Method and apparatus for reconstructing magnetic resonance echo data based on deep learning

Abstract

This application provides a method, apparatus, storage medium, and computer program product for reconstructing nuclear magnetic resonance (NMR) echo data based on deep learning, belonging to the field of NMR logging data processing. The method includes: extracting a target echo sequence data of a preset input length from the original echo sequence data, with the target echo sequence data located at the end of the original echo sequence data; inputting the target echo sequence data into a pre-trained echo data reconstruction model to output predicted echo sequence data; splicing the predicted echo sequence data to the end of the original echo sequence data to obtain spliced ​​echo sequence data; generating training samples based on the spliced ​​echo sequence data and retraining the echo data reconstruction model using the training samples; repeating the above steps until the spliced ​​echo sequence data reaches the preset sequence length to complete the reconstruction of the original echo sequence data. This application can improve the accuracy of reconstructed echo data.

Description

Technical Field

[0001] This application relates to the field of nuclear magnetic resonance logging data processing technology, specifically to a method and apparatus for reconstructing nuclear magnetic resonance echo data based on deep learning. Background Technology

[0002] Nuclear magnetic resonance (NMR) logging technology, with its ability to directly detect formation pore fluid information, has become an indispensable tool in the evaluation of complex oil and gas reservoirs. This technology obtains the lateral relaxation time by measuring the relaxation signal of hydrogen nuclei in the formation. T 2) Distribution spectrum, which provides key reservoir parameters such as porosity, permeability, mobile fluid saturation and bound water saturation, has wide application value in the field of oil and gas exploration and development.

[0003] During nuclear magnetic resonance (NMR) logging data acquisition, the instrument emits a sequence of radio frequency pulses into the formation, exciting hydrogen nuclei in the formation pore fluid to generate NMR signals. Subsequently, it acquires echo sequences generated by the relaxation process of these hydrogen nuclei. These echo sequences are a series of time-series, multi-exponentially decaying signals, typically containing thousands of echo points in their complete length. To obtain reliable data... T 2. Spectral inversion results indicate that complete echo sequence data needs to be collected during well logging operations, resulting in long measurement times and high power consumption per measurement.

[0004] However, cableless NMR logging instruments rely on downhole batteries for power, resulting in extremely limited energy supply. Acquiring complete echo sequences using traditional methods would severely restrict the effective logging time per run and the ability to cover long well sections continuously. Furthermore, the large amount of data storage would consume significant downhole storage resources, further increasing the system's energy consumption. Therefore, these factors have become key bottlenecks hindering the large-scale application of cableless NMR logging technology.

[0005] To address this, some scholars have proposed using compressed sensing technology to compress and reconstruct nuclear magnetic resonance (NMR) echo data. This involves reducing the data volume through non-uniform sampling of the echo data, and then reconstructing the complete echo signal using optimized algorithms. However, while this method reduces the number of sampling points to some extent, the reconstruction accuracy significantly decreases when the stratum porosity is complex, making it difficult to meet subsequent requirements. T 2. The accuracy requirements of spectral inversion for echo data. Therefore, existing technologies suffer from low accuracy in reconstructing complete echo sequence data from incomplete short echo sequence data. Summary of the Invention

[0006] The purpose of this application is to provide a method, device, storage medium, and computer program product for reconstructing nuclear magnetic resonance echo data based on deep learning, so as to solve the problem of low accuracy of echo data reconstruction in the prior art.

[0007] To achieve the above objectives, the first aspect of this application provides a method for reconstructing nuclear magnetic resonance echo data using deep learning, the method comprising: Obtain the original echo sequence data to be reconstructed; Extract target echo sequence data that meets the preset input length from the original echo sequence data. The target echo sequence data is located at the end of the original echo sequence data. The target echo sequence data is input into a pre-trained echo data reconstruction model to output predicted echo sequence data. The predicted echo sequence data is appended to the end of the original echo sequence data to obtain the spliced ​​echo sequence data; Training samples are generated based on spliced ​​echo sequence data, and the echo data reconstruction model is retrained using the training samples to obtain the retrained echo data reconstruction model. The original echo sequence data is updated to spliced ​​echo sequence data, and the pre-trained echo data reconstruction model is updated to the retrained echo data reconstruction model. The above steps are repeated until the spliced ​​echo sequence data reaches the preset sequence length to complete the reconstruction of the original echo sequence data.

