, the initially acquired planar echo imaging diffusion weighted imaging is calibrated by using distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging, including:

[0040] Perform abnormal parameter correction and global Gaussian smoothing on the distortion parameters;

[0041] The planar echo imaging diffusion weighted imaging image is resampled three-dimensionally according to the distortion parameters after smoothing to achieve correction of geometric distortion.

[0042] Preferably, the training preprocessing method and the inference preprocessing method are respectively directed to the training phase and the inference phase of the model.

[0043] In a second aspect, an embodiment of the present invention provides a device for correcting geometric distortion of MRI images, comprising:

[0044] An acquisition module, used for acquiring planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device;

[0045] A data preprocessing module, used for preprocessing the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method;

[0046] A distortion parameter prediction module is used to input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters;

[0047] The geometric distortion calibration module is used to calibrate the initially acquired planar echo imaging diffusion weighted imaging through distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging.

[0048] In a third aspect, an embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory is used to store one or more computer programs; when the one or more computer programs stored in the memory are executed by the processor, the electronic device can implement any possible design method of the first aspect.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a processor, the method as described in any one of the above embodiments is implemented.

[0050] In a fifth aspect, an embodiment of the present invention further provides a computer program product, which, when executed on an electronic device, enables the electronic device to execute any possible design method of any of the above aspects.

[0051] Beneficial effects: The present invention adopts deep learning technology, which can automatically learn the complex features and patterns of MRI images, especially showing strong adaptability in the correction of geometric distortion of planar echo diffusion-weighted imaging images. By using fast spin echo diffusion-weighted imaging (TSE-DWI) images as reference images and combining advanced deep neural network structures (such as convolution, residual convolution, sliding window transformation module, etc.), the present invention can accurately identify and correct various types of geometric distortions, significantly improving the accuracy and consistency of the correction results;

[0052] Compared with the traditional method based on hardware adjustment, the present invention does not rely on expensive hardware equipment adjustment, but realizes image correction through software algorithm, which has significant cost advantages. This method can not only reduce the maintenance cost of MRI equipment, but also reduce the debugging requirements of clinical operators on equipment hardware, thus lowering the technical threshold.

[0053] The present invention uses deep learning technology to correct image distortion, which has strong versatility and adaptability. Whether it is different models of MRI equipment or individual differences of different patients, the fully trained model can effectively adapt and provide consistent and reliable correction effects, which makes the present invention widely applicable in multiple clinical scenarios and has strong promotion value.

[0054] Image processing methods based on deep learning can realize the automatic correction process, reduce manual intervention, and greatly improve processing efficiency. The trained model can complete the correction of newly acquired images in a short time, saving doctors' diagnosis time. This can significantly improve work efficiency and reduce the risk of misdiagnosis and missed diagnosis in clinical applications, especially in emergency and high-load diagnosis and treatment scenarios;

[0055] Since geometric distortion will directly affect the accurate presentation of anatomical structures and lesion areas in MRI images, the present invention can provide more realistic and accurate image data by effectively correcting the geometric distortion in planar echo imaging diffusion-weighted imaging images, thereby improving the doctor's diagnostic accuracy and helping the doctor to better judge the location, size and shape of the lesion, and then formulate a more reasonable treatment plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Provide a schematic diagram of the method framework for the present invention;

[0057] Figure 2Provide an overall flow chart for the present invention;

[0058] Figure 3 A two-dimensional schematic diagram of the principle of a method for cutting out an effective area of ​​an image is provided for the present invention;

[0059] Figure 4 A schematic diagram of the network structure of a distortion parameter prediction model is provided for the present invention;

[0060] Figure 5 A schematic diagram of the training process of the distortion parameter prediction model is provided for the present invention;

[0061] Figure 6 is a schematic diagram of the structure of a device provided by an embodiment of the present application;

[0062] Figure 7 This is a structural block diagram of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the technical solution of the present invention clearer, the present invention is further described in detail below with reference to the specific embodiments of the accompanying drawings.

