Deep learning-based magnetic resonance image feature extraction method and system

Abstract

The invention relates to the technical field of image data processing, in particular to a magnetic resonance image feature extraction method and system based on deep learning, and the method comprises the following steps: obtaining original K space data, extracting the numerical components of a real part and an imaginary part to construct a real part data matrix and an imaginary part data matrix, performing cross convolution on the initialized real part and imaginary part convolution kernel weight parameters and the real part and imaginary part data matrixes to generate four groups of intermediate feature maps, calculating a difference value of the first and second intermediate feature maps to obtain an output feature real part map, calculating a sum value of the third and fourth intermediate feature maps to obtain an output feature imaginary part map, and outputting the output feature real part map and the output feature imaginary part map. A deep complex feature vector is generated by performing linear rectification non-linear activation mapping. According to the invention, by constructing the complex convolution operation, the depth feature mining of the magnetic resonance complex data can be realized by using the phase and amplitude information of the magnetic resonance image, and the precision and robustness of feature extraction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a method and system for extracting magnetic resonance image features based on deep learning. Background Technology

[0002] Image data processing technology encompasses the general technical scope of analyzing, processing, encoding, and transforming digital image signals using computer hardware and software systems. Traditional magnetic resonance image feature extraction methods typically involve first preprocessing the raw magnetic resonance image data with a Gaussian or median filter for denoising; then using threshold segmentation algorithms or edge detection operators to delineate specific regions of interest in the image; subsequently, calculating texture parameters such as contrast and entropy using the gray-level co-occurrence matrix; or obtaining statistical values ​​such as the mean, variance, and skewness of pixel gray levels through histogram statistics; and finally, concatenating these manually defined mathematical indicators into a feature vector.

[0003] Traditional magnetic resonance imaging feature extraction relies on Gaussian or median filters for denoising preprocessing and threshold segmentation to define regions of interest. It also calculates gray-level co-occurrence matrices or histogram statistical indicators to assemble feature vectors. This over-reliance on manually set mathematical indicators and preprocessing steps can only extract surface texture and statistical information, ignoring the high-dimensional nonlinear features contained in the original image data. Furthermore, it is prone to losing detailed information during denoising and segmentation, resulting in limited feature expression capabilities. This makes it difficult to fully explore the potential deep pathological patterns and related information in the original data, and thus cannot meet the complex and ever-changing clinical diagnostic needs. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method and system for extracting magnetic resonance image features.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based magnetic resonance image feature extraction method, comprising the following steps: S1: Obtain the original K-space data, extract the real part and imaginary part of the original K-space data, and construct the real part data matrix and the imaginary part data matrix; S2: Initialize the real part convolution kernel weight parameters and the imaginary part convolution kernel weight parameters. Convolve the real part data matrix with the real part convolution kernel weight parameters to generate a first intermediate feature map. Convolve the imaginary part data matrix with the imaginary part convolution kernel weight parameters to generate a second intermediate feature map. Convolve the real part data matrix with the imaginary part convolution kernel weight parameters to generate a third intermediate feature map. Convolve the imaginary part data matrix with the real part convolution kernel weight parameters to generate a fourth intermediate feature map. S3: Calculate the difference between the first intermediate feature map and the second intermediate feature map to generate the real part of the output feature map, and calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate the imaginary part of the output feature map; S4: By performing linear rectification nonlinear activation mapping on the real part graph of the output features and the imaginary part graph of the output features, values ​​less than zero are set to zero and values ​​greater than or equal to zero are retained to generate a deep complex feature vector.

[0006] As a further aspect of the present invention, step S1 specifically comprises: S11: The radio frequency receiving coil of the magnetic resonance imaging equipment is used to collect echo signals under differentiated frequency encoding and phase encoding. The analog signal is converted into a discrete complex domain numerical sequence through the analog-to-digital conversion interface. The header information of the sequence is analyzed to determine the imaging field size and sampling point density, and the original K-space data is obtained. S12: Traverse each complex sampling point in the original K-space data, use the complex separation algorithm to extract the real part scalar value and imaginary part scalar value corresponding to multiple sampling points, and perform Z-score-based standardization processing on the extracted real part and imaginary part values ​​to generate standardized real part and imaginary part values. S13: Based on the imaging field size, the standardized real and imaginary numerical components are mapped to a two-dimensional grid coordinate system, and the missing values ​​in the unsampled area are filled by bilinear interpolation. The processed component values ​​are then reorganized into a two-dimensional tensor structure according to the row and column index order to construct the real data matrix and the imaginary data matrix respectively.

[0007] As a further aspect of the present invention, step S2 specifically comprises: S21: Set the kernel size, stride parameter and zero padding number of the convolutional neural network layer, use the Glorot uniform distribution initialization strategy to generate random floating-point values ​​that conform to the normal distribution law, and assign the random floating-point values ​​to the weight variables of the real part channel and the imaginary part channel respectively, and initialize the weight parameters of the real part convolutional kernel and the weight parameters of the imaginary part convolutional kernel. S22: Invoke the two-dimensional convolution operator in the deep learning computing framework, slide the real part convolution kernel weight parameters on the real part data matrix with a set stride and perform a dot product summation operation to generate the first intermediate feature map, slide the imaginary part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the second intermediate feature map, slide the imaginary part convolution kernel weight parameters on the real part data matrix and perform a dot product summation operation to generate the third intermediate feature map, and simultaneously slide the real part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the fourth intermediate feature map.