[0008] In this application, generating training samples based on spliced ​​echo sequence data includes: obtaining a first sliding window that meets a preset retraining length; starting from the beginning position of the spliced ​​echo sequence data, extracting retraining input samples and retraining output samples from the spliced ​​echo sequence data based on the first sliding window; moving the first sliding window backward by a first preset sliding distance, and extracting retraining input samples and retraining output samples again at the new position; repeating the above steps until the first sliding window moves to the end position of the spliced ​​echo sequence data, and summarizing all retraining input samples and retraining output samples to obtain training samples.

[0009] In this application, the method further includes a pre-training step for the echo data reconstruction model, comprising: obtaining a second sliding window that meets a preset pre-training length; starting from the beginning position of the original echo sequence data, extracting pre-training input samples and pre-training output samples from the original echo sequence data based on the second sliding window; moving the second sliding window backward by a second preset sliding distance, and extracting pre-training input samples and pre-training output samples again at the new position; repeating the above steps until the second sliding window moves to the end position of the original echo sequence data, summarizing all pre-training input samples and pre-training output samples to obtain pre-training samples, and pre-training the echo data reconstruction model based on the pre-training samples.

[0010] In this application, the echo data reconstruction model includes: an input layer for receiving pre-trained samples; a convolutional feature extraction module connected after the input layer, comprising a first convolutional layer, a first activation function, a second convolutional layer, and a second activation function connected sequentially; the first convolutional layer performs convolution processing on the pre-trained samples to output first convolutional features; the first activation function performs nonlinear transformation on the first convolutional features to output first nonlinear features; the second convolutional layer performs convolution processing on the first nonlinear features to output second convolutional features; the second activation function performs nonlinear transformation on the second convolutional features to output convolutional output features; and an attention mechanism module connected after the convolutional feature extraction module, comprising a global average pooling layer, a first fully connected layer, a third activation function, and a second fully connected layer. The system consists of a fourth activation function and a weighting unit; a global average pooling layer is used to perform global average pooling on the convolutional output features along the sequence dimension to generate channel description vectors; a first fully connected layer and a third activation function are used to perform dimensionality reduction and nonlinear transformation on the channel description vectors to generate intermediate features; a second fully connected layer and a fourth activation function are used to perform dimensionality increase and normalization on the intermediate features to generate attention weights; a weighting unit is used to perform channel-wise dot product between the attention weights and the convolutional output features to output an enhanced feature map; a long short-term memory network layer, connected after the attention mechanism module, is used to extract temporal features from the enhanced feature map along the sequence direction to generate a temporal feature vector; and an output layer, connected after the long short-term memory network layer, is used to generate a predicted echo sequence based on the temporal feature vector.

[0011] In this application, the first convolutional layer uses a 5×1 convolutional kernel and 16 channels, and the second convolutional layer uses a 3×1 convolutional kernel and 64 channels; the long short-term memory network layer contains long short-term memory network units with 200 hidden nodes; the output layer contains a dropout layer with a dropout rate of 0.2 and a fully connected layer with the number of neurons equal to the length of the predicted echo sequence.

[0012] In this application, before extracting the target echo sequence data that meets the preset input length from the original echo sequence data, the method further includes: normalizing the original echo sequence data.

[0013] In this application, before extracting the target echo sequence data that meets the preset input length from the original echo sequence data, the method further includes: performing denoising processing on the original echo sequence data to be reconstructed when the signal-to-noise ratio of the original echo sequence data is less than a preset threshold.

[0014] A second aspect of this application provides a deep learning-based reconstruction apparatus for nuclear magnetic resonance echo data, comprising: a memory configured to store instructions; and a processor configured to retrieve instructions from the memory and, when executing the instructions, to implement the aforementioned deep learning-based reconstruction method for nuclear magnetic resonance echo data.

[0015] A third aspect of this application provides a machine-readable storage medium storing instructions for causing a machine to execute the aforementioned deep learning-based NMR echo data reconstruction method.

[0016] The fourth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned deep learning-based method for reconstructing nuclear magnetic resonance echo data.