[0064] Example 1

[0065] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0066] In view of the problems existing in the existing technology, such as Figure 1 As shown, a method for correcting geometric distortion of MRI images comprises the steps of:

[0067] Step S101, acquiring planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device; planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging are two commonly used diffusion weighted imaging methods to obtain images; echo planar (EPI) imaging often has significant distortion in DWI scanning, so it needs to be corrected, and the main purpose of acquiring these two imaging data is to use their characteristics under different imaging modes to reduce distortion effects and improve image quality;

[0068] Step S102, preprocessing the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method; the training preprocessing method is a preprocessing performed by a machine learning model, which may include data standardization, denoising, image alignment, etc., with the purpose of converting input data into a format that can be better used for prediction. The purpose of preprocessing is to provide clearer and more consistent image data for subsequent steps, especially in the process involving machine learning models, the quality of input data has a huge impact on the effect of the model;

[0069] Step S103, input the preprocessed image into a trained distortion parameter prediction model to obtain distortion parameters; input the preprocessed image into a trained distortion parameter prediction model, which is based on a deep learning convolutional neural network (CNN). The distortion parameters refer to the geometric distortion in the image caused by the device or imaging method (such as EPI scanning). The predicted distortion parameters will be provided to the next step for accurate image correction.

[0070] Step S104, calibrating the initially acquired planar echo imaging diffusion weighted imaging by using distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging, and correcting the original planar echo imaging diffusion weighted imaging by using the distortion parameters to obtain a distortion-free image. The key to this step is to use an image processing algorithm (such as affine transformation, nonlinear transformation, etc.) to correct the geometric distortion in the image based on the predicted distortion parameters to ensure that the final image is more accurate and clear.

[0071] In some preferred embodiments, acquiring planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device includes:

[0072] Diffusion-weighted imaging of echo planar imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion-weighted imaging through an echo planar imaging sequence;

[0073] Fast spin echo diffusion weighted imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion weighted imaging through a fast spin echo sequence.

[0074] Specifically, echo-planar imaging (EPI) is a fast magnetic resonance imaging technique, especially suitable for diffusion-weighted imaging (DWI). Its basic principle is to generate high-spatial-resolution images by quickly acquiring signals from multiple slices. When performing diffusion-weighted imaging, EPI can quickly capture the diffusion characteristics of water molecules in tissues.

[0075] Turbo Spin Echo (TSE) is another sequence method used for diffusion-weighted imaging. Compared with planar echo imaging, TSE can achieve higher image quality through longer acquisition time and optimized sequence design. It is particularly advantageous in solving the distortion and artifacts that may occur in planar echo imaging.

[0076] In some preferred embodiments, the training preprocessing method includes image range normalization, random image contrast transformation, random image rotation, random image flipping, image effective area cutting, image resolution resampling, and random image fixed size block cutting;

[0077] Inference preprocessing methods include image range normalization, image effective area cutting, and image resolution resampling.

[0078] Specifically, the training preprocessing method includes the following methods:

[0079] Image range normalization maps the value range of the image to between 0 and 1. This normalization operation can reduce the numerical differences between different images, so that subsequent model training is not affected by the value range of the original image.

[0080] Random transformation of image contrast increases the diversity of images and enhances the generalization ability of the model. By linearly stretching or scaling the pixel values ​​of the image, the contrast of the image is adjusted. This transformation helps to simulate scenes under different image quality conditions.

[0081] Random image rotation enhances the robustness of the model, enabling the model to process images in different directions. The image is rotated three-dimensionally at random angles to simulate rotation transformations in different directions and enhance the transformation invariance of the model.

[0082] Random image flipping further enhances the diversity of training data by randomly flipping images left to right, front to back, and top to bottom. In this way, the symmetry training of images can be increased and the generalization ability of the model can be improved.

[0083] The effective area of ​​the image is cut out to remove irrelevant areas of the image (such as the background), focus on the effective part of the image, find the common foreground area in the fast spin echo diffusion weighted imaging image and the planar echo imaging diffusion weighted imaging image (that is, the important area contained in both), and then crop it according to the minimum circumscribed rectangle of the effective area, which can reduce the interference of background noise on subsequent processing.