[0008] As a further aspect of the present invention, step S3 specifically comprises: S31: Obtain the first intermediate feature map and the second intermediate feature map, perform a point-by-point subtraction operation at the matrix element level, subtract the value of the corresponding position in the second intermediate feature map from the value of each pixel position in the first intermediate feature map, and generate the output feature real part map. S32: Obtain the third intermediate feature map and the fourth intermediate feature map, perform a point-by-point addition operation at the matrix element level, and superimpose the value of each pixel position of the third intermediate feature map with the value of the corresponding position of the fourth intermediate feature map to generate the output feature imaginary part map.

[0009] As a further aspect of the present invention, step S4 specifically comprises: S41: Perform nonlinear rectification operations on the real part map and the imaginary part map of the output features respectively, filter out and retain pixel values ​​greater than or equal to zero in the matrix through comparison operation, and forcibly replace all pixel values ​​less than zero in the matrix with zero values ​​to generate the activated real part feature map and imaginary part feature map. As a further aspect of the present invention, step S4 specifically comprises: S41: Perform nonlinear rectification operations on the real part map and the imaginary part map of the output features respectively, filter out and retain pixel values ​​greater than or equal to zero in the matrix through comparison operation, and forcibly replace all pixel values ​​less than zero in the matrix with zero values ​​to generate the activated real part feature map and imaginary part feature map. S42: Perform global average pooling on the activated real feature map and imaginary feature map, calculate the pixel mean in multiple feature channels to compress the spatial dimension, and concatenate and fuse the compressed real feature data and imaginary feature data according to the channel dimension. Flatten the fused multidimensional data into a one-dimensional array through a fully connected layer to generate the deep complex feature vector.

[0010] As a further aspect of the present invention, the standardization process of S12 includes: The distribution of modulus values ​​of all sampling points in the original K-space data is statistically analyzed. The global mean and global standard deviation of the distribution are calculated. The real and imaginary numerical components are center-shifted using the global mean. The shifted component values ​​are scaled and normalized using the global standard deviation. A statistical parameter mapping table for inverse standardization is established to generate standardized component data.

[0011] As a further aspect of the present invention, the dot product summation process of S22 includes: Based on a pre-defined sliding window mechanism, local receptive field data blocks corresponding to the convolution kernel size are extracted from the input matrix. A weighted sum between the local receptive field data blocks and their corresponding convolution kernel weights is calculated using matrix multiplication. A pre-defined bias term is then introduced to linearly correct the weighted sum. The calculation logic satisfies the following formula: ; Generate feature map pixel values; in, Represents the coordinates in the output feature map The pixel value at that location, and These represent the height and width of the convolution kernel, respectively. The stride parameter represents the convolution operation. This represents the value of the input matrix at the corresponding position in the sliding window. Represents the position of the convolution kernel The weight value at the location, This represents the bias parameter used to adjust the activation threshold.

[0012] As a further aspect of the present invention, the point-by-point subtraction operation process of S31 includes: Obtain the first pixel matrix in the first intermediate feature map and the second pixel matrix in the second intermediate feature map, check whether their dimensions are consistent. If they are consistent, start a parallel computing thread, perform arithmetic difference calculation on the floating-point value of each corresponding coordinate position, use the calculated difference to construct the current matrix element, and write it into the target memory address to generate the output feature real part map.

[0013] As a further aspect of the present invention, the nonlinear rectification operation process of S41 includes: The activation threshold of the linear rectifier unit is set to zero. The value of each pixel in the input feature map is traversed. The sign attribute of the value is identified by the condition judgment instruction. If the value is negative, the gradient propagation path corresponding to it is blocked and the output value is set to zero. If the value is positive, its original linear response characteristics are kept unchanged, and the corrected feature map is generated.

