MRI (Magnetic Resonance Imaging) image processing method for artificial cochlea implantation patient

Abstract

According to the MRI image processing method for the artificial cochlea implantation patient provided by the embodiment of the invention, firstly, a multi-modal input mechanism of high-field MRI, low-field MRI and CT images is introduced, and on the basis of fully keeping the advantages of each modal information, the multi-modal input mechanism of the high-field MRI, the low-field MRI and the CT images is introduced; the innovatively designed adaptive weighted fusion network is utilized to realize deep fusion and redundant information suppression of cross-modal features, the limitation of single-modal imaging is overcome, and the definition, the structural integrity and the contrast ratio of a fused image are improved. And then aiming at the problem that metal artifacts are difficult to model accurately in a traditional method, a dynamic magnetic field disturbance model based on a time sequence is constructed, and artifact distribution which is more in line with a real scene is simulated by combining factors such as head movement and magnetic field heterogeneity. On the basis, an attention enhanced artifact correction network is further designed, a local nonlinear artifact area can be accurately recognized and removed, the image purification effect is remarkably improved, and the generalization ability is high.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to a method for processing MRI images of patients with cochlear implants. Background Technology

[0002] With the rapid development of medical imaging technology, magnetic resonance imaging (MRI) has been widely used in clinical fields such as neurology, otology, and neurology, especially in the assessment of neural pathways and analysis of temporal bone structure before and after cochlear implantation. However, cochlear implants and other implants often contain metal components, which can interact strongly with high-frequency magnetic fields during MRI imaging, leading to severe metal artifacts that seriously interfere with the accuracy of anatomical observation and diagnosis.

[0003] Current research primarily employs metal artifact suppression sequences or low-field MRI scanning strategies to reduce artifact effects. However, these methods typically involve increased imaging time, decreased image resolution, or are only applicable to specific equipment and sequences. Furthermore, while low-field MRI has lower sensitivity to metal artifacts, its image quality is relatively limited, making it difficult to meet the demands of high-precision diagnosis. Therefore, effectively fusing high-field MRI, low-field MRI, and other modalities (such as CT) images to remove metal artifacts while preserving crucial anatomical information has become a key technical challenge in this field.

[0004] In recent years, the application of deep learning in medical image processing has been expanding. However, most artifact restoration methods only address static image artifacts, failing to fully consider the temporal dynamics and physical modeling characteristics of metal artifacts. Furthermore, traditional image enhancement or segmentation methods often rely on large amounts of labeled data, while the labeling cost of cochlear implant images is high and the quality varies, limiting the generalization ability of supervised learning. Therefore, there is an urgent need for an innovative technical approach that integrates multi-source information and introduces physical modeling and self-supervised mechanisms to achieve high-quality restoration and structural enhancement of post-cochlear implant images.

[0005] Based on this background, this invention proposes a complete image processing system that combines multimodal MRI image input, dynamic magnetic field perturbation modeling, metal artifact generation and correction, and multi-task self-supervised learning. By establishing a time-series-driven artifact physical model, coupled with an adaptive weighted modal fusion network and an artifact correction network, and finally combining structural enhancement and edge restoration mechanisms, the system can significantly improve MRI image quality and the accuracy of postoperative assessment. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method for processing MRI images of patients with cochlear implants.

[0007] This invention provides a method for processing MRI images of patients with cochlear implants, comprising:

[0008] S1. After acquiring high-field MRI images, low-field MRI images, and CT images, the image data are preprocessed.

[0009] S2. The preprocessed image data is processed through an adaptive weighted fusion network to perform modality fusion and deep feature learning, and the fused image data is output.

[0010] S3. Based on time series, dynamic magnetic field disturbance modeling is performed to generate metal artifacts, and then the metal artifacts are corrected through an artifact correction network.

[0011] S4. The anatomical structure restoration and enhancement processing of the image data after artifact correction is performed by a multi-task self-supervised deep learning model to output the final image data.

[0012] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S1 specifically includes:

[0013] S11. Acquire high-field MRI images Low-field MRI images and CT images Then, adaptive standardization is performed on it: ;in, For modality Standardized images, For modality Original input image, For modality The mean of the ROI region, For modality The standard deviation of the ROI region, It is the minimum value;

[0014] S12. Generate corresponding modality embedding vectors based on the image modalities of the standardized image, and then perform image denoising processing on the standardized image and the modality embedding vectors using a conditional denoising autoencoder network to obtain the denoised image. , and ;

[0015] S13. Register the denoised MRI image to the CT image coordinate space using a registration network, and output the registered image in CT space.

[0016] S14. Construct a geometric transformation field that causes implant shaking due to head movement, and then perform perturbation image space sampling; construct a metal artifact simulation field, and output the preprocessed image data after artifact superposition with the perturbation image space.