[0084] Image resolution resampling unifies the image resolution to a standard spatial resolution. Through trilinear interpolation, the image is resampled and its resolution is adjusted to 1mm*1mm*1mm. This operation ensures that all input images have consistent resolution, which facilitates subsequent model training and reasoning.

[0085] Combination Figure 3 , randomly cut out fixed-size blocks of the image to enhance the diversity of the data and avoid the model from being overly dependent on the overall structure of the image. A fixed-size image block is randomly cut out from the image. If the image block exceeds the boundary of the original image, it is filled with zero values. The purpose of this is to generate more local area data and avoid over-reliance on the global features of the image.

[0086] The inference preprocessing methods include the following:

[0087] The specific method of image range normalization is the same as that in training preprocessing. By normalizing the pixel values ​​of the image to the interval [0,1], the range of image data is standardized during the inference process.

[0088] The operation of cutting out the effective area of ​​the image is the same as that in the training preprocessing. The image is cropped according to the common effective area, and only the meaningful part of the image is retained for subsequent processing.

[0089] Image resolution resampling For images in the inference process, the same resolution standard (1mm*1mm*1mm) as that in the training phase is also used to ensure that the model can obtain a consistent data format during inference.

[0090] In some preferred embodiments, the preprocessed image is input into a trained distortion parameter prediction model, and the distortion parameter prediction model is trained before obtaining the distortion parameters, including:

[0091] Acquiring historical magnetic resonance imaging data, including echo-planar imaging diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging data;

[0092] Screening and preprocessing of historical MRI data;

[0093] The sorted data is divided into a training set, a validation set, and a test set. The training set is used for model training, the validation set is used to adjust the model parameters and evaluate the model performance during the training process, and the test set is used to finally evaluate the generalization ability and correction effect of the model.

[0094] During the training process, the validation set data is used to evaluate the model regularly, and the model's hyperparameters are adjusted based on the evaluation results;

[0095] When the performance of the model on the validation set reaches the preset threshold or no longer improves, training is stopped and the model parameters at this time are saved.

[0096] Specifically, a large amount of existing magnetic resonance imaging data is obtained, including planar echo imaging diffusion-weighted imaging and fast spin echo diffusion-weighted imaging, and the data is screened and preprocessed. Samples that meet the requirements are screened out from the collected MRI images. Images with poor quality, images with excessive distortion, or other images that do not meet the standards need to be eliminated to ensure the quality of the training data;

[0097] Training set: Assign most of the images to the training set. The model will use this data for learning. The training set contains images and their labels (e.g., images with distortion and images after correction). The model will adjust its parameters on this data.

[0098] Validation set: A portion of the original data is extracted as the validation set. The validation set does not participate in the training of the model, but is regularly used to evaluate the performance of the model during the training process. The functions of the validation set are:

[0099] Evaluate the real-time performance of the model;

[0100] Adjust model hyperparameters (such as learning rate, batch size, etc.);

[0101] Check whether the model is overfitting;

[0102] Test set: The test set is part of the data used to ultimately evaluate the generalization ability of the model. The test set does not participate in training and verification. Its main purpose is to evaluate the performance of the model on unknown data.

[0103] During the training process, the validation set data is used regularly to evaluate the performance of the model. According to the performance of the model on the validation set, the model's accuracy, recall rate, loss function and other indicators are evaluated. Hyperparameter adjustment is to adjust the model's hyperparameters through the evaluation results of the validation set. Common hyperparameters include:

[0104] Learning rate: affects the speed at which the model weights are updated. By adjusting the learning rate, the model can be ensured to converge at an appropriate speed.

[0105] Batch size: controls the number of samples used in each training. A smaller batch size usually brings more randomness and may help the model escape from the local optimum.

