A signal equalization method for MRI imaging
Through the combination of the two-dimensional Fourier transform algorithm, deep learning model CNN and GAN evaluation model, the problem of signal inhomogeneity and artifacts in MRI imaging is solved, the generation of high-quality images is achieved, and the diagnosis accuracy of MRI imaging in experimental animals is improved. Sex and efficiency.
Patent Information
- Application Number
- CN202410645289.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-23
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-05-23
AI Technical Summary
During MRI imaging, the volume, organ or tissue size, breathing, heartbeat and motion artifacts of laboratory animals, as well as differences in magnetic fields and tissues of individuals lead to uneven signals, resulting in blurred images or ghosting phenomena. Difficult to effectively solve.
Using two-dimensional Fourier transform algorithm, deep learning model CNN, GAN evaluation model and optimization function, artifact recognition, signal inhomogeneity analysis, image correction and quality evaluation are carried out, through noise suppression, uniformity correction, artifact removal and edge Enhanced algorithm to generate high-quality image signals.
It improves the quality of MRI imaging, enhances image clarity and detail visibility, improves diagnosis accuracy and efficiency, and promotes the clinical application of animal models.
Smart Images

Figure CN118628594B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of magnetic resonance imaging, and in particular to a signal equalization method for MRI imaging. Background Art
[0002] MRI imaging technology can conduct in-depth research on the physiology and pathology of laboratory animals and provide a detailed understanding of the anatomical structure of animal models, which is crucial for the creation and research of disease models. It can also study changes in animal models under specific physiological or pathological conditions through dynamic and static imaging, track disease progression in genetically modified animal models, and reveal the molecular mechanisms of the disease. MRI imaging technology is of great benefit to the study and understanding of the physiological and pathological states of experimental animals. It can identify and quantify lesions in animal models, and promote medical research and the development of new therapies.
[0003] For example, application number CN202280018914.3 discloses an equalization technology for magnetic resonance imaging (MRI) signals, which includes a magnetic resonance system and method. By improving the performance of the MRI coil and the collaboration of the calibration antenna, an equalization filter is provided in the electronic signal processing chain to equalize the MR channel on the MR frequency band to solve the problem of insensitivity of the MRI imaging system caused by uneven excitation of the radio frequency (RF) coil. However, when performing MRI imaging on animals of different species in the laboratory, the experimental animals to be tested are affected by the volume, organ or tissue size, the living body's breathing, heartbeat, any limb movement motion artifacts, and even the individual magnetic field of the experimental animals and the difference in magnetic susceptibility between different tissues. Therefore, blurred images or ghosting phenomena will be generated due to signal unevenness during the MRI imaging process.
[0004] The existing technology has the following shortcomings: when performing MRI imaging on laboratory animals, in addition to improving the uniformity and strength of the magnetic field by improving the MRI equipment, using ablators, or improving the coil design suitable for use with experimental animals, there are also problems such as signal unevenness, artifacts, low recognition rate, and unclear images caused by artifacts caused by the different types of laboratory animals, the environment of the breeding space, and animal behavior, as well as analysis and correction defects in image processing technology.
[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention
[0006] The purpose of the present invention is to provide a signal equalization method for MRI imaging. The present invention solves the problems in the above-mentioned background technology by adopting a two-dimensional Fourier transform algorithm, a deep learning model CNN, a GAN evaluation model and an optimization function.
[0007] In order to achieve the above object, the present invention provides the following technical solution: a signal equalization method for MRI imaging, comprising the following steps:
[0008] S1. Pretreatment: Establish an animal model and provide an imaging environment. Adjust the scanning parameters of the MRI equipment according to the imaging site of the experimental animal, including magnetic field intensity, gradient pulse intensity, and sequence parameters;
[0009] S2. Scan and collect data: Start the pre-processed MRI equipment and scan the animal model. Receive the high-sensitivity coil adapted to the experimental animal to collect MRI data. Apply the two-dimensional Fourier transform algorithm to the MRI data for image reconstruction to obtain the image signal to be evaluated.
[0010] S3, Artifact Identification and Signal Inhomogeneity Analysis: The deep learning model CNN is used to detect abnormal signals such as signal inhomogeneity and artifacts in the image to be evaluated;
[0011] S4. Image correction processing: Use noise suppression, uniformity correction, artifact removal, and edge enhancement algorithms to perform image correction and multimodal fusion processing on artifacts identified in the image to be evaluated and signal inhomogeneity analyzed to obtain high-quality image signals;
[0012] S5. Image quality assessment: Build a generative adversarial network (GAN) evaluation model to evaluate the image clarity and overall signal balance parameters of high-quality images. The parameters include clarity, contrast, signal balance, and artifacts, and generate quality assessment coefficients.
[0013] S6. Feedback monitoring and model optimization: Compare the quantitative indicators SNR and CNR based on the quality assessment coefficient, and feed them back to the artifact identification and signal inhomogeneity analysis steps for continuous optimization. Iterate the image correction and quality assessment processing to obtain the final MRI image signal.
[0014] Optionally, the steps of establishing an animal model and providing an imaging environment are as follows:
[0015] According to the pathological requirements of experimental animals for MRI imaging research, select the tissues, organs or body parts of the experimental animals, use anesthetics, sedatives or experimental animal positioning devices to keep the experimental animals still and comfortable, without physiological stress reactions and in a stable state of physical signs, and then perform magnetic resonance imaging examinations, and perform pathological analysis of the experimental animals by identifying the imaging images;
[0016] The imaging environment was the same as the temperature, humidity, air pressure, air flow rate and audio space in the living environment of the experimental animals.