[0017] As described above and in any possible implementation, a further implementation is provided, wherein S12 specifically includes:

[0018] S121. Input the one-hot encoded vector representing the image modality into the embedding matrix to obtain the modality embedding vector;

[0019] S122. The encoder obtains a feature map from the standardized image through convolutional encoding, generates channel scaling and offset factors through modality embedding vectors, and obtains a modulated feature map after linear modulation of the intermediate feature map; expressed by the formula: ,in, This represents the modulated feature map. This represents the intermediate feature map. Represents the modal embedding vector. and These represent the channel scaling factor and offset factor generated by the fully connected network, respectively.

[0020] S123. Input the modulated feature map into the attention bottleneck module; the attention bottleneck module first calculates the channel statistics vector through global average pooling: ;Then,

[0021] The channel statistics vector, channel scaling factor, and offset factor are added together to form the fusion vector. This can be expressed as a formula: ;then,

[0022] fusion vector The input is fed into a multilayer perceptron, and the channel attention weights are calculated using activation and sigmoid functions. This can be expressed as a formula: , and These are the parameters for the fully connected layer. It is the ReLU activation function. To represent the Sigmoid function;

[0023] The decoder uses multi-layer transposed convolutions for reconstruction, and combines skip connections to recover structural details to obtain the denoised image. , and ;

[0024] The conditional denoising autoencoder network includes an encoder, a bottleneck attention module, and a decoder.

[0025] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the loss function of the conditional denoising autoencoder network is expressed as:

[0026] ; , and These are the weighting coefficients for the mean squared error loss, the structural similarity loss, and the edge preservation regularization term, respectively.

[0027] The mean squared error loss can be expressed by the formula:

[0028] , The total number of pixels in the image. The denoised image generated by the neural network pixel value, For the standardized image Each pixel value;

[0029] The structural similarity loss is expressed by the formula:

[0030] , , and They are respectively and Local mean calculated using a sliding window and They are respectively and The local variance, For covariance, and All are stability constants;

[0031] The edge preservation regularization term is expressed by the formula:

[0032] , For the first The denoised image value of each pixel. and The values ​​of adjacent pixels;

[0033] After the loss function is calculated, the parameters of the conditional denoising autoencoder network are updated using the Adam optimizer.

[0034] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the registration network is trained in the following manner:

[0035] Randomly select a set of denoised images from the dataset. , and The input is fed into the registration network to calculate the transformation matrix. and The image is then transformed and the registered image is output. and ;

[0036] Through registration loss function Calculate the total registration loss and minimize the total loss function;

[0037] in, and Here are the weighting coefficients for mean squared error and mutual information loss, where mean squared error loss is... , and For the first Pixel coordinates of each pixel, mutual information loss , This represents the joint probability distribution of the image pixels;

[0038] By calculating the gradient of the network parameters with respect to minimizing the total loss function, backpropagation is performed to optimize the network weights, thus obtaining the final registration network.

[0039] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S14 specifically includes:

[0040] S141. Input the denoised image. and ;

[0041] S142, Constructing a geometric transformation field , The range of motion of the head. The frequency of head movements. For each frame of the image, the sampling time is... For random noise;

[0042] S143. Based on the geometric transformation field, the denoised image... and Spatial sampling is performed to obtain the perturbed image. ,in, , and It is an image coordinate transformation caused by head movement;

[0043] S143. Construct a simulation field for metal artifacts. The simulation formula is as follows:

[0044] ,in, This indicates the simulation of metal artifacts. Indicates a mask for the metallic region. Indicates the intensity of metal artifacts. Indicates the location of the implant. Indicates the degree of artifact diffusion;

[0045] S144. After superimposing the simulated metal artifact field with the perturbed image, output the preprocessed image data. ,in, This is the preprocessed image data after artifact overlay. This is a function for generating random noise.

[0046] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S2 specifically includes:

[0047] S21. Acquire the preprocessed high-field MRI image, low-field MRI image, and CT image, and input them into the corresponding modal feature encoders to extract their respective modal features. , and ;

[0048] S22. Perform global average pooling on the features of each modality and obtain the adaptive fusion weights for each modality through a learnable linear transformation and the Softmax function. , and , ;

[0049] S23. The modal features described are weighted and fused according to the adaptive fusion weights to obtain the fused feature map. , Include , and , Include , and ;

[0050] S24. After processing the fused feature map into the bottleneck attention module, the enhanced fused depth feature is obtained. The bottleneck attention module, through Perform calculations;

[0051] S25. The fusion depth features Input decoder network, output fused image .