[0014] A deep learning-based magnetic resonance image feature extraction system is provided. This system implements the aforementioned deep learning-based magnetic resonance image feature extraction method. The system includes: The data acquisition and preprocessing module is used to acquire raw K-space data, extract the real and imaginary numerical components of the raw K-space data, and perform data formatting and standardization processing to construct real and imaginary data matrices. The complex convolution operation module is used to initialize the real part convolution kernel weight parameters and the imaginary part convolution kernel weight parameters, convolve the real part data matrix with the real part convolution kernel weight parameters to generate a first intermediate feature map, convolve the imaginary part data matrix with the imaginary part convolution kernel weight parameters to generate a second intermediate feature map, convolve the real part data matrix with the imaginary part convolution kernel weight parameters to generate a third intermediate feature map, and convolve the imaginary part data matrix with the real part convolution kernel weight parameters to generate a fourth intermediate feature map; The feature difference fusion module is used to calculate the difference between the first intermediate feature map and the second intermediate feature map to generate an output feature real part map, and to calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate an output feature imaginary part map. The nonlinear activation mapping module is used to generate a deep complex feature vector by performing linear rectification nonlinear activation mapping on the real part graph of the output feature and the imaginary part graph of the output feature, setting values ​​less than zero to zero and retaining values ​​greater than or equal to zero.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the real and imaginary numerical components of the original K-space data are obtained and corresponding data matrices are constructed. After initializing the weight parameters, cross-convolution operations of the real and imaginary parts are performed respectively. The real and imaginary part maps of the output features are generated by calculating the difference and sum of the intermediate feature maps. Linear rectification nonlinear activation mapping is performed on the output feature maps to retain positive values. The operation mechanism based on the complex domain can completely preserve the phase and amplitude information in the original magnetic resonance data, avoid the signal loss caused by traditional modulus processing, effectively extract and fuse the correlation features between the real and imaginary parts, and enhance the ability of the feature vector to represent deep tissue structures and pathological details. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the main flow of the magnetic resonance image feature extraction method of the present invention; Figure 2 This is a schematic diagram of the original K-space data acquisition and preprocessing process of the present invention; Figure 3 This is a schematic diagram of the complex field convolution operation and intermediate feature extraction process of the present invention; Figure 4 This is a schematic diagram of the process of generating complex feature maps by fusing intermediate feature maps according to the present invention; Figure 5 This is a schematic diagram of the nonlinear activation and feature vector generation process of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a deep learning-based method for extracting features from magnetic resonance images, comprising the following steps: S1: Obtain the original K-space data, extract the real and imaginary numerical components of the original K-space data, and construct the real and imaginary data matrices; The specific steps of S1 are as follows: S11: The radio frequency receiving coil of the magnetic resonance imaging equipment is used to collect echo signals under differentiated frequency encoding and phase encoding. The analog signal is converted into a discrete complex domain numerical sequence through the analog-to-digital conversion interface. The sequence header information is analyzed to determine the imaging field size and sampling point density, and the original K-space data is obtained. S12: Traverse each complex sampling point in the original K-space data, use the complex separation algorithm to extract the real part scalar value and imaginary part scalar value corresponding to multiple sampling points, and perform Z-score-based standardization on the extracted real part and imaginary part numerical components to generate standardized real part and imaginary part numerical components. The standardization process for S12 includes: The distribution of modulus values ​​of all sampling points in the original K-space data is statistically analyzed. The global mean and global standard deviation of the distribution are calculated. The real and imaginary numerical components are shifted by centering using the global mean. The shifted component values ​​are scaled and normalized using the global standard deviation. A statistical parameter mapping table for inverse standardization is established to generate standardized component data. S13: Based on the imaging field size, the standardized real and imaginary numerical components are mapped to a two-dimensional grid coordinate system. The missing values ​​in the unsampled area are filled by bilinear interpolation. The processed component values ​​are reorganized into a two-dimensional tensor structure according to the row and column index order, and the real and imaginary data matrices are constructed respectively.

[0020] Regarding the process of acquiring and analyzing signals using magnetic resonance imaging (MRI) equipment in S11, the main MRI magnet is first activated to generate a static magnetic field of 1.5T or 3.0T. A gradient coil applies a trapezoidal gradient magnetic field for spatial positioning. The radio frequency (RF) receiving coil, after being excited by a preset pulse sequence, receives the RF echo signal generated by the nuclear magnetic resonance of biological tissue. This analog signal is amplified by a preamplifier and then transmitted to an analog-to-digital converter (ADC). The ADC discretizes the continuous analog voltage signal at a sampling rate of 10MHz and converts the voltage amplitude of each sampling point into a 16-bit or 32-bit binary value, forming a complex domain numerical sequence. Subsequently, the system reads the metadata information in the header of this sequence file and locates the tag group defined in the DICOM protocol. Specifically, it reads tag (0018, 0050) to obtain the slice thickness parameter, tag (0028, 0030) to obtain the pixel pitch, and tag (0018, 1310) to obtain the acquisition matrix size, which is then set as follows: The system reads tags (0018, 0088) to obtain the imaging field of view (FOV) size parameters. Based on the row and column number parameters parsed above, the system allocates corresponding continuous storage space in memory and fills the discretized complex domain numerical sequence in row-major order to complete the memory loading of the original K-space data.

[0021] The aforementioned raw K-space data refers to the raw frequency domain data that is directly received by the radio frequency coil and digitized during magnetic resonance imaging. Its coordinate system is defined in the spatial frequency domain and contains all the amplitude and phase information required for image reconstruction.