[0017] Optionally, the steps of collecting the MRI data are as follows:
[0018] During an MRI scan, the magnetic field and radio frequency (RF) pulses excite the hydrogen nuclei in the body. When the hydrogen nuclei return to equilibrium, they release a signal that is MRI data and is calibrated as S(u,v). n , and is captured by the receiving coil and stored in the parameter space of k-space. K-space is a frequency space that includes low-frequency information and high-frequency information of the image in the MRI data. (u, v) represents a point data in k-space, n represents N points in k-space, S(u, v) n It is represented as the set of all MRI data stored in k-space;
[0019] Before image reconstruction, the MRI data stored in the k-space must be corrected and filled.
[0020] Optionally, the steps of performing image reconstruction using the two-dimensional Fourier transform algorithm are as follows:
[0021] The MRI data S(u,v) after correction and filling is n , MRI data are converted into image space using a two-dimensional Fourier transform algorithm;
[0022] Then the conversion formula of the two-dimensional Fourier transform algorithm is Where F(x,y) represents the data of a point (u,v) in k-space that is transformed into the corresponding point in the image space through two-dimensional Fourier transform, f(u,v) represents the data of a point (u,v) in k-space, u and v represent frequency variables, x and y represent spatial variables, j represents the imaginary unit, ∫ represents the integration symbol, and -∞ and +∞ represent the integration area;
[0023] The MRI data S(u,v) n The complex image transformed one by one into the image space is F(x,y) n , then the complex image F(x,y) n Including amplitude and phase, that is, the image signal to be evaluated is F(x,y) n .
[0024] Optionally, the logical steps of the deep learning model CNN to detect abnormal signals are as follows:
[0025] Collect several image signals F(x,y) to be evaluated n , and the image signal to be evaluated F(x,y) n The data set is formed after normalization processing, which includes images of normal signals and abnormal images containing inhomogeneity and artifacts;
[0026] Construct a CNN model, and use the divided training set in the input data set for training. The convolutional layer in the CNN is used to evaluate the input image signal F(x,y) n For feature extraction, the convolution calculation formula is Where T(i,j) l Represented as the feature map output of layer l at position (i, j), I(x,y) l-1 Represented as the input image or feature map of the l-1 layer, K(ix,jy) l Indicates the size of the convolution kernel, b l Represented as a bias constant, σ represents an activation function that can map the input feature to a value between 0 and 1;
[0027] The pooling layer in CNN performs pooling processing to reduce the dimension of the feature map and retain abnormal information. The pooling calculation formula is P(i,j)=max(x,y)×σ(s+x,s+y)I(x,y), where P(i,j) represents the output after pooling, max(x,y) represents the maximum value, σ represents the activation function, s represents the window step size, and I(x,y) represents the input value in the pooling window.
[0028] Optionally, the fully connected layer in the CNN maps the trained distributed feature representation to the sample label space and outputs the prediction result. The calculation formula of the fully connected layer is Q(i,j)=w x,y ×I′(x,y)+b0, where Q(i,j) represents the output vector, w x,y It is represented as a weight matrix, I′(x,y) is represented as the input vector of the feature map, and b0 is represented as the bias vector;
[0029] During CNN training, the cross entropy loss function is introduced for evaluation based on the difference between the output value and the actual label. The calculation formula of the cross entropy loss function is: In the formula, Loss is the cross entropy loss value, Q is the true label, and Q k It represents the predicted label output by the model, n represents the number of samples, and p(Q k ) The probability distribution of the model prediction for the kth sample;
[0030] According to the loss function value Loss, the weights and biases in the network are updated through the back propagation algorithm to reduce the prediction error. The model performance is evaluated using the validation set, and the hyperparameters are adjusted to optimize the model. The gradient descent algorithm is used to minimize the loss function. The calculation formula for updating the weight in the gradient descent algorithm is W(i,j) l+1 =W(i,j) l -η×Loss, where W(i,j) l+1Represented as the weight value adjusted by the gradient descent algorithm, W(i,j) l It represents the weight value of the current layer l, and η represents the learning rate;
[0031] After the model training is completed and fully evaluated and tuned, the test set divided in the dataset is used to test the generalization ability of the CNN model, that is, the ability to identify and detect abnormal signals such as signal inhomogeneity and artifacts.
[0032] Optionally, the steps of image correction and multimodal fusion processing are as follows:
[0033] First, the image signal F(x,y) to be evaluated n Perform noise suppression processing to generate a noise suppression signal, calibrated as G(x′,y′), and use the noise suppression Gaussian filter algorithm to process it, that is, the calculation formula is G(x′,y′)=F(x,y) n ×K(x′,y′), and K Where F(x,y) n Denotes the input image, K(x′, y′) denotes the Gaussian kernel, ξ denotes the Gaussian width, i and j denote the coordinates after the convolution layer operation, and G(x′, y′) denotes the output image after a noise suppression signal.