[0052] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes:

[0053] S31. Establish a dynamic perturbation field model based on the spatial mask function and time series characteristics, and generate a dynamic magnetic field perturbation field. ,in, For the first The amplitude coefficient of each disturbance component, For the first The frequency of each disturbance component For the first The initial phase of each disturbance component, This represents the total number of disturbance components. This is a mask function for the metal implantation area;

[0054] S32. Generate image data containing metal artifacts based on the dynamic magnetic field disturbance field and the fused image, wherein... , For image data containing metal artifacts, For the merged image, The imaginary unit, It is the gyromagnetic ratio constant;

[0055] S33. Construct an artifact correction network for image data containing artifacts. After correction, output the image data with artifact correction. ;

[0056] The artifact correction network includes an encoder, an attention module, and a decoder;

[0057] The loss function of the artifact correction network is: , For mean square error loss, For attention loss in artifact regions, For structural similarity loss, , and The weights are the weights for each loss.

[0058] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S4 specifically includes:

[0059] S41. Input the image data after artifact correction. The shared encoder in a multi-task self-supervised deep learning model extracts shared features, i.e. ;

[0060] S42. The anatomical structure recovery decoder of the multi-task self-supervised deep learning model recovers the enhanced anatomical structure image from shared features, i.e. ;

[0061] S43. The edge prediction decoder of the multi-task self-supervised deep learning model recovers anatomical edge information through edge supervision, that is... ;

[0062] S44. The local contrast enhancement module of the multi-task self-supervised deep learning model enhances the image contrast of local regions and outputs the final image data. ;

[0063] in, Features extracted by the shared encoder The output image after anatomical reconstruction. For edge prediction output, This is the image after local contrast enhancement. This is the contrast enhancement factor. For shared encoder network functions, Decoder function for anatomical structure recovery. For edge prediction decoder functions, For trainable parameters in a shared encoder, Trainable parameters in the anatomical structure recovery decoder. These are trainable parameters in the edge prediction decoder;

[0064] The loss function of a multi-task self-supervised deep learning model is , , and The weights for each loss are: For structural recovery loss, For edge prediction loss, This is due to the loss in contrast enhancement.

[0065] On the other hand, embodiments of the present invention provide an MRI image processing system for cochlear implant patients, the system comprising:

[0066] The preprocessing module is used to preprocess the image data after acquiring high-field MRI images, low-field MRI images, and CT images;

[0067] The fusion module is used to perform modality fusion and deep feature learning on the preprocessed image data through an adaptive weighted fusion network, and output the fused image data.

[0068] The correction module generates metal artifacts by modeling dynamic magnetic field disturbances based on time series, and then corrects the metal artifacts through an artifact correction network.

[0069] The processing module uses a multi-task self-supervised deep learning model to perform anatomical structure restoration and enhancement on the artifact-corrected image data, outputting the final image data.

[0070] One of the above technical solutions has the following beneficial effects:

[0071] (1) By introducing a multimodal input mechanism of high-field MRI, low-field MRI and CT images, while fully preserving the advantages of each modality, the innovatively designed adaptive weighted fusion network is used to achieve deep fusion of cross-modal features and suppression of redundant information, overcoming the limitations of single-modal imaging and improving the clarity, structural integrity and contrast of the fused image.

[0072] (2) To address the problem that traditional methods struggle to accurately model metal artifacts, this invention constructs a time-series-based dynamic magnetic field perturbation model. By incorporating factors such as head movement and magnetic field non-uniformity, it simulates artifact distributions that better reflect real-world scenarios. Furthermore, an attention-enhanced artifact correction network is designed, which can accurately identify and remove local nonlinear artifact regions, significantly improving image cleanup.

[0073] (3) This invention uses a multi-task self-supervised structure restoration network to perform deep reconstruction and anatomical structure enhancement on the corrected image. The network takes structure restoration, edge prediction and local contrast enhancement as synergistic goals, and still has strong generalization ability under low labeling or unlabeled conditions. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 This is a flowchart illustrating an MRI image processing method for a cochlear implant patient provided in an embodiment of the present invention.

[0076] Figure 2 A schematic block diagram of an MRI image processing system for cochlear implant patients provided in an embodiment of the present invention;

[0077] Figure 3 This is a schematic diagram of the hardware structure of an MRI image processing system for cochlear implant patients provided in an embodiment of the present invention. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0079] Please refer to Figure 1 This is a flowchart illustrating a method for processing MRI images of patients undergoing cochlear implantation, provided by an embodiment of the present invention. As shown in the figure, the method includes the following steps:

[0080] S1. After acquiring high-field MRI images, low-field MRI images, and CT images, the image data are preprocessed.

[0081] S2. The preprocessed image data is processed through an adaptive weighted fusion network to perform modality fusion and deep feature learning, and the fused image data is output.

[0082] S3. Based on time series, dynamic magnetic field disturbance modeling is performed to generate metal artifacts, and then the metal artifacts are corrected through an artifact correction network.

[0083] S4. The anatomical structure restoration and enhancement processing of the image data after artifact correction is performed by a multi-task self-supervised deep learning model to output the final image data.