[0022] For the data traversal and standardization process in S12, the program uses a double loop structure to traverse the two-dimensional array corresponding to the original K-space data; For the array index is Each complex sampling point The system calls the complex number arithmetic instruction set to read the real part of the low-order byte and the imaginary part of the high-order byte at the memory address, respectively. Let the original complex number sampling points be... Then extract the scalar values ​​respectively. As the original value of the real part As the original values ​​of the imaginary part, they are stored in the temporary matrix of the real part. With the imaginary temporary matrix ; Next, Z-score-based normalization is performed; The system first and All valid sampling points are traversed and counted; Define the total number of sampling points as ; Calculate the distribution parameters of the real part data: Initialize the accumulator to 0, and then... Sum of all values ​​and divide by Obtain the global average value of the real part ; Calculate each value with Square the differences, sum these squared values ​​and divide by Then, perform a square root operation on the result to obtain the global standard deviation of the real part. ; Similarly, calculate the global average value of the imaginary part. Global standard deviation of imaginary part To ensure the physical meaning of the data and the feasibility of subsequent inversion, the system establishes a statistical parameter mapping table in memory to record... and Four floating-point values; Subsequently, Each value in Perform standardized operations; the operation logic is as follows: ; right Each value in implement .

[0023] The Z-score-based standardization process mentioned above refers to a preprocessing method that transforms data of different magnitudes into a unified standard score. By subtracting the mean and dividing by the standard deviation, the processed data conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the influence of data magnitude on subsequent feature extraction.

[0024] Assuming we select the central region of K-space Taking a local real data block as an example, the data set is... Total number of sampling points ; First, calculate the global average. : ; Then calculate the variance and global standard deviation. :variance Standard deviation ; Finally, calculate the first data point. Standardized values: ; This result indicates that the original numerical values After standardization, it is transformed into a negative value approximately 1.3867 standard deviations away from the mean, effectively eliminating the influence of dimensions.

[0025] Regarding the mapping and interpolation process in S13, the system uses the FOV size and target image matrix size obtained in S11. Establish a two-dimensional Cartesian grid coordinate system; The standardized K-space sampling points are mapped to grid nodes according to their frequency encoding and phase encoding gradient values; For unsampled grid points, i.e. missing values, caused by non-Cartesian sampling, the system uses a bilinear interpolation algorithm to fill them; Set the interpolation search radius to Each pixel unit, for coordinates Search for the Euclidean distance of the missing value at the specified location. The four nearest known sampling points within the range ; Get the values ​​of these 4 points and its distance from the target point ; Calculate weights and normalize the weights ; The target point value is calculated as follows ; After filling in all missing values, the system sorts them from row 0 to row 1. Rows, each row from column 0 to column 1 The column indexing order reorganizes discrete values ​​in memory into a compact two-dimensional tensor structure; Real-part tensors are labeled as real-part data matrices. The imaginary part of the data tensor is labeled as an imaginary part matrix. .

[0026] The bilinear interpolation algorithm mentioned above is a mathematical method for estimating the value of an unknown point in a two-dimensional grid by performing linear interpolation in two directions. It uses the values ​​of four known sampling points around it and their spatial relationship to calculate the value of the target position by weighting, thereby ensuring the continuity of the image data.

[0027] Please see Figure 1 and Figure 3 S2: Initialize the weight parameters of the real part convolution kernel and the weight parameters of the imaginary part convolution kernel. Convolve the real part data matrix with the weight parameters of the real part convolution kernel to generate the first intermediate feature map. Convolve the imaginary part data matrix with the weight parameters of the imaginary part convolution kernel to generate the second intermediate feature map. Convolve the real part data matrix with the weight parameters of the imaginary part convolution kernel to generate the third intermediate feature map. Convolve the imaginary part data matrix with the weight parameters of the real part convolution kernel to generate the fourth intermediate feature map. The specific steps of S2 are as follows: S21: Set the kernel size, stride parameter and zero padding number of the convolutional neural network layer. Use the Glorot uniform distribution initialization strategy to generate random floating-point values ​​that conform to the normal distribution law, and assign the random floating-point values ​​to the weight variables of the real part channel and the imaginary part channel respectively. Initialize the weight parameters of the real part convolutional kernel and the imaginary part convolutional kernel. S22: Call the two-dimensional convolution operator in the deep learning computing framework, slide the real part convolution kernel weight parameters on the real part data matrix with a set stride and perform a dot product summation operation to generate the first intermediate feature map, slide the imaginary part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the second intermediate feature map, slide the imaginary part convolution kernel weight parameters on the real part data matrix and perform a dot product summation operation to generate the third intermediate feature map, and simultaneously slide the real part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the fourth intermediate feature map; The dot product summation process for S22 includes: Based on a pre-defined sliding window mechanism, local receptive field data blocks corresponding to the convolution kernel size are extracted from the input matrix. A weighted sum between the local receptive field data blocks and their corresponding convolution kernel weights is calculated using matrix multiplication. A pre-defined bias term is then introduced to linearly correct the weighted sum. The calculation logic satisfies the following formula: ; Generate feature map pixel values; in, Represents the coordinates in the output feature map The pixel value at that location, and These represent the height and width of the convolution kernel, respectively. The stride parameter represents the convolution operation. This represents the value of the input matrix at the corresponding position in the sliding window. Represents the position of the convolution kernel The weight value at the location, This represents the bias parameter used to adjust the activation threshold.