[0034] Then, the uniformity correction of the primary noise suppression signal G(x′, y′) is performed to generate a secondary correction signal, which is calibrated as Fu(x′, y′). The surface fitting method is used to process the primary noise suppression signal G(x′, y′). The integral calculation formula is: Get the feature map Fe(x,y), then perform surface fitting on the feature map Fe(x,y), then the fitting calculation formula is Gu(x,y)=∑Fe(x,y)×W(xi,yj), get the estimated image Gu(x,y), and add the estimated image Gu(x,y) back to the original image, then the calculation formula of the output image is Fu(x′,y′)=Gu(x,y)+G(x′,y′), get the quadratic correction signal Fu(x′,y′), where ψ represents the control parameter of the neighborhood, W(xi,yj) represents the surface fitting operator, xi and yj represent the distance in the horizontal and vertical directions in the two-dimensional image respectively;
[0035] The quadratic corrected signal Fu(x′,y′) is then used to remove artifacts from the image to generate a tertiary de-artifacted signal, which is calibrated as M(x,y) and processed using the median filter method, i.e., M(x,y) = median[Fu(x′-k0,y′-l0)], where median represents the median, [Fu(x′-k0,y′-l0)] represents the intensity value of the quadratic corrected signal Fu(x′,y′) at position (x′,y′), and k0 and l0 represent the size of the neighborhood.
[0036] Secondly, the image after the three-time artifact removal signal M(x,y) is edge enhanced to generate a four-time edge enhancement signal, calibrated as L(x,y), and the Laplace filter is used to enhance the edge of the image, that is, Where, Expressed as the Laplace operator;
[0037] Finally, the image processed by the four-time edge enhancement signal L(x,y) is subjected to multimodal fusion to generate the final high-quality image signal, calibrated as D(x,y), and the image features are fused using the weighted average method, that is, Where W i0 It is represented as the weight of the i0th modality, and n0 is represented as the number of modalities that need to be fused.
[0038] Optionally, the steps for establishing the adversarial network GAN evaluation model are as follows:
[0039] Collect the image signal F(x,y) to be evaluated n And a data training set of high-quality image signals D(x,y), and a GAN model is constructed, including a generator and a discriminator;
[0040] Define the loss function for the GAN model, and use the peak signal-to-noise ratio (PSNR) loss function.
[0041] Use the data training set to train the GAN model. During the training process, the generator and discriminator will be optimized alternately;
[0042] Use an independent test set to evaluate model performance and make necessary adjustments to the GAN model based on the evaluation results;
[0043] After the GAN model training is completed, the discriminator is used to provide quality assessment for the new image, and the output evaluation result is used as the quality assessment coefficient.
[0044] Optionally, the logic for obtaining the quality assessment coefficient is as follows:
[0045] According to the image signal F(x,y) to be evaluated n And high-quality image signal D(x,y), define the peak signal-to-noise ratio PSNR loss function, then the loss function formula is and Where L(MSE) is the difference between the predicted high-quality image signal D(x,y) and the real image signal to be evaluated F(x,y) n The mean square error between them, N represents the number of samples, L(PSNR) represents the peak signal-to-noise ratio, MAX I Expressed as the maximum pixel value of the image;
[0046] Determine the GAN objective function to reflect the adversarial process between the generator and the discriminator. The objective function calculation formula is f(S,P)=min{S}×E(lnP(D(x,y)))×(D(x,y))+min{P,Z}×E(ln(1-P(S(Z)))), where f(S,P) represents the objective function, S and P represent the generator and discriminator respectively, D(x,y) represents the real image of the input high-quality image signal, Z represents the input noise of the generator, and E represents the expected value;
[0047] The quality assessment coefficient is the true output probability of the discriminator for the input image, that is, the calculation formula of the quality assessment coefficient is δ(QEC)=100P(D(x,y)), and P(D(x,y))∈[0,1], where δ(QEC) represents the quality assessment coefficient after the discriminator output mapping, and P(D(x,y)) represents the true output probability of the discriminator for the input image.
[0048] Optionally, the steps of comparing the quality assessment coefficient with the quantitative indicators SNR and CNR are as follows:
[0049] The linear regression analysis method was used to compare the quality evaluation coefficient δ(QEC) with the quantitative indicators SNR and CNR to analyze the influence of the quantitative indicators SNR and CNR on the quality evaluation coefficient δ(QEC). The calculation formula of the linear regression analysis method is δ(QEC) = β0 + β1 × SNR + β2 × CNR + ε0, where β0 represents the intercept term, ε0 represents the error term, and β1 and β2 represent the slopes of the quantitative indicators SNR and CNR, respectively.
[0050] According to the quality evaluation coefficient δ(QEC), quantization index SNR and CNR, an optimization function is defined and calibrated as g(λ), which is used to iteratively calculate and adjust the CNN model parameters to improve image quality. The calculation formula of the optimization function is: Where t represents the number of iterations, It is represented as the sum of the number of iterations T, λ is the adjustment factor, Δ is the actual value of the current quality evaluation coefficient δ(QEC), quantization indicators SNR and CNR, and g(λ) is the improvement amount of the current optimal value.
[0051] In the above technical solution, the technical effects and advantages provided by the present invention are:
[0052] The present invention uses a two-dimensional Fourier transform algorithm and a deep learning model CNN method to perform artifact recognition and analyze abnormal signals with uneven signals on MRI data collected by magnetic resonance imaging equipment during scanning of experimental animals. It also uses noise suppression, uniformity correction, artifact removal, and edge enhancement algorithms to perform image correction and multimodal fusion processing on the image signals to be evaluated. This improves the quality of MRI imaging, speeds up the processing speed of recognition and analysis, provides a more comprehensive diagnostic view of experimental animals, and promotes the development of research and clinical applications.