[0084] Specifically, S1 includes:

[0085] S11. Acquire high-field MRI images Low-field MRI images and CT images Then, adaptive standardization is performed on it: ;in, For modality Standardized images, For modality Original input image, For modality The mean of the ROI region, For modality The standard deviation of the ROI region, It is a local minimum.

[0086] The purpose of this processing is to ensure the stability and accuracy of standardization by using an automatic segmentation network to detect brain regions or ROIs around implants, since high-field MRI images, low-field MRI images, and CT images have different intensity ranges and distribution characteristics.

[0087] S12. Based on the imaging modality of the standardized image, generate the corresponding modality coding vector. Then, input the standardized image and the modality coding vector into a conditional denoising autoencoder network for image denoising processing to obtain the denoised image. , and .

[0088] In practice, high-field MRI images are subject to Rician noise, low-field MRI images are more significantly affected by Gaussian artifacts, and CT images are typically affected by Poisson noise. Traditional general-purpose denoising algorithms (such as mean filtering, Gaussian filtering, and BM3D) cannot adapt to different noise types and distributions under different modalities, so conditional denoising autoencoders are introduced.

[0089] Specifically, S12 includes:

[0090] S121. Input the one-hot encoded vector representing the image modality into the embedding matrix to obtain the modality embedding vector;

[0091] S122. The encoder obtains a feature map from the standardized image through convolutional encoding, generates channel scaling and offset factors through modality embedding vectors, and obtains a modulated feature map after linear modulation of the intermediate feature map; expressed by the formula: ,in, This represents the modulated feature map. This represents the intermediate feature map. Represents the modal embedding vector. and These represent the channel scaling factor and offset factor generated by the fully connected network, respectively.

[0092] S123. Input the modulated feature map into the attention bottleneck module; the attention bottleneck module first calculates the channel statistics vector through global average pooling: ;Then,

[0093] The channel statistics vector, channel scaling factor, and offset factor are added together to form the fusion vector. This can be expressed as a formula: ;then,

[0094] fusion vector The input is fed into a multilayer perceptron, and the channel attention weights are calculated using activation and sigmoid functions. This can be expressed as a formula: , and These are the parameters for the fully connected layer. It is the ReLU activation function. To represent the Sigmoid function;

[0095] The decoder uses multi-layer transposed convolutions for reconstruction, and combines skip connections to recover structural details to obtain the denoised image. , and ;

[0096] The conditional denoising autoencoder network comprises an encoder, a bottleneck attention module, and a decoder. Its specific operation is as follows: the encoder extracts image features and modulates them using modality vectors; the bottleneck module performs channel attention weighting; and the decoder reconstructs the denoised image based on the modulated feature map. The denoising network first uses the encoder to convolve-encode the input image to obtain a feature map, then injects the modality vectors into the intermediate features via broadcasting. Next, the channel attention module of the bottleneck layer is compressed to adaptively adjust the modality response. The decoder uses multi-layer transposed convolutions for reconstruction, combined with skip connections to restore structural details.

[0097] The conditional denoising autoencoder network is trained in the following way:

[0098] (1) Construct the loss function:

[0099] The loss function of a conditional denoising autoencoder network is expressed as:

[0100] ; , and These are the weighting coefficients for the mean squared error loss, the structural similarity loss, and the edge preservation regularization term, respectively.

[0101] The mean squared error loss can be expressed by the formula:

[0102] , The total number of pixels in the image. The denoised image generated by the neural network pixel value, For the standardized image Each pixel value;

[0103] The structural similarity loss is expressed by the formula:

[0104] , , and They are respectively and Local mean calculated using a sliding window and They are respectively and The local variance, For covariance, and All are stability constants;

[0105] The edge preservation regularization term is expressed by the formula:

[0106] , For the first The denoised image value of each pixel. and The values ​​of adjacent pixels;

[0107] After the loss function is calculated, the parameters of the conditional denoising autoencoder network are updated using the Adam optimizer.

[0108] (2) Construct a training set containing the validation set of the dataset, and randomly select modal images from the dataset to add noise. You can choose any of the Rician, Gaussian and Poisson modal images.

[0109] The denoising network first uses an encoder to encode the noise-increased input image through convolution to obtain a feature map, and then injects the modality vector into the intermediate features. Next, it uses a channel attention module in the compression bottleneck layer to adaptively adjust the modality response. The decoder uses multi-layer transposed convolution for reconstruction and combines skip connections to restore structural details and output the denoised image data.

[0110] (3) Substitute each parameter and image data into the loss function of the conditional denoising autoencoder network, and continue to substitute them into the training until the loss function converges, then the training of the loss function is completed.

[0111] S13. Register the denoised MRI image to the CT image coordinate space using a registration network, and output the registered image in CT space.

[0112] The goal of using a registration network is to achieve spatial alignment between three modalities (high-field MRI, low-field MRI, and CT), ensuring spatial consistency between images and the effectiveness of multimodal image fusion.