[0028] Regarding the parameter initialization process in S21, the system first sets the hyperparameters of the convolutional layer: Kernel height ,width Input channel number Number of output channels Based on the Glorot uniform distribution initialization strategy, calculate the boundary range for random number generation; definition ,definition ; Calculate the distribution limit value : ; The system calls the pseudo-random number generator, within the interval Generate a sequence of random floating-point numbers that conforms to a uniform distribution. These random values ​​are then filled into the real part of the convolution kernel weight tensor. and imaginary part convolution kernel weight tensor middle; Here, the real and imaginary channels are initialized with independent random seeds to ensure the diversity of initial feature extraction.

[0029] The Glorot uniform distribution initialization strategy mentioned above refers to a weight initialization method that aims to maintain the uniformity of the variance of activation values ​​in each layer of a neural network. By dynamically setting the range of random weight values ​​according to the number of input and output connections, it prevents the problems of gradient vanishing or gradient explosion in the early stages of deep network training.

[0030] For the process of performing convolution and dot product summation in S22, the system constructs four parallel convolution calculation paths in the computation graph architecture; Taking the generation of the first intermediate feature map by convolving the real part data matrix with the real part convolution kernel weight parameters as an example, the execution logic is explained in detail; Set step size parameters Zero fill quantity This means filling the edges of the input matrix with zero values ​​to maintain its size; The system will Real part of the convolution kernel Covering the real data matrix top left corner coordinates ; According to the formula Perform the operation to generate feature map pixel values.

[0031] in, Represents the coordinates in the output feature map The pixel value at that location; and These represent the height and width of the convolutional kernel, respectively, and are set to 3 here; The stride parameter represents the convolution operation, and here it is set to 1; This represents the value of the input matrix at the corresponding sliding window position, used to provide local receptive field data; Represents the position of the convolution kernel The weight values ​​at each location are used to extract specific spatial feature patterns; This represents the bias parameter used to adjust the activation threshold, initialized to 0.0.

[0032] Selecting the input matrix One Local receptive field and initialization convolution kernel Perform calculations; bias term Set as ; The input data and weight data are shown in Table 1 and Table 2.

[0033] Table 1. Input matrix local region data table; As shown in Table 1, the local region of the input matrix contains 9 standardized floating-point values, reflecting the local feature distribution in the K-space or image domain.

[0034] Table 2. Convolution kernel weight data; As shown in Table 2, the convolution kernel weight parameters contain 9 initialized floating-point values, which are used to extract features from the input data.

[0035] Will and Multiply and sum the elements at corresponding positions: ; Calculate item by item: Row0 is ; Row1 item is ; Row2 item is ; Summation ; Substitute the bias term into the formula: ; The result shows that the matching response intensity of this local region of the input matrix with the convolution kernel feature pattern is 0.95, and this value will be written into the corresponding pixel position of the first intermediate feature map. The system repeats the above process, utilizing and Generate the first intermediate feature map ; use and Generate the second intermediate feature map ; use and Generate the third intermediate feature map ; use and Generate the fourth intermediate feature map ; All intermediate feature maps are temporarily stored in the video memory buffer.

[0036] Please see Figure 1 and Figure 4 S3: Calculate the difference between the first intermediate feature map and the second intermediate feature map to generate the real part of the output feature map, and calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate the imaginary part of the output feature map. The specific steps for S3 are as follows: S31: Obtain the first intermediate feature map and the second intermediate feature map, perform a point-by-point subtraction operation at the matrix element level, subtract the value of the corresponding position in the second intermediate feature map from the value of each pixel position in the first intermediate feature map, and generate the output feature real part map. The point-by-point subtraction operation process of S31 includes: Obtain the first pixel matrix in the first intermediate feature map and the second pixel matrix in the second intermediate feature map, check whether their dimensions are consistent. If they are consistent, start a parallel computing thread, perform arithmetic difference calculation on the floating-point value of each corresponding coordinate position, use the calculated difference to construct the current matrix element, and write it into the target memory address to generate the output feature real part map. S32: Obtain the third and fourth intermediate feature maps, perform point-by-point addition operations at the matrix element level, and superimpose the value of each pixel position in the third intermediate feature map with the value of the corresponding position in the fourth intermediate feature map to generate the output imaginary feature map.

[0037] Regarding the process of generating the real part graph of the output features in S31, this step aims to simulate the real part calculation rules of complex multiplication; First intermediate feature map The second intermediate feature map is the product of the real part of the input and the real part of the weights. The product term corresponding to the imaginary part input and the imaginary part weight; The system first locks from the video memory address. and The memory block; verify whether the dimensions of the two are strictly consistent, both being memory blocks. After successful verification, initiate the parallel computing command. For coordinates in the feature map Read each point numerical value and numerical value ; Perform arithmetic subtraction: ; Calculation results Write to a new memory address and construct the output feature real part map. .

[0038] Continuing the example from S2, assume that in coordinates At that point, the calculation result of the first intermediate feature map is as described above. ; Assume the value obtained by convolution at the same position in the second intermediate feature map is ; The value of the real part of the output feature map at this position is: ; The result shows that after processing the real part operation logic of complex field convolution, the real part feature response value after fusion at this position is 0.50.