[0053] The present invention also uses the GAN evaluation model and optimization function to further equalize the signal and generate high-quality images, realizes the function of adjusting model parameters to improve image quality, enhances image clarity and visibility of details, and achieves the diversity effect of data set training, which helps to train more robust diagnostic models and improves the accuracy and efficiency of clinical diagnosis in animal models. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 This is a flow chart of the signal equalization method for MRI imaging of the present invention. DETAILED DESCRIPTION
[0055] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.
[0056] The present invention provides Figure 1 A signal equalization method for MRI imaging is shown, comprising the following steps:
[0057] S1. Pretreatment: Establish an animal model and provide an imaging environment. Adjust the scanning parameters of the MRI equipment according to the imaging site of the experimental animal, including magnetic field intensity, gradient pulse intensity, and sequence parameters;
[0058] Specifically, the steps for establishing an animal model and providing an imaging environment are as follows:
[0059] According to the pathological requirements of experimental animals for MRI imaging research, select the tissues, organs or body parts of the experimental animals, use anesthetics, sedatives or experimental animal positioning devices to keep the experimental animals still and comfortable, without physiological stress reactions and in a stable state of physical signs, and then perform magnetic resonance imaging examinations, and perform pathological analysis of the experimental animals by identifying the imaging images;
[0060] The imaging environment is the same as the experimental animals' living environment in terms of temperature, humidity, air pressure, air flow rate, and audio space to reduce the animals' physiological stress response.
[0061] The scanning parameters of MRI equipment include magnetic field strength, gradient pulse strength and sequence. Among them, the magnetic field strength is related to the signal strength and image resolution. Therefore, a high magnetic field strength of 3T-7T is selected according to the imaging part and body size of the experimental animal; the gradient pulse strength and sequence are related to the image resolution and scanning time, so the pulse repetition time, bandwidth, radio frequency pulse length, radio frequency RF pulse sequence, pulse sequence schedule, and uneven static magnetic field and radio frequency field are adjusted.
[0062] Specifically, the steps for adjusting the scanning parameters of the MRI device are as follows:
[0063] Determine the volume of the imaging site of the experimental animal and select an appropriate voxel size, ranging from 50 to 200 μm, and apply a T1-weighted MRI pulse sequence. At the same time, set the echo time TE and repetition time TR of the T1-weighted MRI pulse sequence;
[0064] Increase the number of averages and apply presaturation pulses and spectral fat suppression;
[0065] We customize coils to suit the specific animal size and imaging area, and synchronize the breathing and heartbeat monitoring equipment of the experimental animals with the MRI equipment scan.
[0066] S2. Scan and collect data: Start the pre-processed MRI equipment and scan the animal model. Receive the high-sensitivity coil adapted to the experimental animal to collect MRI data. Apply the two-dimensional Fourier transform algorithm to the MRI data for image reconstruction to obtain the image signal to be evaluated.
[0067] Specifically, the steps for collecting MRI data are as follows:
[0068] During an MRI scan, the magnetic field and radio frequency (RF) pulses excite the hydrogen nuclei in the body. When the hydrogen nuclei return to equilibrium, they release a signal that is MRI data and is calibrated as S(u,v). n , and is captured by the receiving coil and stored in the parameter space of k-space. K-space is a frequency space that includes low-frequency information and high-frequency information of the image in the MRI data. (u, v) represents a point data in k-space, n represents N points in k-space, S(u, v) n It is represented as the set of all MRI data stored in k-space;
[0069] Before image reconstruction, the MRI data stored in the k-space must be corrected and filled. The correction process includes field inhomogeneity correction and removal of k-space centerline artifacts. Field inhomogeneity correction uses a polynomial fitting algorithm to solve the image brightness changes caused by magnetic field inhomogeneity. Removal of k-space centerline artifacts uses a low-pass filter to reduce sampling artifacts. Filling processing uses linear interpolation to fill in missing data to improve image quality.
[0070] Specifically, the steps of image reconstruction using the two-dimensional Fourier transform algorithm are as follows:
[0071] The MRI data S(u,v) after correction and filling is n , MRI data are converted into image space using a two-dimensional Fourier transform algorithm;
[0072] Then the conversion formula of the two-dimensional Fourier transform algorithm is Where F(x,y) represents the data of a point (u,v) in k-space that is transformed into the corresponding point in the image space through two-dimensional Fourier transform, f(u,v) represents the data of a point (u,v) in k-space, u and v represent frequency variables, x and y represent spatial variables, j represents the imaginary unit, ∫ represents the integration symbol, and -∞ and +∞ represent the integration area;
[0073] The MRI data S(u,v) n The complex image transformed one by one into the image space is F(x,y) n , then the complex image F(x,y) n Including amplitude and phase, that is, the image signal to be evaluated is F(x,y) n .
[0074] S3, Artifact Identification and Signal Inhomogeneity Analysis: The deep learning model CNN is used to detect abnormal signals such as signal inhomogeneity and artifacts in the image to be evaluated;
[0075] Specifically, the logical steps of the deep learning model CNN to detect abnormal signals are as follows:
[0076] Collect several image signals F(x,y) to be evaluated n , and the image signal to be evaluated F(x,y) n The data set is formed after normalization processing, which includes images of normal signals and abnormal images containing inhomogeneity and artifacts;
[0077] Construct a CNN model, and use the divided training set in the input data set for training. The convolutional layer in the CNN is used to evaluate the input image signal F(x,y) nFor feature extraction, the convolution calculation formula is Where T(i,j) l Represented as the feature map output of layer l at position (i, j), I(x,y) l-1 Represented as the input image or feature map of the l-1 layer, K(ix,jy) l Indicates the size of the convolution kernel, b l Represented as a bias constant, σ represents an activation function that can map the input feature to a value between 0 and 1;
[0078] The pooling layer in CNN performs pooling processing to reduce the dimension of the feature map and retain abnormal information. The pooling calculation formula is P(i,j)=max(x,y)×σ(s+x,s+y)I(x,y), where P(i,j) represents the output after pooling, max(x,y) represents the maximum value, σ represents the activation function, s represents the window step size, and I(x,y) represents the input value in the pooling window.