[0113] The registration network includes:

[0114] The encoder extracts high-level features from the image. It consists of several convolutional layers (Conv), each followed by a ReLU activation function.

[0115] Transform estimator: Uses fully connected (FC) layers to estimate the spatial transformation matrix (usually a rigid or affine transformation) from features extracted from the encoder.

[0116] Decoder: Applies the transformation to the denoised MRI image through deconvolution or upsampling and generates the registered image.

[0117] The registration network was trained in the following manner:

[0118] Randomly select a set of denoised images from the dataset. , and The input is fed into the registration network to calculate the transformation matrix. and The image is then transformed and the registered image is output. and ;

[0119] Through registration loss function Calculate the total registration loss and minimize the total loss function;

[0120] in, and Here are the weighting coefficients for mean squared error and mutual information loss, where mean squared error loss is... , and For the first Pixel coordinates of each pixel, mutual information loss , This represents the joint probability distribution of the image pixels;

[0121] By calculating the gradient of the network parameters with respect to minimizing the total loss function, backpropagation is performed to optimize the network weights, thus obtaining the final registration network.

[0122] S13 specifically includes:

[0123] The registration process involves transforming the matrix. Aligning MRI images to the coordinate space of CT images involves the following process: , , This indicates an image transformation operation, and options include rigid transformation, affine transformation, etc. and The transformation matrix;

[0124] Transformation matrix and Estimate using the following methods: ,

[0125] and These are the feature representations of high-field and low-field MRI images after being encoded, respectively. It is a transformation estimation module, which can be a fully connected network, a transformation learning module, etc.

[0126] S14. Construct a geometric transformation field that causes implant shaking due to head movement, and then perform perturbation image space sampling; construct a metal artifact simulation field, and output the preprocessed image data after artifact superposition with the perturbation image space.

[0127] The purpose of S14 is to simulate implant jitter and metal artifact perturbation caused by head movement and to overlay these artifacts in the image so that the subsequent network can learn how to remove these artifacts.

[0128] S14 specifically includes:

[0129] S141. Input the denoised image. and .

[0130] As the patient's head moves, the metal implant will vibrate slightly, causing a slight spatial displacement and distortion in the image of the metal area. We use a geometric transformation field to simulate this process.

[0131] S142, Constructing a geometric transformation field , The range of motion of the head. The frequency of head movements. For each frame of the image, the sampling time is... The random noise term simulates irregular implant vibration.

[0132] Based on the above geometric transformation field Spatial sampling is performed on the denoised MRI image. This means that each pixel location... According to the perturbation amount of the transformed field Repositioning is performed to obtain the perturbed image. .

[0133] S143. Based on the geometric transformation field, the denoised image... and Spatial sampling is performed to obtain the perturbed image. ,in, , and It is an image coordinate transformation caused by head movement.

[0134] To simulate the artifacts caused by metal implants, we constructed a metal artifact field that simulates the superposition effect of metal artifacts in the image based on the spatial changes of the implant area.

[0135] S143. Construct a simulation field for metal artifacts. The simulation formula is as follows:

[0136] ,in, This indicates the simulation of metal artifacts. Indicates a mask for the metallic region. Indicates the intensity of metal artifacts. Indicates the location of the implant. Indicates the degree of artifact diffusion;

[0137] S144. After superimposing the simulated metal artifact field with the perturbed image, output the preprocessed image data. ,in, This is the preprocessed image data after artifact overlay. This is a function for generating random noise.

[0138] Specifically, S2 includes:

[0139] S21. Acquire the preprocessed high-field MRI image, low-field MRI image, and CT image, and input them into the corresponding modal feature encoders to extract their respective modal features. , and ;

[0140] Each modal input extracts features through an independent convolutional encoder, and the modal feature encoding network is a stacked structure containing Conv+BN+ReLU.

[0141] S22. Perform global average pooling on the features of each modality and obtain the adaptive fusion weights for each modality through a learnable linear transformation and the Softmax function. , and , ;

[0142] In this step, we introduce an attention weight module to calculate the fusion weights for each modality based on the channel attention mechanism: , For global average pooling, the feature map is compressed into a vector; It is a learnable linear transformation.

[0143] S23. The modal features described are weighted and fused according to the adaptive fusion weights to obtain the fused feature map. , Include , and , Include , and ;

[0144] S24. After processing the fused feature map into the bottleneck attention module, the enhanced fused depth feature is obtained. The bottleneck attention module, through Perform calculations;

[0145] S25. The fusion depth features Input decoder network, output fused image .