[0106] Network structure: adjust the network depth, width or convolution kernel size, etc.;

[0107] After each training cycle, the performance on the validation set is evaluated. If the performance of the model on the validation set reaches the preset performance standard (for example, the accuracy exceeds 95%), the training process will be terminated. If the performance of the model on the validation set no longer improves and there is no improvement after multiple consecutive iterations, the "early stopping" strategy can be used to stop training to prevent overfitting. When training stops, the model weights and parameters at this time are saved. The saved model can be used for subsequent reasoning, testing, or deployed in the actual environment for prediction.

[0108] In some preferred embodiments, Figure 5 , input the preprocessed image into the trained distortion parameter prediction model, and train the distortion parameter prediction model before obtaining the distortion parameters, including:

[0109] The loss function of model training includes three loss functions: structural similarity index, mutual information, and normalized cross-correlation. After weighting, the final loss function formula is:

[0110] ;

[0111] W1, W2, and W3 are the weight coefficients of structural similarity index, mutual information, and normalized cross-correlation, respectively. SSIM is the structural similarity index loss function, MI is the mutual information loss function, and NCC is the normalized cross-correlation loss function.

[0112] The optimizer for model training uses a combination of adaptive momentum estimation with weight decay optimizer and forward-looking optimization.

[0113] Specifically, during the training process, the loss function includes structural similarity index, mutual information and normalized cross-correlation for image processing tasks, which are explained as follows: Structural Similarity Index is an indicator used to measure the similarity between two images. It takes into account aspects such as image brightness, contrast and structure. In image reconstruction tasks, the structural similarity index can help evaluate the structural similarity between the restored image and the original image, and is particularly sensitive to the texture and structural details of the image; Mutual Information measures the amount of information shared between two images. It is a classic method for evaluating the similarity between images, especially suitable for image registration and alignment. It is used to evaluate the overlap or information sharing between the restored image and the target image. Normalized cross-correlation Cross-Correlation is a way to measure the similarity between two images. It calculates the normalized cross-correlation of the two images to get their similarity. In image alignment and registration tasks, normalized cross-correlation is a commonly used similarity measure that can help ensure that the model tries to keep the image as close to the target image as possible when restoring it. The three loss functions of structural similarity index, mutual information and normalized cross-correlation are weighted to get a final loss function, where W1, W2, and W3 are the weights of the three loss functions, which determine the importance of each loss function in the final loss calculation. Specifically: W1, W2, W3: The adjustment of these weight values ​​will affect the contribution of each loss function during the training process. For example, if you want the model to pay more attention to the restoration of image structure, you can increase the weight of the structural similarity index; if you want to strengthen the image registration effect, you can increase the weight of the mutual information loss function or the normalized cross-correlation.

[0114] The adaptive momentum estimation with weight decay optimizer is a variant of the Adam optimizer, specifically designed to improve the problem of weight decay (L2 regularization). In the Adam optimizer, weight decay usually affects the gradient calculation when updating parameters, while the adaptive momentum estimation with weight decay optimizer separates weight decay from the gradient update step, which helps to improve the stability and generalization of training. Proactive optimization is an improved optimization strategy designed to improve the performance of the optimizer, especially when using an adaptive optimizer. Proactive optimization maintains the "historical" solutions of multiple optimizers and regularly adjusts the current solution to prevent falling into local optimality, which can improve the stability of the model during training and accelerate convergence.

[0115] In some preferred embodiments, Figure 4 , input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters, including:

[0116] Inputting planar echo imaging diffusion weighted imaging and fast spin echo diffusion weighted imaging for preprocessing to obtain preprocessed images;

[0117] Preprocessing the planar echo imaging diffusion weighted imaging to obtain a preprocessed image and preprocessing the fast spin echo diffusion weighted imaging to obtain a preprocessed image for stitching operation;

[0118] The three-scale encoding features are obtained through three sliding window transformation modules and pooling layers in sequence;

[0119] The encoded features of the three scales are then passed through a decoding module consisting of a residual convolution, a rectified linear unit activation layer, and an upsampling layer to obtain the decoded features;

[0120] The three decoded features are concatenated, decoded through a convolutional layer, and upsampled 4 times to obtain the distortion parameters.