[0017] The above technical solution first acquires the original echo sequence data to be reconstructed; then, it extracts target echo sequence data that meets a preset input length and is located at the end of the original echo sequence data. Subsequently, the target echo sequence data is input into a pre-trained echo data reconstruction model to output predicted echo sequence data; the predicted echo sequence data is then concatenated to the end of the original echo sequence data to obtain concatenated echo sequence data. Training samples are then generated based on the concatenated echo sequence data, and the echo data reconstruction model is retrained using these training samples to obtain a retrained echo data reconstruction model. This retrained model learns more content from the echo sequence data. Therefore, since the sample data used to retrain the echo data reconstruction model includes newly predicted echo sequence data, retraining the model using concatenated echo data containing newly predicted echo sequence data can yield a more accurate echo data reconstruction model. Finally, the original echo sequence data is updated with the spliced ​​echo sequence data, and the pre-trained echo data reconstruction model is updated with the retrained echo data reconstruction model. This process is repeated until the spliced ​​echo sequence data reaches the preset sequence length, thus completing the reconstruction of the original echo sequence data. Therefore, by continuously updating the spliced ​​echo sequence data and updating the retrained echo data reconstruction model with a more accurate model, a "prediction-training-re-prediction" cyclical architecture can be achieved. By continuously increasing the number of samples learned by the echo data reconstruction model, the accuracy of the model can be continuously adjusted, ultimately resulting in higher overall accuracy in the reconstruction of the original echo sequence data.

[0018] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1The schematic diagram illustrates a flowchart of a deep learning-based method for reconstructing nuclear magnetic resonance echo data according to an embodiment of this application; Figure 2 This illustration schematically shows a bimodal structure constructed according to an embodiment of this application. T 2. Spectral model diagram; Figure 3 This is a comparison chart of echo data obtained by intelligent reconstruction of echo sequence data with a signal-to-noise ratio of 30 and actual noise-free echo sequence data; Figure 4 This is a comparison chart of echo data obtained by intelligent reconstruction of echo sequence data with a signal-to-noise ratio of 20 and actual noise-free echo sequence data; Figure 5 This is a comparison chart of echo data obtained by intelligent reconstruction of echo sequence data with a signal-to-noise ratio of 10 and actual noise-free echo sequence data; Figure 6 The corresponding echo sequences were obtained by inverting echo sequence data with a signal-to-noise ratio of 30, known echo sequence data, and the complete echo sequence obtained through intelligent reconstruction. T 2 spectra; Figure 7 The corresponding echo sequences were obtained by inverting echo sequence data with a signal-to-noise ratio of 20, known echo sequence data, and the complete echo sequence obtained through intelligent reconstruction. T 2 spectra; Figure 8 The corresponding echo sequences are obtained by inverting echo sequence data with a signal-to-noise ratio of 10, known echo sequence data, and the complete echo sequence obtained through intelligent reconstruction. T 2. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0023] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0024] Figure 1 The illustration schematically shows a flowchart of a deep learning-based method for reconstructing nuclear magnetic resonance echo data according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a method for reconstructing nuclear magnetic resonance echo data based on deep learning. Taking the application of this method to a processor as an example, the reconstruction method may include the following steps: Step S101: Obtain the original echo sequence data to be reconstructed.

[0025] It is understandable that the original echo sequence data is incomplete and requires reconstruction of unknown sequence data.

[0026] Specifically, the method by which this application obtains the original echo sequence data to be reconstructed includes using a cable to send the detection instrument downhole to acquire the data.

[0027] In this embodiment of the application, the original echo sequence data to be reconstructed is used as known data to serve as the data basis for the subsequent reconstruction of the remaining data.

[0028] Step S102: Extract target echo sequence data that meets the preset input length from the original echo sequence data. The target echo sequence data is located at the end of the original echo sequence data.

[0029] It can be understood that the preset input length is the length of the echo sequence data used in the pre-trained echo data reconstruction model. The target echo sequence data is the echo sequence data input into the pre-trained echo data reconstruction model.

[0030] Specifically, this application collects target echo sequence data located at the end of the original echo sequence data, so as to predict subsequent echo sequence data of a certain length based on known echo sequence data of a preset input length.

[0031] In this embodiment of the application, since the known echo sequence data (target echo sequence data) used for prediction is a part of the original echo sequence data, the amount of data is controllable. This avoids inputting a large amount of known echo sequence data into the pre-trained echo data reconstruction model, which would otherwise result in low data prediction efficiency.

[0032] Step S103: Input the target echo sequence data into the pre-trained echo data reconstruction model to output the predicted echo sequence data.

[0033] It is understandable that the predicted echo sequence data is the echo sequence data output by the model reconstructing the echo data from the pre-trained echo data.

[0034] Specifically, this predicted echo sequence data is only a portion of the unknown echo sequence data. The original echo sequence data can be 300 bytes long. The last 150 target echo sequence data points are extracted from the original echo sequence data and used as input data for the pre-trained echo data reconstruction model to output a predicted echo sequence data of length 50 bytes.