[0146] The purpose of S3 is to accurately model dynamic magnetic field disturbances using time-series data, simulate dynamic artifacts caused by metallic implants during real magnetic resonance imaging (MRI) scans, and then remove these artifacts through a specially designed artifact correction network to restore medical image quality. S3 specifically includes:

[0147] S31. Establish a dynamic perturbation field model based on the spatial mask function and time series characteristics, and generate a dynamic magnetic field perturbation field. ,in, For the first The amplitude coefficient of each disturbance component, For the first The frequency of each disturbance component For the first The initial phase of each disturbance component, This represents the total number of disturbance components. This is a mask function for the metal implantation area;

[0148] S32. Generate image data containing metal artifacts based on the dynamic magnetic field disturbance field and the fused image, wherein... , For image data containing metal artifacts, For the merged image, The imaginary unit, It is the gyromagnetic ratio constant;

[0149] S33. Construct an artifact correction network for image data containing artifacts. After correction, output the image data with artifact correction. ;

[0150] The artifact correction network includes an encoder, an attention module, and a decoder;

[0151] The loss function of the artifact correction network is: , For mean square error loss, For attention loss in artifact regions, For structural similarity loss, , and The weights are the weights for each loss.

[0152] The purpose of S4 is to further restore and enhance the anatomical structures of medical images after artifact correction, thereby improving image clarity, structural accuracy, and clinical applicability. Specifically, S4 includes:

[0153] S41. Input the image data after artifact correction. The shared encoder in a multi-task self-supervised deep learning model extracts shared features, i.e. ;

[0154] S42. The anatomical structure recovery decoder of the multi-task self-supervised deep learning model recovers the enhanced anatomical structure image from shared features, i.e. ;

[0155] S43. The edge prediction decoder of the multi-task self-supervised deep learning model recovers anatomical edge information through edge supervision, that is... ;

[0156] S44. The local contrast enhancement module of the multi-task self-supervised deep learning model enhances the image contrast of local regions and outputs the final image data. ;

[0157] in, Features extracted by the shared encoder The output image after anatomical reconstruction. For edge prediction output, This is the image after local contrast enhancement. This is the contrast enhancement factor. For shared encoder network functions, Decoder function for anatomical structure recovery. For edge prediction decoder functions, For trainable parameters in a shared encoder, Trainable parameters in the anatomical structure recovery decoder. These are trainable parameters in the edge prediction decoder;

[0158] The loss function of a multi-task self-supervised deep learning model is , , and The weights for each loss are: For structural recovery loss, For edge prediction loss, This is due to the loss in contrast enhancement.

[0159] Through the above steps, the present invention achieves the following technical effects:

[0160] (1) By introducing a multimodal input mechanism of high-field MRI, low-field MRI and CT images, while fully preserving the advantages of each modality, the innovatively designed adaptive weighted fusion network is used to achieve deep fusion of cross-modal features and suppression of redundant information, overcoming the limitations of single-modal imaging and improving the clarity, structural integrity and contrast of the fused image.

[0161] (2) To address the problem that traditional methods struggle to accurately model metal artifacts, this invention constructs a time-series-based dynamic magnetic field perturbation model. By incorporating factors such as head movement and magnetic field non-uniformity, it simulates artifact distributions that better reflect real-world scenarios. Furthermore, an attention-enhanced artifact correction network is designed, which can accurately identify and remove local nonlinear artifact regions, significantly improving image cleanup.

[0162] (3) Furthermore, this invention uses a multi-task self-supervised structural restoration network to perform deep reconstruction and anatomical enhancement on the corrected image. This network has structural restoration, edge prediction and local contrast enhancement as synergistic goals, and still has strong generalization ability under low labeling or unlabeled conditions.

[0163] The present invention further provides an apparatus embodiment for implementing the steps and methods in the above method embodiments.

[0164] Please refer to Figure 2 This is an MRI image processing system for cochlear implant patients provided in an embodiment of the present invention. The system includes:

[0165] The preprocessing module 210 is used to preprocess the image data after acquiring high-field MRI images, low-field MRI images and CT images;

[0166] The fusion module 220 is used to perform modal fusion and deep feature learning on the preprocessed image data through an adaptive weighted fusion network, and output the fused image data.

[0167] The correction module 230 generates metal artifacts by modeling dynamic magnetic field disturbances based on time series, and then corrects the metal artifacts through an artifact correction network.

[0168] The processing module 230 performs anatomical structure restoration and enhancement processing on the artifact-corrected image data using a multi-task self-supervised deep learning model, and outputs the final image data.

[0169] Because each unit module in this embodiment can execute Figure 1 For the methods shown, parts not described in detail in this embodiment, please refer to the [examples / descriptions]. Figure 1 Related explanations.

[0170] Please refer to Figure 3 This is a schematic diagram of the hardware structure of an MRI image processing system for cochlear implant patients provided in an embodiment of the present invention. The data prediction device includes at least one processor and a memory. The at least one processor is coupled to the memory and is used to read and execute instructions in the memory to perform the MRI image processing method for cochlear implant patients provided in an embodiment of the present invention.