[0079] The fully connected layer in CNN maps the distributed feature representation obtained through training to the sample label space and outputs the prediction result. The calculation formula of the fully connected layer is Q(i,j)=w x,y ×I′(x,y)+b0, where Q(i,j) represents the output vector, w x,y It is represented as a weight matrix, I′(x,y) is represented as the input vector of the feature map, and b0 is represented as the bias vector;
[0080] During CNN training, the cross entropy loss function is introduced for evaluation based on the difference between the output value and the actual label. The calculation formula of the cross entropy loss function is: In the formula, Loss is the cross entropy loss value, Q is the true label, and Q k It represents the predicted label output by the model, n represents the number of samples, and p(Q k ) The probability distribution of the model prediction for the kth sample;
[0081] According to the loss function value Loss, the weights and biases in the network are updated through the back propagation algorithm to reduce the prediction error. The model performance is evaluated using the validation set, and the hyperparameters are adjusted to optimize the model. The gradient descent algorithm is used to minimize the loss function. The calculation formula for updating the weight in the gradient descent algorithm is W(i,j) l+1 =W(i,j) l -η×Loss, where W(i,j) l+1 Represented as the weight value adjusted by the gradient descent algorithm, W(i,j) l It represents the weight value of the current layer l, and η represents the learning rate;
[0082] After the model training is completed and fully evaluated and tuned, the test set divided in the dataset is used to test the generalization ability of the CNN model, that is, the ability to identify and detect abnormal signals such as signal inhomogeneity and artifacts, for anomaly detection in the image to be evaluated.
[0083] S4. Image correction processing: Use noise suppression, uniformity correction, artifact removal, and edge enhancement algorithms to perform image correction and multimodal fusion processing on artifacts identified in the image to be evaluated and signal inhomogeneity analyzed to obtain high-quality image signals;
[0084] Specifically, the steps of image correction and multimodal fusion processing are as follows:
[0085] First, the image signal F(x,y) to be evaluated n Noise suppression is performed to generate a noise suppression signal, which is calibrated as G(x′,y′) to improve image quality and prepare for subsequent processing. The noise suppression Gaussian filtering algorithm is used for processing, that is, the calculation formula is G(x′,y′)=F(x,y) n ×K(x′,y′), and Where F(x,y) n Denotes the input image, K(x′, y′) denotes the Gaussian kernel, ξ denotes the Gaussian width, i and j denote the coordinates after the convolution layer operation, and G(x′, y′) denotes the output image after a noise suppression signal.
[0086] Then, the uniformity correction of the primary noise suppression signal G(x′, y′) is performed to generate a secondary correction signal, which is calibrated as Fu(x′, y′) to eliminate the uneven brightness caused by uneven illumination during the scanning process of the MRI equipment. The surface fitting method is used to process the primary noise suppression signal G(x′, y′). The integral calculation formula is: Get the feature map Fe(x,y), then perform surface fitting on the feature map Fe(x,y), then the fitting calculation formula is Gu(x,y)=∑Fe(x,y)×W(xi,yj), get the estimated image Gu(x,y), and add the estimated image Gu(x,y) back to the original image, then the calculation formula of the output image is Fu(x′,y′)=Gu(x,y)+G(x′,y′), get the quadratic correction signal Fu(x′,y′), where ψ represents the control parameter of the neighborhood, W(xi,yj) represents the surface fitting operator, xi and yj represent the distance in the horizontal and vertical directions in the two-dimensional image respectively;
[0087] The quadratic correction signal Fu(x′,y′) is then used to remove artifacts from the image to generate a tertiary artifact removal signal, which is calibrated as M(x,y). This signal is used to correct equipment defects or eliminate artifacts caused by the movement of experimental animals during scanning. The median filter method is used, that is, M(x,y) = median[Fu(x′-k0,y′-l0)], where median represents the median value, [Fu(x′-k0,y′-l0)] represents the intensity value of the quadratic correction signal Fu(x′,y′) at position (x′,y′), and k0 and l0 represent the size of the neighborhood.
[0088] Secondly, the image after the three-time artifact removal signal M(x,y) is edge enhanced to generate a four-time edge enhancement signal, which is calibrated as L(x,y) to improve the image clarity and edge recognition ability. The Laplace filter is used to enhance the edge of the image, that is, Where, Expressed as the Laplace operator;
[0089] Finally, the image processed by the four edge enhancement signals L(x, y) is multimodally fused to generate the final high-quality image signal, which is calibrated as D(x, y) to obtain richer information and improve image quality. The weighted average method is used to fuse image features, that is, Where, It is represented as the weight of the i0th modality, and n0 is represented as the number of modalities that need to be fused.
[0090] S5. Image quality assessment: Build a generative adversarial network (GAN) evaluation model to evaluate the image clarity and overall signal balance parameters of high-quality images. The parameters include clarity, contrast, signal balance, and artifacts, and generate quality assessment coefficients.