[0035] In this embodiment, to ensure prediction efficiency, the length of the target echo sequence data input into the pre-trained echo data reconstruction model is limited to a preset input length. Therefore, the effective prediction length of the echo data reconstruction model is limited. This design can guarantee the accuracy of the echo sequence data within the effective prediction length and also ensure the prediction efficiency of the echo data reconstruction model.

[0036] Step S104: The predicted echo sequence data is spliced ​​to the end of the original echo sequence data to obtain spliced ​​echo sequence data.

[0037] It can be understood that spliced ​​echo sequence data is echo sequence data that includes both original echo sequence data and predicted echo sequence data.

[0038] Specifically, let's take an example to illustrate this step: the predicted echo sequence data of 50 data points is spliced ​​to the end of the original echo sequence data of 300 data points to obtain spliced ​​echo sequence data of 350 data points.

[0039] In the embodiments of this application, after forming spliced ​​echo sequence data, the spliced ​​echo sequence data can be used as known echo sequence data to continue predicting the remaining unknown echo data.

[0040] Step S105: Generate training samples based on the spliced ​​echo sequence data, and retrain the echo data reconstruction model using the training samples to obtain the retrained echo data reconstruction model.

[0041] Specifically, the echo data reconstruction model is retrained by generating training samples based on spliced ​​echo sequence data (350 data points in length). The training samples can be generated by taking multiple sample data of different lengths from the spliced ​​echo sequence data of 350 data points as input and output samples.

[0042] In this embodiment, since the sample data used to retrain the echo data reconstruction model includes newly predicted echo sequence data, the echo data reconstruction model learns more content from the echo sequence data. Therefore, retraining the echo data reconstruction model based on spliced ​​echo data including newly predicted echo sequence data can yield a more accurate echo data reconstruction model.

[0043] Step S106: Update the original echo sequence data to spliced ​​echo sequence data, and update the pre-trained echo data reconstruction model to the retrained echo data reconstruction model. Repeat the above steps until the spliced ​​echo sequence data reaches the preset sequence length to complete the reconstruction of the original echo sequence data.

[0044] It is understandable that the preset sequence length is the length of the complete echo sequence data corresponding to the original echo sequence data.

[0045] Specifically, for example: after retraining the echo data reconstruction model, extract echo sequence data that meets the preset input length (e.g., 150) and is located at the end of the original echo sequence data (201st echo sequence data - 350th echo sequence data) from the spliced ​​echo sequence data with a data point length of 350. The echo sequence data (data point length of 350) is input into the retrained echo data reconstruction model to output another predicted echo sequence data (data point length of 50). The other predicted echo sequence data is concatenated to the end of the concatenated echo sequence data (data point length of 350) to obtain another concatenated echo sequence data (data point length of 400). Training samples are generated based on the other concatenated echo sequence data, and the echo data reconstruction model is retrained again using the training samples to obtain the retrained echo data reconstruction model. The concatenated echo sequence data is updated with the other concatenated echo sequence data, and the retrained echo data reconstruction model is updated with the retrained echo data reconstruction model. The above steps are repeated until the echo sequence data reaches the preset sequence length to complete the reconstruction of the original echo sequence data.

[0046] In this embodiment, by continuously updating the spliced ​​echo sequence data with another spliced ​​echo sequence data, and updating the retrained echo data reconstruction model with a retrained echo data reconstruction model, a "prediction-training-re-prediction" cyclical architecture can be achieved. Simultaneously, by continuously increasing the number of samples learned by the echo data reconstruction model, the accuracy of the echo data reconstruction model can be continuously adjusted, resulting in higher overall accuracy in the reconstruction of the original echo sequence data.