[0121] Specifically, the network accepts two types of MRI data: surface echo imaging diffusion-weighted imaging and fast spin echo diffusion-weighted imaging;

[0122] Preprocess each input image (echo-planar imaging diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging), usually including image normalization, noise removal, size adjustment and other operations. The preprocessed image will be input into the network for subsequent processing;

[0123] Stitching two pre-processed images (echo-planar imaging diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging), which can be horizontal or vertical, is aimed at merging images from two different sources into a complete image input network, so as to utilize the complementary information provided by different imaging methods;

[0124] Sliding window transformation module:

[0125] The spliced ​​images are input into the sliding window transformation module. The sliding window transformation module is based on the structure of the self-attention mechanism, which is specialized in processing the long-distance dependencies in the image and can better capture the global and local information in the image.

[0126] The model uses three sliding window transformation modules, each of which extracts features of different scales. In this way, the network can extract multi-scale information of the image.

[0127] Pooling layer: The output feature map of each sliding window transformation module will pass through the pooling layer. The pooling operation helps to reduce the spatial size and the size of the feature map, thereby reducing the computational complexity and the risk of overfitting;

[0128] The decoding module consists of residual convolution, rectified linear unit activation layer and upsampling layer. Residual convolution helps alleviate the gradient vanishing problem in deep networks, and the rectified linear unit activation function adds nonlinearity, allowing the model to capture more complex patterns.

[0129] The upsampling layer is used to restore the feature map to a higher spatial resolution, so that the model can better restore the details of the original image;

[0130] The three decoded feature maps are concatenated. Through this concatenation, the network can combine the encoded information from different scales to obtain a richer feature representation.

[0131] The concatenated feature map is further decoded through a convolutional layer. The convolution operation helps to extract the final required features:

[0132] The decoded feature map is upsampled by 4 times through the upsampling layer to restore to the required image size;

[0133] Ultimately, the model outputs distortion parameters that are used to correct the distortion of the input image and help restore the original, undistorted image.

[0134] In some preferred embodiments, the performance test of the trained model also includes an inference phase, which includes:

[0135] The initial echo-planar imaging diffusion-weighted imaging needs to be annotated;

[0136] The trained model is tested using the annotated echo planar imaging diffusion weighted imaging to obtain the corrected echo planar imaging diffusion weighted imaging;

[0137] The error index between the corrected echo planar imaging diffusion weighted imaging and the true echo planar imaging diffusion weighted imaging was calculated.

[0138] Specifically, the labeling step refers to manually labeling the initial echo planar imaging diffusion weighted imaging image or automatically labeling it through some method. The goal of labeling is to provide a "target" or "label" for each image, that is, to indicate the true state of the image or the reference image. These labels will be used as true values ​​or target images for comparison with the output of the model during the training process. In this process, images can be manually labeled based on expert knowledge or existing data sets, or pre-trained models can be used to assist in labeling. The task of the model is to use the image's labeling information to correct the distorted part based on the features learned in the training data, and generate an output that is as close to the true undistorted image as possible. After correction, the generated image should have higher quality and less distortion.

[0139] The trained deep learning model is used to test the annotated echo planar imaging diffusion weighted imaging images. The purpose of this step is to verify whether the model can accurately correct the distortion in the image when processing new images, and to calculate the error between the corrected echo planar imaging diffusion weighted imaging and the real echo planar imaging diffusion weighted imaging image, in order to quantify the quality of the correction.

[0140] In some preferred embodiments, the initially acquired echo planar imaging diffusion weighted imaging is calibrated by using distortion parameters to obtain distortion-free echo planar imaging diffusion weighted imaging, including:

[0141] Perform abnormal parameter correction and global Gaussian smoothing on the distortion parameters;

[0142] The planar echo imaging diffusion weighted imaging image is resampled three-dimensionally according to the distortion parameters after smoothing to achieve correction of geometric distortion.