[0091] Specifically, the steps for establishing the adversarial network GAN evaluation model are as follows:
[0092] Collect the image signal F(x,y) to be evaluated n A data training set of high-quality image signals D(x,y) is used to build a GAN model, which includes a generator and a discriminator. The generator generates images and the discriminator is responsible for evaluating the quality of the images.
[0093] Define a loss function for the GAN model to evaluate the discriminator's ability to distinguish real and generated images, and the generator's ability to produce high-quality images. The loss function uses the peak signal-to-noise ratio (PSNR) loss function.
[0094] The GAN model is trained with a data training set. During the training process, the generator and discriminator are optimized alternately, where the generator tries to generate more and more realistic images, while the discriminator strives to distinguish between real images and generated images more accurately.
[0095] Use an independent test set to evaluate model performance and make necessary adjustments to the GAN model based on the evaluation results to improve the accuracy and reliability of the evaluation;
[0096] After the GAN model training is completed, the discriminator is used to provide quality assessment for the new image, and the output evaluation result is used as the quality assessment coefficient.
[0097] Specifically, the logic for obtaining the quality assessment coefficient is as follows:
[0098] According to the image signal F(x,y) to be evaluated n And high-quality image signal D(x,y), define the peak signal-to-noise ratio PSNR loss function, then the loss function formula is and Where L(MSE) is the difference between the predicted high-quality image signal D(x,y) and the real image signal to be evaluated F(x,y) n The mean square error between them, N represents the number of samples, L(PSNR) represents the peak signal-to-noise ratio, MAX I Expressed as the maximum pixel value of the image;
[0099] Determine the GAN objective function to reflect the adversarial process between the generator and the discriminator. The objective function calculation formula is f(S,P)=min{S}×E(lnP(D(x,y)))×(D(x,y))+min{P,Z}×E(ln(1-P(S(Z)))), where f(S,P) represents the objective function, S and P represent the generator and discriminator respectively, D(x,y) represents the real image of the input high-quality image signal, Z represents the input noise of the generator, and E represents the expected value;
[0100] The quality assessment coefficient is the true output probability of the discriminator for the input image, that is, the calculation formula of the quality assessment coefficient is δ(QEC)=100P(D(x,y)), and P(D(x,y))∈[0,1], where δ(QEC) represents the quality assessment coefficient after the discriminator output mapping, and P(D(x,y)) represents the true output probability of the discriminator for the input image. When the value of δ(QEC) approaches 100, the clarity, contrast, signal balance and artifact quality of the output image are better.
[0101] S6. Feedback monitoring and model optimization: Compare the quantitative indicators SNR and CNR based on the quality assessment coefficient, and feed them back to the artifact identification and signal inhomogeneity analysis steps for continuous optimization. Iterate the image correction and quality assessment processing to obtain the final MRI image signal.
[0102] Specifically, the steps for comparing the quality assessment coefficients to the quantitative indicators SNR and CNR are as follows:
[0103] The linear regression analysis method was used to compare the quality evaluation coefficient δ(QEC) with the quantitative indicators SNR and CNR to analyze the degree of influence of the quantitative indicators SNR and CNR on the quality evaluation coefficient δ(QEC), that is, to analyze the degree of influence on the quality of MRI images. The calculation formula of the linear regression analysis method is δ(QEC) = β0 + β1 × SNR + β2 × CNR + ε0, where β0 represents the intercept term, ε0 represents the error term, and β1 and β2 represent the slopes of the quantitative indicators SNR and CNR, respectively.
[0104] According to the quality evaluation coefficient δ(QEC), quantization index SNR and CNR, an optimization function is defined and calibrated as g(λ), which is used to iteratively calculate and adjust the CNN model parameters to improve image quality. The calculation formula of the optimization function is: Where t represents the number of iterations, It is expressed as the sum of the number of iterations T, λ is the adjustment factor, Δ is the actual value of the current quality evaluation coefficient δ(QEC), the quantization indicators SNR and CNR, and g(λ) is the improvement amount of the current optimal value, which is used to balance the optimization of the quality evaluation coefficient δ(QEC) and maintain the quantization indicators SNR and CNR within the ideal range.
[0105] An embodiment of the present invention provides a signal equalization method for MRI imaging. The specific method and process are detailed in the embodiment of the signal equalization method for MRI imaging, which will not be repeated here.