[0039] For the process of generating the imaginary part graph of the output feature in S32, this step simulates the imaginary part calculation rules of complex multiplication; The third intermediate feature map The fourth intermediate feature map corresponds to the product of the real input and the imaginary weights. The product term corresponding to the imaginary part input and the real part weight; System read and For the data, perform element-wise addition operations at the matrix element level; for each corresponding coordinate Read the value and ; Perform arithmetic addition: ; Calculation results Write to the target memory region to construct the output feature imaginary part graph. .

[0040] The pointwise addition operation mentioned above refers to the mathematical operation process of adding the values ​​of corresponding elements between two matrices or tensors of the same dimension to generate a new matrix or tensor of the same dimension, which is used to fuse information from different feature channels.

[0041] Assuming in coordinates At this location, the value of the third intermediate feature map is ; Assume the value of the fourth intermediate feature map is ; The value of the output feature's imaginary part at that position is: ; The result indicates that the imaginary part feature response value after fusion at this location is 0.60; At this point, the system has completed a single-layer complex convolution operation, generating a dataset containing... and Complex feature pairs.

[0042] Please see Figure 1 and Figure 5S4: By performing linear rectification nonlinear activation mapping on the real part graph and the imaginary part graph of the output features, values ​​less than zero are set to zero and values ​​greater than or equal to zero are retained to generate deep complex feature vectors. The specific steps for S4 are as follows: S41: Perform nonlinear rectification operations on the real part map and the imaginary part map of the output features respectively. Filter out and retain pixel values ​​greater than or equal to zero in the matrix through comparison operation, and forcibly replace all pixel values ​​less than zero in the matrix with zero values ​​to generate the activated real part feature map and imaginary part feature map. The nonlinear rectification operation process of S41 includes: The activation threshold of the linear rectifier unit is set to zero. The value of each pixel in the input feature map is traversed. The sign attribute of the value is identified by the condition judgment instruction. If the value is negative, the gradient propagation path corresponding to it is blocked and the output value is set to zero. If the value is positive, its original linear response characteristics are kept unchanged, and the corrected feature map is generated. S42: Perform global average pooling on the activated real and imaginary feature maps, calculate the pixel mean in multiple feature channels to compress the spatial dimension, and then concatenate and fuse the compressed real and imaginary feature data according to the channel dimension. The fused multidimensional data is flattened into a one-dimensional array through a fully connected layer to generate a deep complex feature vector.

[0043] For the process of performing nonlinear rectification in S41, the system analyzes the real part diagram of the output characteristic curve generated in S3. With output feature imaginary part graph Process it; This operation aims to introduce nonlinear factors to enhance the model's ability to express complex features; Set the activation threshold to ; System loading Enter the calculation unit and iterate through each floating-point value within it. ; Execution condition judgment logic: judgment Is it valid? If true, then perform the zeroing operation and output the value. And mark the gradient at that position as 0 during backpropagation; If not immediately If the original value is not found, the output value will remain unchanged. ; The system Processing to generate activated real feature maps ,right Processing to generate the activated imaginary part feature map .

[0044] The aforementioned nonlinear rectification operation refers to a nonlinear transformation process that uses the linear rectified unit ReLU as the activation function to map all negative inputs to zero while keeping positive inputs unchanged. This operation can effectively solve the gradient vanishing problem in deep network training and increase the sparsity of the network.

[0045] Referencing the results from the S3 example: Output feature real part map in The value at is The output feature map is in The value at is ; Assume the output feature real part map is on another coordinate system. The value at is ; Perform ReLU activation: Location Real part: The result is retained as ; Location Virtual part: The result is retained as ; Location Real part: The result is replaced with ; This result indicates that the location The negative response feature is suppressed, eliminating its interference with subsequent levels.

[0046] Regarding the global average pooling and vector generation process in S42, the system first performs... and Perform spatial dimension compression; Assuming the feature map size is The number of channels is ; For the Each channel calculates all values ​​within that channel. The arithmetic mean of the pixel values; the calculation formula is: ; Through this operation, the two-dimensional feature map of each channel is compressed into a scalar value; The real feature map is converted into a 32-dimensional real feature vector. The imaginary feature map is converted into a 32-dimensional imaginary feature vector. ; Subsequently, the system performs splicing and fusion; Will and Connect along the channel dimension; Define the fused vector Its length is ; index to storage The value, index to storage The value; Finally, the data flows through the fully connected layer; In this embodiment, the 64-dimensional vector is directly output as a deep complex feature vector for subsequent image classification or reconstruction tasks.

[0047] The aforementioned global average pooling refers to a feature dimensionality reduction operation that compresses a three-dimensional feature map tensor into a one-dimensional feature vector by calculating the average value of the feature map in the spatial dimension, thereby reducing the number of model parameters and enhancing the model's robustness to spatial transformations.

[0048] Assuming that after global average pooling, a portion of the channel data is extracted as shown in Table 3.