[0047] The above technical solution first acquires the original echo sequence data to be reconstructed, and then extracts the target echo sequence data that meets the preset input length and is located at the end of the original echo sequence data. Subsequently, the target echo sequence data is input into a pre-trained echo data reconstruction model to output predicted echo sequence data. This predicted echo sequence data is then concatenated to the end of the original echo sequence data to obtain concatenated echo sequence data. Training samples are then generated based on the concatenated echo sequence data, and the echo data reconstruction model is retrained using these training samples to obtain a retrained echo data reconstruction model. This retrained model learns more content from the echo sequence data. Therefore, since the sample data used to retrain the echo data reconstruction model includes newly predicted echo sequence data, retraining the model using concatenated echo data containing newly predicted echo sequence data can yield a more accurate echo data reconstruction model. Finally, the original echo sequence data is updated with the spliced ​​echo sequence data, and the pre-trained echo data reconstruction model is updated with the retrained echo data reconstruction model. This process is repeated until the spliced ​​echo sequence data reaches the preset sequence length, thus completing the reconstruction of the original echo sequence data. Therefore, by continuously updating the spliced ​​echo sequence data and updating the retrained echo data reconstruction model with a more accurate model, a "prediction-training-re-prediction" cyclical architecture can be achieved. By continuously increasing the number of samples learned by the echo data reconstruction model, the accuracy of the model can be continuously adjusted, ultimately resulting in higher overall accuracy in the reconstruction of the original echo sequence data.

[0048] In one embodiment, generating training samples based on spliced ​​echo sequence data includes: obtaining a first sliding window that meets a preset retraining length; starting from the beginning position of the spliced ​​echo sequence data, extracting retraining input samples and retraining output samples from the spliced ​​echo sequence data based on the first sliding window; moving the first sliding window backward by a first preset sliding distance, and extracting retraining input samples and retraining output samples again at the new position; repeating the above steps until the first sliding window moves to the end position of the spliced ​​echo sequence data, and summarizing all retraining input samples and retraining output samples to obtain training samples.

[0049] It can be understood that the preset retraining length includes the length of the retraining input samples and the length of the retraining output samples. The retraining input samples are the samples used to retrain the echo data reconstruction model, and their length can be 150. The retraining output samples are the output samples used to train the echo data reconstruction model, and their length can be 50. The first sliding window is a sliding window constructed during the retraining phase, and it can be a truncated window with a length of 200. The first preset sliding distance is a pre-set sliding distance, which can be 1, used to move the first sliding window to capture more retraining input and output samples.

[0050] Specifically, this application obtains a first sliding window that meets a preset retraining length (which can be 200). Starting from the beginning position of the spliced ​​echo sequence data (the first echo sequence data), retraining input samples (the 1st echo sequence data to the 150th echo sequence data) and retraining output samples (the 151st echo sequence data to the 200th echo sequence data) are extracted from the spliced ​​echo sequence data based on the first sliding window. The first sliding window is then moved backward by a first preset sliding distance (which can be 1), and retraining input samples (the 2nd echo sequence data to the 151st echo sequence data) and retraining output samples (the 152nd echo sequence data to the 201st echo sequence data) are extracted again at the new position. The above steps are repeated until the first sliding window moves to the end position of the spliced ​​echo sequence data, and all retraining input samples and retraining output samples are summarized to obtain training samples.

[0051] In this embodiment, the training set is reconstructed using newly spliced ​​echo sequence data for retraining. The resulting echo data reconstruction model can then more comprehensively learn the actual attenuation characteristics of the echo sequence data in real time, ensuring the accuracy of long sequence reconstruction. Simultaneously, this mechanism reuses the predicted echo sequence data generated during the iteration process, achieving real-time updates to the echo data reconstruction model at extremely low cost, without relying on a large number of offline-accumulated labeled samples, thus improving the efficiency of echo sequence data utilization.

[0052] In one embodiment, the method further includes a pre-training step for the echo data reconstruction model, comprising: obtaining a second sliding window that meets a preset pre-training length; starting from the beginning position of the original echo sequence data, extracting pre-training input samples and pre-training output samples from the original echo sequence data based on the second sliding window; moving the second sliding window backward by a second preset sliding distance, and extracting pre-training input samples and pre-training output samples again at the new position; repeating the above steps until the second sliding window moves to the end position of the original echo sequence data, summarizing all pre-training input samples and pre-training output samples to obtain pre-training samples, and pre-training the echo data reconstruction model based on the pre-training samples.

[0053] It is understood that the preset pre-training length includes the length of the pre-training input samples and the length of the pre-training output samples. The pre-training input samples are used to pre-train the echo data reconstruction model, and their length can be 150. The pre-training output samples are used to train the echo data reconstruction model, and their length can be 50. The second sliding window is a sliding window constructed during the pre-training stage. The second sliding window can be the same as the first sliding window, and can be a truncated window with a length of 200. The second preset sliding distance is a pre-set sliding distance, which can be 1, used to move the second sliding window to capture more pre-training input and output samples.