[0171] Thirdly, embodiments of the present invention provide a computer-readable medium. The computer-readable medium stores program code that, when executed on a computer, causes the computer to perform the MRI image processing method for cochlear implant patients provided in embodiments of the present invention.

[0172] At the hardware level, the device may include a processor, and optionally an internal bus, network interface, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk drive. Of course, the device may also include other hardware required for other business operations.

[0173] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be categorized into address bus, data bus, control bus, etc.

[0174] Memory is used to store programs. Specifically, programs may include program code, which includes computer operation instructions. Memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0175] The steps of the method disclosed in the embodiments of this invention can be directly manifested as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0176] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0177] For ease of description, the above apparatus is described by dividing it into various units or modules according to their functions. Of course, in implementing this invention, the functions of each unit or module can be implemented in one or more software and / or hardware.

[0178] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0179] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0181] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0183] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0184] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0185] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0186] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0187] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0188] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0189] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for processing MRI images of patients with cochlear implants, characterized in that, The method includes: S1. After acquiring high-field MRI images, low-field MRI images, and CT images, the image data are preprocessed. S2. The preprocessed image data is processed through an adaptive weighted fusion network to perform modality fusion and deep feature learning, and the fused image data is output. S3. Based on time series, dynamic magnetic field disturbance modeling is performed to generate metal artifacts, and then the metal artifacts are corrected through an artifact correction network. S4. The anatomical structure restoration and enhancement processing of the image data after artifact correction is performed by a multi-task self-supervised deep learning model to output the final image data.

2. The MRI image processing method for cochlear implant patients according to claim 1, characterized in that, S1 specifically includes: S11. Acquire high-field MRI images Low-field MRI images and CT images Then, adaptive standardization is performed on it: ;in, For modality Standardized images, For modality Original input image, For modality The mean of the ROI region, For modality The standard deviation of the ROI region, It is the minimum value; S12. Based on the imaging modality of the standardized image, generate the corresponding modality coding vector. Then, input the standardized image and the modality coding vector into a conditional denoising autoencoder network for image denoising processing to obtain the denoised image. , and ; S13. Register the denoised MRI image to the CT image coordinate space using a registration network, and output the registered image in CT space. S14. Construct a geometric transformation field that causes implant shaking due to head movement, and then perform perturbation image space sampling; construct a metal artifact simulation field, and output the preprocessed image data after artifact superposition with the perturbation image space.

3. The MRI image processing method for cochlear implant patients according to claim 2, characterized in that, S12 specifically includes: S121. Input the one-hot encoded vector representing the image modality into the embedding matrix to obtain the modality embedding vector; S122. The encoder obtains a feature map from the standardized image through convolutional encoding, generates channel scaling and offset factors through modality embedding vectors, and obtains a modulated feature map after linear modulation of the intermediate feature map; expressed by the formula: ,in, This represents the modulated feature map. This represents the intermediate feature map. Represents the modal embedding vector. and These represent the channel scaling factor and offset factor generated by the fully connected network, respectively. S123. Input the modulated feature map into the attention bottleneck module; the attention bottleneck module first calculates the channel statistics vector through global average pooling: ;Then, The channel statistics vector, channel scaling factor, and offset factor are added together to form the fusion vector. This can be expressed as a formula: ;then, fusion vector The input is fed into a multilayer perceptron, and the channel attention weights are calculated using activation and sigmoid functions. This can be expressed as a formula: , and These are the parameters for the fully connected layer. It is the ReLU activation function. To represent the Sigmoid function; The decoder uses multi-layer transposed convolutions for reconstruction, and combines skip connections to recover structural details to obtain the denoised image. , and ; The conditional denoising autoencoder network includes an encoder, a bottleneck attention module, and a decoder.

4. The MRI image processing method for cochlear implant patients according to claim 3, characterized in that, The loss function of the conditional denoising autoencoder network is expressed as: ; , and These are the weighting coefficients for the mean squared error loss, the structural similarity loss, and the edge preservation regularization term, respectively. The mean squared error loss can be expressed by the formula: , The total number of pixels in the image. The denoised image generated by the neural network pixel value, For the standardized image Each pixel value; The structural similarity loss is expressed by the formula: , , and They are respectively and Local mean calculated using a sliding window and They are respectively and The local variance, For covariance, and All are stability constants; The edge preservation regularization term is expressed by the formula: , For the first The denoised image value of each pixel. and The values ​​of adjacent pixels; After the loss function is calculated, the parameters of the conditional denoising autoencoder network are updated using the Adam optimizer.