[0049] Table 3. Numerical table of feature vectors after pooling; Table 3 shows the average real and imaginary parts of some channels after pooling.

[0050] After performing the concatenation operation, the resulting deep complex feature vector The first few elements are: Index0 is 0.45, which comes from the real part Ch0; Index1 is 0.00, from the real part Ch1; Index31 is 1.05, from the real part Ch31; Index32 is 0.12, from the imaginary part Ch0; Index33 is 0.88, from the imaginary part Ch1; Index63 is 0.00, which comes from the imaginary part Ch31; This result shows that the real and imaginary features in a physical sense are integrated into the same high-dimensional vector space, thus fully preserving the amplitude and phase statistical characteristics of the magnetic resonance signal.

[0051] A deep learning-based magnetic resonance image feature extraction system is provided. This system is used to execute the aforementioned deep learning-based magnetic resonance image feature extraction method. The system includes: The data acquisition and preprocessing module is used to acquire raw K-space data, extract the real and imaginary numerical components of the raw K-space data, and perform data formatting and standardization to construct the real and imaginary data matrices. The complex convolution operation module is used to initialize the real part convolution kernel weight parameters and the imaginary part convolution kernel weight parameters, convolve the real part data matrix with the real part convolution kernel weight parameters to generate the first intermediate feature map, convolve the imaginary part data matrix with the imaginary convolution kernel weight parameters to generate the second intermediate feature map, convolve the real part data matrix with the imaginary convolution kernel weight parameters to generate the third intermediate feature map, and convolve the imaginary part data matrix with the real convolution kernel weight parameters to generate the fourth intermediate feature map. The feature difference fusion module is used to calculate the difference between the first intermediate feature map and the second intermediate feature map to generate the output real part map, and to calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate the output imaginary part map. The nonlinear activation mapping module is used to generate deep complex feature vectors by performing linear rectification nonlinear activation mapping on the real part graph and the imaginary part graph of the output features, setting values ​​less than zero to zero and retaining values ​​greater than or equal to zero.

[0052] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A deep learning-based method for extracting features from magnetic resonance images, characterized in that, Includes the following steps: S1: Obtain the original K-space data, extract the real part and imaginary part of the original K-space data, and construct the real part data matrix and the imaginary part data matrix; S2: Initialize the real part convolution kernel weight parameters and the imaginary part convolution kernel weight parameters. Convolve the real part data matrix with the real part convolution kernel weight parameters to generate a first intermediate feature map. Convolve the imaginary part data matrix with the imaginary part convolution kernel weight parameters to generate a second intermediate feature map. Convolve the real part data matrix with the imaginary part convolution kernel weight parameters to generate a third intermediate feature map. Convolve the imaginary part data matrix with the real part convolution kernel weight parameters to generate a fourth intermediate feature map. S3: Calculate the difference between the first intermediate feature map and the second intermediate feature map to generate the real part of the output feature map, and calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate the imaginary part of the output feature map; S4: By performing linear rectification nonlinear activation mapping on the real part graph of the output features and the imaginary part graph of the output features, values ​​less than zero are set to zero and values ​​greater than or equal to zero are retained to generate a deep complex feature vector.

2. The deep learning-based magnetic resonance image feature extraction method according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: The radio frequency receiving coil of the magnetic resonance imaging equipment is used to collect echo signals under differentiated frequency encoding and phase encoding. The analog signal is converted into a discrete complex domain numerical sequence through the analog-to-digital conversion interface. The header information of the sequence is analyzed to determine the imaging field size and sampling point density, and the original K-space data is obtained. S12: Traverse each complex sampling point in the original K-space data, use the complex separation algorithm to extract the real part scalar value and imaginary part scalar value corresponding to multiple sampling points, and perform Z-score-based standardization processing on the extracted real part and imaginary part values ​​to generate standardized real part and imaginary part values. S13: Based on the imaging field size, the standardized real and imaginary numerical components are mapped to a two-dimensional grid coordinate system, and the missing values ​​in the unsampled area are filled by bilinear interpolation. The processed component values ​​are then reorganized into a two-dimensional tensor structure according to the row and column index order to construct the real data matrix and the imaginary data matrix respectively.

3. The deep learning-based magnetic resonance image feature extraction method according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: Set the kernel size, stride parameter and zero padding number of the convolutional neural network layer, use the Glorot uniform distribution initialization strategy to generate random floating-point values ​​that conform to the normal distribution law, and assign the random floating-point values ​​to the weight variables of the real part channel and the imaginary part channel respectively, and initialize the weight parameters of the real part convolutional kernel and the weight parameters of the imaginary part convolutional kernel. S22: Invoke the two-dimensional convolution operator in the deep learning computing framework, slide the real part convolution kernel weight parameters on the real part data matrix with a set stride and perform a dot product summation operation to generate the first intermediate feature map, slide the imaginary part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the second intermediate feature map, slide the imaginary part convolution kernel weight parameters on the real part data matrix and perform a dot product summation operation to generate the third intermediate feature map, and simultaneously slide the real part convolution kernel weight parameters on the imaginary part data matrix and perform a dot product summation operation to generate the fourth intermediate feature map.