[0054] Specifically, this application obtains a second sliding window that meets a preset pre-training length (which can be 200). Starting from the beginning position of the original echo sequence data (the first echo sequence data), pre-training input samples (the 1st echo sequence data to the 150th echo sequence data) and pre-training output samples (the 151st echo sequence data to the 200th echo sequence data) are extracted from the spliced ​​echo sequence data based on the second sliding window. The second sliding window is then moved backward by a second preset sliding distance (which can be 1), and pre-training input samples (the 2nd echo sequence data to the 151st echo sequence data) and pre-training output samples (the 152nd echo sequence data to the 201st echo sequence data) are extracted again at the new position. The above steps are repeated until the second sliding window moves to the end position of the original echo sequence data. All pre-training input samples and pre-training output samples are then summarized to obtain pre-training samples, and the echo data reconstruction model is pre-trained based on the pre-training samples.

[0055] In this embodiment of the application, the echo data reconstruction model is pre-trained using known original echo sequence data, which can ensure the accuracy of the first data predicted by the echo data reconstruction model.

[0056] In one embodiment, the echo data reconstruction model includes: an input layer for receiving pre-trained samples; a convolutional feature extraction module connected after the input layer, comprising a first convolutional layer, a first activation function, a second convolutional layer, and a second activation function connected in sequence; the first convolutional layer performs convolution processing on the pre-trained samples to output a first convolutional feature; the first activation function performs a nonlinear transformation on the first convolutional feature to output a first nonlinear feature; the second convolutional layer performs convolution processing on the first nonlinear feature to output a second convolutional feature; the second activation function performs a nonlinear transformation on the second convolutional feature to output a convolutional output feature; and an attention mechanism module connected after the convolutional feature extraction module, comprising a global average pooling layer, a first fully connected layer, a third activation function, and a second fully connected layer. The system consists of a first fully connected layer, a fourth activation function, and a weighting unit. A global average pooling layer performs global average pooling on the convolutional output features along the sequence dimension to generate channel description vectors. A first fully connected layer and a third activation function perform dimensionality reduction and nonlinear transformation on the channel description vectors to generate intermediate features. A second fully connected layer and a fourth activation function perform dimensionality increase and normalization on the intermediate features to generate attention weights. A weighting unit performs channel-wise multiplication of the attention weights with the convolutional output features to output an enhanced feature map. A long short-term memory (LSTM) network layer, connected after the attention mechanism module, extracts temporal features from the enhanced feature map along the sequence direction to generate a temporal feature vector. An output layer, connected after the LSM network layer, generates a predicted echo sequence based on the temporal feature vector.

[0057] It can be understood that the first convolutional layer is a convolutional layer used to perform convolution processing on pre-trained samples. The first activation function is an activation function that performs a nonlinear transformation on the output of the first convolutional layer. The first nonlinear feature is the output feature of the first nonlinear activation function. The second convolutional layer is a convolutional layer used to perform convolution processing on the first nonlinear feature. The second convolutional feature is the feature output by the second convolutional layer. The second nonlinear activation function is an activation function used to perform a nonlinear transformation on the second convolutional feature. The global average pooling layer is a pooling layer that performs global average pooling processing on the output of the second nonlinear activation function. In the prior art, pooling layers also include max pooling layers and global max pooling layers, while this application uses a global average pooling layer. The channel description vector is the output feature of the global average pooling layer. The intermediate features are the features output by the first fully connected layer and the third activation function. The attention weights are the output weights of the second fully connected layer and the fourth activation function. The feature map is the output feature of the weighted unit. The temporal feature vector is the output feature of the long short-term memory network layer. The predicted echo sequence is the output feature of the output layer.

[0058] Specifically, the input layer receives training samples. It is connected to the first convolutional layer in the convolutional feature extraction module. Simultaneously, the first convolutional layer, the first activation function, the second convolutional layer, and the second activation function are connected sequentially. The input to the first convolutional layer is the pre-training samples, and its output is the first convolutional feature. The input to the first activation function is the first convolutional feature, and its output is the first non-linear feature. The input to the second convolutional layer is the first non-linear feature, and its output is the second convolutional feature. The input to the second activation function is the second convolutional feature, and its output is the convolutional output feature. The second activation function in the convolutional feature extraction module is connected to the global average pooling layer of the attention mechanism module. The input to the global average pooling layer is the convolutional output feature, and its output is the channel description vector. The input to the first fully connected layer is the channel description vector, and the output of the third activation function is the intermediate feature. The output of the first fully connected layer is connected to the input of the third activation function. The input to the second fully connected layer is the intermediate feature, and the output of the fourth activation function is the attention weight. The output of the second fully connected layer is connected to the input of the fourth activation function. The input to the weighted unit is the attention weight and the convolutional output feature, and its output is the enhanced feature map. The Long Short-Term Memory (LSTM) network layer is connected to the weighted units of the attention mechanism module. The input is the enhanced feature map, and the output is a temporal feature vector. The output layer is connected after the LSM network layer, with the input being the temporal feature vector and the output being the predicted echo sequence. This embodiment also includes a loss function, using the error between the actual echo sequence and the predicted echo sequence as the loss function. The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm, and the network parameters are iteratively updated using the Adam optimization algorithm. Through multiple rounds of iterative training, the prediction error is continuously reduced, improving the model's ability to fit the NMR echo sequence. When the loss function drops to a preset threshold or reaches the maximum number of training rounds, model training is stopped, and the final echo prediction model is obtained.