[0143] Specifically, distortion parameter correction usually involves adjusting these parameters so that the geometric distortion of the image is corrected, and some algorithms (such as median filtering or local constraint correction, etc.) are used to smooth or correct outliers in the parameters. Gaussian smoothing is a commonly used image smoothing technique, which is mainly used for denoising and reducing the influence of local noise in the image. Here, Gaussian smoothing is applied to the distortion parameters to reduce the influence of local instability and noise on the distortion correction results. By applying the Gaussian kernel to smooth the distortion parameters, the spatial distribution of the distortion parameters can be made smoother and more consistent. The role of the Gaussian filter is to suppress the changes in the local area of ​​the distortion parameters and make the smoothed parameters more suitable for geometric transformation. The next step is to use these smoothed distortion parameters for three-dimensional resampling of the image, and use the three-dimensional spatial transformation algorithm to spatially adjust the input planar echo imaging diffusion weighted imaging image. Through the resampling algorithm, each pixel in the image can be repositioned so that each point in the image is mapped to its correct position according to the corrected distortion parameters.

[0144] In some preferred embodiments, the training preprocessing method and the inference preprocessing method are respectively directed to the training phase and the inference phase of the model.

[0145] Specifically, the training preprocessing method is used to prepare data in the training phase to adapt it to the needs of the model and improve the training efficiency and generalization ability of the model. The inference preprocessing method is used to process the input data in the inference phase to ensure that the model can accurately predict the input data.

[0146] In some preferred embodiments, Figure 6 , and also proposes a device for correcting geometric distortion of MRI images, comprising:

[0147] An acquisition module 301 is used to acquire planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device;

[0148] A data preprocessing module 302 is used to preprocess the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method;

[0149] The distortion parameter prediction module 303 is used to input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters;

[0150] The geometric distortion calibration module 304 is used to calibrate the initially acquired echo planar imaging diffusion weighted imaging by using distortion parameters to obtain distortion-free echo planar imaging diffusion weighted imaging.

[0151] All relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, and will not be repeated here.

[0152] In some other embodiments of the present invention, an electronic device 400 is disclosed. Figure 7 As shown, the electronic device may include: one or more processors 401; a memory 402; a display 403; one or more applications (not shown); and one or more computer programs 404. The above components may be connected via one or more communication buses 405. The one or more computer programs 404 are stored in the memory 402 and configured to be executed by the one or more processors 401. The one or more computer programs 404 include instructions. The instructions may be used to execute the following: Figures 1 to 5 And each step in the corresponding embodiment.

[0153] Through the description of the above implementation methods, technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device and unit described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0154] Each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0155] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: flash memory, mobile hard disk, read-only memory, random access memory, disk or optical disk and other media that can store program codes.

[0156] The above is only a specific implementation of the embodiment of the present invention, but the protection scope of the embodiment of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiment of the present invention should be included in the protection scope of the embodiment of the present invention. Therefore, the protection scope of the embodiment of the present invention should be based on the protection scope of the claims.

Claims

1. A method for correcting geometric distortion of MRI images, characterized in that: Includes steps: Acquire echo-planar imaging diffusion-weighted images and fast spin-echo diffusion-weighted images from magnetic resonance imaging equipment; Preprocessing the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method; Input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters; The initially acquired planar echo imaging diffusion weighted imaging is calibrated by using distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging.

2. The method according to claim 1, characterized in that Acquire echo-planar diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging from a magnetic resonance imaging device, including: Diffusion-weighted imaging of echo planar imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion-weighted imaging through an echo planar imaging sequence; Fast spin echo diffusion weighted imaging is obtained by configuring a magnetic resonance imaging device and performing diffusion weighted imaging through a fast spin echo sequence.

3. The method according to claim 1, characterized in that The training preprocessing method includes image range normalization, random image contrast transformation, random image rotation, random image flipping, image effective area cutting, image resolution resampling, and random image fixed size block cutting; Inference preprocessing methods include image range normalization, image effective area cutting, and image resolution resampling.