[0106] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
Claims
1. A signal equalization method for MRI imaging, characterized in that: The steps include: S1. Pretreatment: Establish an animal model and provide an imaging environment, and adjust the scanning parameters of the MRI equipment according to the imaging site of the experimental animal, including magnetic field intensity, gradient pulse intensity, and sequence parameters; S2. Scan and collect data: Start the pre-processed MRI device and scan the animal model, receive the high-sensitivity coil adapted to the experimental animal to collect MRI data, and apply the two-dimensional Fourier transform algorithm to the MRI data for image reconstruction to obtain the image signal to be evaluated; S3, Artifact identification and signal inhomogeneity analysis: The deep learning model CNN is used to detect the signal inhomogeneity and abnormal signals of artifacts in the image to be evaluated; The logical steps of the deep learning model CNN to detect abnormal signals are as follows: Collect several image signals F(x,y) to be evaluated n , and the image signal to be evaluated F(x,y) n It contains images of normal signals and abnormal images containing inhomogeneity and artifacts, and forms a data set after normalization processing; Construct a CNN model, and use the divided training set in the input data set for training. The convolutional layer in the CNN evaluates the input image signal F(x,y) n For feature extraction, the convolution calculation formula is Where T(i,j) l Represented as the feature map output of the lth layer at position (i, j), I(x, y) l-1 Represented as the input image or feature map of the l-1th layer, K(ix,jy) l Represents the size of the convolution kernel, b l Represented as a bias constant, σ represents an activation function that can map the input feature to a value between 0 and 1; The pooling layer in CNN performs pooling processing to reduce the dimension of the feature map and retain abnormal information. The pooling calculation formula is P(i,j)=max(x,y)×σ(s+x,s+y)I(x,y), where P(i,j) represents the output after pooling, max(x,y) represents the maximum value, σ represents the activation function, s represents the window step size, and I(x,y) represents the input value in the pooling window; The fully connected layer in CNN maps the distributed feature representation obtained through training to the sample label space and outputs the prediction result. The calculation formula of the fully connected layer is Q(i,j)=w x,y ×I′(x,y)+b0, where Q(i,j) represents the output vector, w x,y It is represented as a weight matrix, I′(x,y) is represented as the input vector of the feature map, and b0 is represented as the bias vector; During CNN training, the cross entropy loss function is introduced for evaluation based on the difference between the output value and the actual label. The calculation formula of the cross entropy loss function is: In the formula, Loss is the cross entropy loss value, Q is the true label, and Q k represents the predicted label output by the model, n represents the number of samples, and p(Q k ) The probability distribution predicted by the model for the kth sample; According to the loss function value Loss, the weights and biases in the network are updated through the back propagation algorithm to reduce the prediction error. The model performance is evaluated using the validation set, and the hyperparameters are adjusted to optimize the model. The gradient descent algorithm is used to minimize the loss function. The calculation formula for updating the weight in the gradient descent algorithm is W(i,j) l+1 =W(i,j) l -η×Loss, where W(i,j) l+1 Represented as the weight value adjusted by the gradient descent algorithm, W(i,j) l It is represented by the weight value of the current layer l, and η is represented by the learning rate; After the model training is completed and fully evaluated and tuned, the generalization ability of the CNN model is tested using the test set divided in the dataset, that is, the ability to identify and detect abnormal signals such as signal inhomogeneity and artifacts; S4, Image correction processing: Use noise suppression, uniformity correction, artifact removal and edge enhancement algorithms to perform image correction and multimodal fusion processing on the artifacts identified in the image to be evaluated and the signal inhomogeneity analyzed to obtain high-quality image signals; The steps of image correction and multimodal fusion processing are as follows: First, the image signal F(x,y) to be evaluated n Noise suppression is performed to generate a noise suppression signal, which is calibrated as G(x′, y′). The noise suppression Gaussian filtering algorithm is used for processing, that is, the calculation formula is G(x′, y′) = F(x, y) n ×K(x′,y′), and Where F(x,y) n Denotes the input image, K(x′, y′) denotes the Gaussian kernel, ξ denotes the Gaussian width, i and j denote the coordinates after the convolution layer operation, and G(x′, y′) denotes the output image after a noise suppression signal. Then, the uniformity correction of the primary noise suppression signal G(x′, y′) is performed to generate a secondary correction signal, which is calibrated as Fu(x′, y′). The surface fitting method is used to process the primary noise suppression signal G(x′, y′). The integral calculation formula is: The feature map Fe(x,y) is obtained, and then the surface fitting is performed on the feature map Fe(x,y). The fitting calculation formula is Gu(x,y)=∑Fe(x,y)×W(xi,yj), and the estimated image Gu(x,y) is obtained. The estimated image Gu(x,y) is added back to the original image, and the calculation formula of the output image is Fu(x′,y′)=Gu(x,y)+G(x′,y′), and the secondary correction signal Fu(x′,y′) is obtained. In the formula, ψ represents the control parameter of the neighborhood, W(xi,yj) represents the surface fitting operator, and xi and yj represent the distance in the horizontal and vertical directions in the two-dimensional image respectively; Then, the quadratic corrected signal Fu(x′, y′) is used to remove artifacts in the image to generate a tertiary artifact-free signal, which is calibrated as M(x, y) and processed by the median filtering method, that is, M(x, y) = median[Fu(x′-k0, y′-l0)], where median means the median, [Fu(x′-k0, y′-l0)] means the intensity value of the quadratic corrected signal Fu(x′, y′) at the position (x′, y′), and k0 and l0 mean the size of the neighborhood. Secondly, the image after the third artifact removal signal M(x,y) is edge enhanced to generate a fourth edge enhancement signal, calibrated as L(x,y), and the Laplace filter is used to enhance the edge of the image, that is, In the formula, Expressed as the Laplace operator; Finally, the image processed by the four-time edge enhancement signal L(x, y) is subjected to multimodal fusion to generate the final high-quality image signal, calibrated as D(x, y), and the image features are fused using the weighted average method, that is, Where W i0 It is represented as the weight of the i0th mode, and n0 is represented as the number of modes to be fused; S5. Image quality assessment: Establish a generative adversarial network (GAN) evaluation model to evaluate the image clarity and overall signal balance parameters of high-quality images, including clarity, contrast, signal balance, and artifacts, and generate quality assessment coefficients; S6. Feedback monitoring and optimization model: Compare the quantitative indicators SNR and CNR according to the quality assessment coefficient, and feed back to the artifact identification and signal inhomogeneity analysis steps for continuous optimization, iterate the image correction and quality assessment processing, and obtain the final MRI image signal; The steps of comparing the quality assessment coefficient with the quantitative indicators SNR and CNR are as follows: The linear regression analysis method is used to compare the quality evaluation coefficient δ(QEC) with the quantitative indicators SNR and CNR to analyze the influence of the quantitative indicators SNR and CNR on the quality evaluation coefficient δ(QEC). The calculation formula of the linear regression analysis method is δ(QEC)=β0+β1×SNR+β2×CNR+ε0, where β0 represents the intercept term, ε0 represents the error term, and β1 and β2 represent the slopes of the quantitative indicators SNR and CNR respectively. According to the quality evaluation coefficient δ(QEC), the quantization indexes SNR and CNR, an optimization function is defined and calibrated as g(λ), which is used to iteratively calculate and adjust the CNN model parameters to improve the image quality. The calculation formula of the optimization function is: In the formula, t represents the number of iterations, It is represented as the sum of the iteration times T, λ is the adjustment factor, Δ is the actual value of the current quality evaluation coefficient δ(QEC), quantization index SNR and CNR, and g(λ) is the improvement of the current optimal value.