[0059] In this embodiment, the local attenuation features of echo data are extracted layer by layer by the convolutional feature extraction module, the feature channels are adaptively weighted and enhanced by the attention mechanism module, and the long short-term memory network layer is used to capture the long-term dependence of echo sequence data. Finally, the output layer generates the predicted echo sequence, thus constructing a complete step from local feature extraction to global temporal modeling of echo sequence data, which effectively improves the accuracy of echo data reconstruction.

[0060] In one embodiment, the first convolutional layer uses a 5×1 convolutional kernel and 16 channels, and the second convolutional layer uses a 3×1 convolutional kernel and 64 channels; the long short-term memory network layer contains long short-term memory network units with 200 hidden nodes; the output layer contains a dropout layer with a dropout rate of 0.2 and a fully connected layer with the number of neurons equal to the length of the predicted echo sequence.

[0061] Specifically, in the echo data reconstruction model of this application, the convolutional feature extraction module adopts a two-layer one-dimensional convolutional structure: the first convolutional layer uses a 5×1 convolutional kernel with 16 channels to capture short-term decay fluctuations in the local area of ​​the echo sequence data; the second convolutional layer uses a 3×1 convolutional kernel with 64 channels to further extract more abstract second convolutional features based on the first convolutional features. Both convolutional layers scan features along the sequence dimension with a stride of 1 to ensure that no local information is missed, and the increase in the number of channels from 16 to 64 expands the feature dimension layer by layer to enhance the representation ability. The long short-term memory network layer has 200 hidden nodes, which is sufficient to carry the long-range temporal information in the echo sequence, while taking into account the computational efficiency of the model and avoiding the risk of overfitting caused by too many nodes. The output layer is a dropout layer with a dropout rate of 0.2 after the long short-term memory network layer, which randomly discards some feature dimensions to suppress overfitting. After that, a fully connected layer with the number of neurons equal to the length of the predicted echo sequence linearly maps the temporal feature vector extracted by the long short-term memory network to the predicted echo sequence.

[0062] In one embodiment, before extracting target echo sequence data that meets a preset input length from the original echo sequence data, the method further includes: normalizing the original echo sequence data.

[0063] Specifically, the normalization process involves dividing each original echo sequence data by the initial value of the original echo sequence data.

[0064] In this embodiment of the application, normalizing the original echo sequence data can reduce the amplitude difference between different samples, improve the stability of the echo reconstruction model training and the reconstruction accuracy of the original echo sequence data.

[0065] In one embodiment, before extracting target echo sequence data that meets a preset input length from the original echo sequence data, the method further includes: performing denoising processing on the original echo sequence data to be reconstructed if the signal-to-noise ratio of the original echo sequence data is less than a preset threshold.

[0066] It is understandable that the preset threshold can be 30.

[0067] Specifically, the noise reduction process includes variational mode decomposition.

[0068] In the embodiments of this application, denoising processing can reduce the impact of noise on the reconstruction of the original echo sequence, improve the stability of the high echo reconstruction model training, and enhance the reconstruction accuracy of the original echo sequence data.

[0069] This application uses the reconstruction of one-dimensional nuclear magnetic resonance data as an example to conduct numerical simulation experiments to verify the effectiveness of this application. In actual processing, for echo sequence data with a signal-to-noise ratio below 30, denoising processing should be performed before reconstruction.