4. The method according to claim 1, characterized in that The preprocessed image is input into the trained distortion parameter prediction model. Before obtaining the distortion parameters, the distortion parameter prediction model is trained, including: Acquiring historical magnetic resonance imaging data, including echo-planar imaging diffusion-weighted imaging and fast spin-echo diffusion-weighted imaging data; Screening and preprocessing of historical MRI data; The sorted data is divided into training set, validation set and test set. The training set is used for model training, the validation set is used to adjust the model parameters and evaluate the model performance during the training process, and the test set is used to finally evaluate the generalization ability and correction effect of the model. During the training process, the validation set data is used to evaluate the model regularly, and the model's hyperparameters are adjusted based on the evaluation results; When the performance of the model on the validation set reaches the preset threshold or no longer improves, training is stopped and the model parameters at this time are saved.

5. The method according to claim 4, characterized in that The preprocessed image is input into the trained distortion parameter prediction model. Before obtaining the distortion parameters, the distortion parameter prediction model is trained, including: The loss function of model training includes three loss functions: structural similarity index, mutual information, and normalized cross-correlation. After weighting, the final loss function formula is: ； W1, W2, and W3 are the weight coefficients of structural similarity index, mutual information, and normalized cross-correlation, respectively. SSIM is the structural similarity index loss function, MI is the mutual information loss function, and NCC is the normalized cross-correlation loss function. The optimizer for model training uses a combination of adaptive momentum estimation with weight decay optimizer and forward-looking optimization.

6. The method according to claim 1, characterized in that Input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters, including: Inputting planar echo imaging diffusion weighted imaging and fast spin echo diffusion weighted imaging for preprocessing to obtain preprocessed images; Preprocessing the planar echo imaging diffusion weighted imaging to obtain a preprocessed image and preprocessing the fast spin echo diffusion weighted imaging to obtain a preprocessed image for stitching operation; The three-scale encoding features are obtained through three sliding window transformation modules and pooling layers in sequence; The encoded features of the three scales are then passed through a decoding module consisting of a residual convolution, a rectified linear unit activation layer, and an upsampling layer to obtain the decoded features; The three decoded features are concatenated, decoded through a convolutional layer, and upsampled 4 times to obtain the distortion parameters.

7. The method according to claim 5, characterized in that The performance test of the trained model also includes the inference phase, which includes: The initial echo-planar imaging diffusion-weighted imaging needs to be annotated; The trained model is tested using the annotated echo planar imaging diffusion weighted imaging to obtain the corrected echo planar imaging diffusion weighted imaging; The error index between the corrected echo planar imaging diffusion weighted imaging and the true echo planar imaging diffusion weighted imaging was calculated.

8. The method according to claim 1, characterized in that The initially acquired echo planar imaging diffusion weighted imaging is calibrated by using distortion parameters to obtain distortion-free echo planar imaging diffusion weighted imaging, including: Perform abnormal parameter correction and global Gaussian smoothing on the distortion parameters; The planar echo imaging diffusion weighted imaging image is resampled three-dimensionally according to the distortion parameters after smoothing to achieve correction of geometric distortion.

9. The method according to claim 7, characterized in that: The training preprocessing method and the inference preprocessing method are respectively aimed at the training phase and the inference phase of the model.

10. A device for correcting geometric distortion of MRI images, characterized in that: include: An acquisition module, used for acquiring planar echo diffusion weighted imaging and fast spin echo diffusion weighted imaging from a magnetic resonance imaging device; A data preprocessing module, used for preprocessing the planar echo imaging diffusion weighted imaging and the fast spin echo diffusion weighted imaging to obtain a preprocessed image, wherein the preprocessing includes a training preprocessing method and an inference preprocessing method; A distortion parameter prediction module is used to input the preprocessed image into the trained distortion parameter prediction model to obtain the distortion parameters; The geometric distortion calibration module is used to calibrate the initially acquired planar echo imaging diffusion weighted imaging through distortion parameters to obtain distortion-free planar echo imaging diffusion weighted imaging.

Images