4. The deep learning-based magnetic resonance image feature extraction method according to claim 1, characterized in that, The specific steps of S3 are as follows: S31: Obtain the first intermediate feature map and the second intermediate feature map, perform a point-by-point subtraction operation at the matrix element level, subtract the value of the corresponding position in the second intermediate feature map from the value of each pixel position in the first intermediate feature map, and generate the output feature real part map. S32: Obtain the third intermediate feature map and the fourth intermediate feature map, perform a point-by-point addition operation at the matrix element level, and superimpose the value of each pixel position of the third intermediate feature map with the value of the corresponding position of the fourth intermediate feature map to generate the output feature imaginary part map.

5. The deep learning-based magnetic resonance image feature extraction method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Perform nonlinear rectification operations on the real part map and the imaginary part map of the output features respectively, filter out and retain pixel values ​​greater than or equal to zero in the matrix through comparison operation, and forcibly replace all pixel values ​​less than zero in the matrix with zero values ​​to generate the activated real part feature map and imaginary part feature map. S42: Perform global average pooling on the activated real feature map and imaginary feature map, calculate the pixel mean in multiple feature channels to compress the spatial dimension, and concatenate and fuse the compressed real feature data and imaginary feature data according to the channel dimension. Flatten the fused multidimensional data into a one-dimensional array through a fully connected layer to generate the deep complex feature vector.

6. The deep learning-based magnetic resonance image feature extraction method according to claim 2, characterized in that, The standardization process of S12 includes: The distribution of modulus values ​​of all sampling points in the original K-space data is statistically analyzed. The global mean and global standard deviation of the distribution are calculated. The real and imaginary numerical components are center-shifted using the global mean. The shifted component values ​​are scaled and normalized using the global standard deviation. A statistical parameter mapping table for inverse standardization is established to generate standardized component data.

7. The deep learning-based magnetic resonance image feature extraction method according to claim 3, characterized in that, The dot product summation process of S22 includes: Based on a pre-defined sliding window mechanism, local receptive field data blocks corresponding to the convolution kernel size are extracted from the input matrix. A weighted sum between the local receptive field data blocks and their corresponding convolution kernel weights is calculated using matrix multiplication. A pre-defined bias term is then introduced to linearly correct the weighted sum. The calculation logic satisfies the following formula: ； Generate feature map pixel values; in, Represents the coordinates in the output feature map The pixel value at that location, and These represent the height and width of the convolution kernel, respectively. The stride parameter represents the convolution operation. This represents the value of the input matrix at the corresponding position in the sliding window. Represents the position of the convolution kernel The weight value at the location, This represents the bias parameter used to adjust the activation threshold.

8. The deep learning-based magnetic resonance image feature extraction method according to claim 4, characterized in that, The point-by-point subtraction operation process of S31 includes: Obtain the first pixel matrix in the first intermediate feature map and the second pixel matrix in the second intermediate feature map, check whether their dimensions are consistent. If they are consistent, start a parallel computing thread, perform arithmetic difference calculation on the floating-point value of each corresponding coordinate position, use the calculated difference to construct the current matrix element, and write it into the target memory address to generate the output feature real part map.

9. The deep learning-based magnetic resonance image feature extraction method according to claim 5, characterized in that, The nonlinear rectification operation process of S41 includes: The activation threshold of the linear rectifier unit is set to zero. The value of each pixel in the input feature map is traversed. The sign attribute of the value is identified by the condition judgment instruction. If the value is negative, the gradient propagation path corresponding to it is blocked and the output value is set to zero. If the value is positive, its original linear response characteristics are kept unchanged, and the corrected feature map is generated.

10. A deep learning-based magnetic resonance image feature extraction system, characterized in that, The system is used to implement the deep learning-based magnetic resonance image feature extraction method according to any one of claims 1-9, and the system comprises: The data acquisition and preprocessing module is used to acquire raw K-space data, extract the real and imaginary numerical components of the raw K-space data, and perform data formatting and standardization processing to construct real and imaginary data matrices. The complex convolution operation module is used to initialize the real part convolution kernel weight parameters and the imaginary part convolution kernel weight parameters, convolve the real part data matrix with the real part convolution kernel weight parameters to generate a first intermediate feature map, convolve the imaginary part data matrix with the imaginary part convolution kernel weight parameters to generate a second intermediate feature map, convolve the real part data matrix with the imaginary part convolution kernel weight parameters to generate a third intermediate feature map, and convolve the imaginary part data matrix with the real part convolution kernel weight parameters to generate a fourth intermediate feature map; The feature difference fusion module is used to calculate the difference between the first intermediate feature map and the second intermediate feature map to generate an output feature real part map, and to calculate the sum of the third intermediate feature map and the fourth intermediate feature map to generate an output feature imaginary part map. The nonlinear activation mapping module is used to generate a deep complex feature vector by performing linear rectification nonlinear activation mapping on the real part graph of the output feature and the imaginary part graph of the output feature, setting values ​​less than zero to zero and retaining values ​​greater than or equal to zero.

Images