2. The signal equalization method for MRI imaging according to claim 1, characterized in that: The steps of establishing an animal model and providing an imaging environment are as follows: According to the pathological requirements of MRI imaging research of experimental animals, select the tissues, organs or body parts of experimental animals, take anesthetics, sedatives or experimental animal positioning devices to keep the experimental animals still and comfortable, without physiological stress response and in a stable state of physical signs, and then conduct magnetic resonance imaging examination, and conduct pathological analysis of experimental animals by identifying imaging pictures; The imaging environment has the same temperature, humidity, air pressure, air flow rate and audio space as the experimental animals’ breeding and living environment.
3. The signal equalization method for MRI imaging according to claim 2, characterized in that: The steps of collecting the MRI data are as follows: During the MRI scan, the magnetic field and radio frequency (RF) pulses excite the hydrogen nuclei in the body. When the hydrogen nuclei return to equilibrium, they release signals that are MRI data and are calibrated as S(u,v). n , and is captured by the receiving coil and stored in the parameter space of k-space. k-space is a frequency space that includes low-frequency information and high-frequency information of the image in the MRI data. (u,v) represents a point data in k-space, n represents N points in k-space, S(u,v) n It is represented as the set of all MRI data stored in k-space; Before image reconstruction, the MRI data stored in the k-space must be corrected and filled.
4. The signal equalization method for MRI imaging according to claim 3, characterized in that: The steps of image reconstruction by the two-dimensional Fourier transform algorithm are as follows: The MRI data S(u,v) after correction and filling is n , the MRI data are converted into image space using a two-dimensional Fourier transform algorithm; Then the conversion formula of the two-dimensional Fourier transform algorithm is Where, F(x,y) represents a point (u,v) in k-space that is transformed into a corresponding point in image space through two-dimensional Fourier transform, f(u,v) represents a point (u,v) in k-space, u and v represent frequency variables, x and y represent space variables, j represents an imaginary unit, ∫ represents the integral symbol, and -∞ and +∞ represent the integral region; The MRI data S(u,v) n The complex image transformed one by one into the image space is F(x,y) n , then the complex image F(x,y) n Including amplitude and phase, that is, the image signal to be evaluated is F(x,y) n .
5. The signal equalization method for MRI imaging according to claim 4, characterized in that: The steps for establishing the adversarial network GAN evaluation model are as follows: Collect the image signal F(x,y) to be evaluated n and a data training set of high-quality image signals D(x,y), and construct a GAN model, including a generator and a discriminator; Define the loss function for the GAN model, and the loss function adopts the peak signal-to-noise ratio PSNR loss function; Train the GAN model with the data training set. During the training process, the generator and discriminator will be optimized alternately; Use an independent test set to evaluate model performance and make necessary adjustments to the GAN model based on the evaluation results; After the GAN model training is completed, the discriminator is used to provide quality assessment for new images, and the output evaluation result is used as the quality assessment coefficient.
6. The signal equalization method for MRI imaging according to claim 5, characterized in that: The acquisition logic of the quality assessment coefficient is as follows: According to the image signal F(x,y) to be evaluated n And high-quality image signal D(x,y), define the peak signal-to-noise ratio PSNR loss function, then the loss function formula is and Where L(MSE) is the difference between the predicted high-quality image signal D(x, y) and the real image signal to be evaluated F(x, y) n The mean square error between them, N represents the number of samples, L (PSNR) represents the peak signal-to-noise ratio, MAX I Represented as the maximum pixel value of the image; Determine the GAN objective function to reflect the adversarial process between the generator and the discriminator. The objective function calculation formula is f(S,P)=min{S}×E(lnP(D(x,y)))×(D(x,y))+min{P,Z}×E(ln(1-P(S(Z)))), where f(S,P) is the objective function, S and P are the generator and the discriminator, respectively, D(x,y) is the real image of the input high-quality image signal, Z is the input noise of the generator, and E is the expected value. The quality assessment coefficient is the true output probability of the discriminator for the input image, that is, the calculation formula of the quality assessment coefficient is δ(QEC)=100P(D(x,y)), and P(D(x,y))∈[0,1], where δ(QEC) represents the quality assessment coefficient after the discriminator output mapping, and P(D(x,y)) represents the true output probability of the discriminator for the input image.