5. The MRI image processing method for cochlear implant patients according to claim 2, characterized in that, The registration network was trained in the following manner: Randomly select a set of denoised images from the dataset. , and The input is fed into the registration network to calculate the transformation matrix. and The image is then transformed and the registered image is output. and ; Through registration loss function Calculate the total registration loss and minimize the total loss function; in, and Here are the weighting coefficients for mean squared error and mutual information loss, where mean squared error loss is... , and For the first Pixel coordinates of each pixel, mutual information loss , This represents the joint probability distribution of the image pixels; By calculating the gradient of the network parameters with respect to minimizing the total loss function, backpropagation is performed to optimize the network weights, thus obtaining the final registration network.

6. The MRI image processing method for patients with cochlear implantation according to claim 2, characterized in that, S14 specifically includes: S141. Input the denoised image. and ; S142, Constructing a geometric transformation field , The range of motion of the head. The frequency of head movements. For each frame of the image, the sampling time is... For random noise; S143. Based on the geometric transformation field, the denoised image... and Spatial sampling is performed to obtain the perturbed image. ,in, , and It is an image coordinate transformation caused by head movement; S143. Construct a simulation field for metal artifacts. The simulation formula is as follows: ,in, This indicates the simulation of metal artifacts. Indicates a mask for the metallic region. Indicates the intensity of metal artifacts. Indicates the location of the implant. Indicates the degree of artifact diffusion; S144. After superimposing the simulated metal artifact field with the perturbed image, output the preprocessed image data. ,in, This is the preprocessed image data after artifact overlay. This is a function for generating random noise.

7. The method for processing MRI images of patients with cochlear implants according to claim 2, characterized in that, S2 specifically includes: S21. Acquire the preprocessed high-field MRI image, low-field MRI image, and CT image, and input them into the corresponding modal feature encoders to extract their respective modal features. , and ; S22. Perform global average pooling on the features of each modality and obtain the adaptive fusion weights for each modality through a learnable linear transformation and the Softmax function. , and , ; S23. The modal features described are weighted and fused according to the adaptive fusion weights to obtain the fused feature map. , Include , and , Include , and ; S24. After processing the fused feature map into the bottleneck attention module, the enhanced fused depth feature is obtained. The bottleneck attention module, through Perform calculations; S25. The fusion depth features Input decoder network, output fused image .

8. The method for processing MRI images of patients undergoing cochlear implantation according to claim 2, characterized in that, S3 specifically includes: S31. Establish a dynamic perturbation field model based on the spatial mask function and time series characteristics, and generate a dynamic magnetic field perturbation field. ,in, For the first The amplitude coefficient of each disturbance component, For the first The frequency of each disturbance component For the first The initial phase of each disturbance component, This represents the total number of disturbance components. This is a mask function for the metal implantation area; S32. Generate image data containing metal artifacts based on the dynamic magnetic field disturbance field and the fused image, wherein... , For image data containing metal artifacts, For the merged image, The imaginary unit, It is the gyromagnetic ratio constant; S33. Construct an artifact correction network for image data containing artifacts. After correction, output the image data with artifact correction. ; The artifact correction network includes an encoder, an attention module, and a decoder; The loss function of the artifact correction network is: , For mean square error loss, For attention loss in artifact regions, For structural similarity loss, , and The weights are the weights for each loss.

9. The method for processing MRI images of patients with cochlear implants according to claim 2, characterized in that, S4 specifically includes: S41. Input the image data after artifact correction. The shared encoder in a multi-task self-supervised deep learning model extracts shared features, i.e. ; S42. The anatomical structure recovery decoder of the multi-task self-supervised deep learning model recovers the enhanced anatomical structure image from shared features, i.e. ; S43. The edge prediction decoder of the multi-task self-supervised deep learning model recovers anatomical edge information through edge supervision, that is... ; S44. The local contrast enhancement module of the multi-task self-supervised deep learning model enhances the image contrast of local regions and outputs the final image data. ; in, Features extracted by the shared encoder The output image after anatomical reconstruction. For edge prediction output, This is the image after local contrast enhancement. This is the contrast enhancement factor. For shared encoder network functions, Decoder function for anatomical structure recovery. For edge prediction decoder functions, For trainable parameters in a shared encoder, Trainable parameters in the anatomical structure recovery decoder. These are trainable parameters in the edge prediction decoder; The loss function of a multi-task self-supervised deep learning model is , , and The weights for each loss are: For structural recovery loss, For edge prediction loss, This is due to the loss in contrast enhancement.

10. An MRI image processing system for a cochlear implant patient using the method of any one of claims 1-9, characterized in that, The system includes: The preprocessing module is used to preprocess the image data after acquiring high-field MRI images, low-field MRI images, and CT images; The fusion module is used to perform modality fusion and deep feature learning on the preprocessed image data through an adaptive weighted fusion network, and output the fused image data. The correction module generates metal artifacts by modeling dynamic magnetic field disturbances based on time series, and then corrects the metal artifacts through an artifact correction network. The processing module uses a multi-task self-supervised deep learning model to perform anatomical structure restoration and enhancement on the artifact-corrected image data, outputting the final image data.

Images