[0070] The numerical simulation experiment first constructs a bimodal structure dominated by the small aperture. T 2. Spectral model, its T The peak values ​​of the two spectra correspond to 5ms and 100ms respectively, and the amplitude ratio of the corresponding portions of the small aperture and the large aperture is 6:4. (Setting...) T The number of sampling points was 128, the echo interval was 0.2ms, and the number of echoes was 2500. Then, a certain amount of Gaussian white noise was added to the forward modeling results.

[0071] Figure 2 It is a bimodal structure constructed by numerical simulation experiments. T 2.

[0072] Figures 3-5 The method involves extracting the first 300 points of echo data with different signal-to-noise ratios, reconstructing them based on the echo data reconstruction model to obtain the remaining echo data, and comparing the reconstruction results with the noise-free echo, as shown in the figure. Figure 3 This is a comparison chart of the reconstruction results of echo sequence data with a signal-to-noise ratio of 30; Figure 4 This is a comparison of the reconstruction results of echo sequence data with a signal-to-noise ratio of 20; Figure 5 This is a comparison of the reconstruction results of echo sequence data with a signal-to-noise ratio of 10.

[0073] Depend on Figures 3-5 It can be seen that the intelligent reconstruction method proposed in this application has high accuracy and is close to the constructed forward echo.

[0074] Figures 6-8 The corresponding data obtained by inverting echo sequence data T 2. Figure 6 The corresponding data is obtained by inverting echo sequence data with a signal-to-noise ratio of 30, 500 known echo sequence data, and the complete echo sequence data obtained through intelligent reconstruction. T 2 spectra; Figure 7 The corresponding data is obtained by inverting echo sequence data with a signal-to-noise ratio of 20, 500 known echo sequence data, and the complete echo sequence data obtained through intelligent reconstruction. T 2 spectra; Figure 8 It is the corresponding T2 spectrum obtained by inverting echo sequence data with a signal-to-noise ratio of 10, 500 known echo sequence data, and complete echo sequence data obtained by intelligent reconstruction.

[0075] In one embodiment, this application provides a deep learning-based nuclear magnetic resonance echo data reconstruction apparatus, comprising: a memory configured to store instructions; and a processor configured to retrieve instructions from the memory and, when executing the instructions, to implement the aforementioned deep learning-based nuclear magnetic resonance echo data reconstruction method.

[0076] In one embodiment, this application provides a machine-readable storage medium storing instructions for causing a machine to execute the aforementioned deep learning-based NMR echo data reconstruction method.

[0077] In one embodiment, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned deep learning-based nuclear magnetic resonance echo data reconstruction method.

[0078] 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 process, method, article, or apparatus. Unless otherwise specified, 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 that element.

[0079] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for reconstructing nuclear magnetic resonance echo data based on deep learning, characterized in that, The method includes: Obtain the original echo sequence data to be reconstructed; Extract a target echo sequence data that meets a preset input length from the original echo sequence data, wherein the target echo sequence data is located at the end of the original echo sequence data; The target echo sequence data is input into a pre-trained echo data reconstruction model to output predicted echo sequence data. The predicted echo sequence data is appended to the end of the original echo sequence data to obtain the appended echo sequence data; Training samples are generated based on the spliced ​​echo sequence data, and the echo data reconstruction model is retrained using the training samples to obtain the retrained echo data reconstruction model. The original echo sequence data is updated to the spliced ​​echo sequence data, and the pre-trained echo data reconstruction model is updated to the retrained echo data reconstruction model. The above steps are repeated until the spliced ​​echo sequence data reaches the preset sequence length to complete the reconstruction of the original echo sequence data.

2. The reconstruction method according to claim 1, characterized in that, Generating training samples based on the spliced ​​echo sequence data includes: Obtain the first sliding window that meets the preset retraining length; Starting from the beginning position of the spliced ​​echo sequence data, retraining input samples and retraining output samples are extracted from the spliced ​​echo sequence data based on the first sliding window; Move the first sliding window backward by a first preset sliding distance, and re-capture the retraining input sample and the retraining output sample at the new position; Repeat the above steps until the first sliding window moves to the end of the spliced ​​echo sequence data, and summarize all the retraining input samples and retraining output samples to obtain the training samples.

3. The reconstruction method according to claim 1, characterized in that, The method also includes a pre-training step for the echo data reconstruction model, including: Obtain a second sliding window that meets the preset pre-training length; Starting from the beginning position of the original echo sequence data, pre-training input samples and pre-training output samples are extracted from the original echo sequence data based on the second sliding window; The second sliding window is moved backward by a second preset sliding distance, and the pre-trained input sample and the pre-trained output sample are re-captured at the new position; Repeat the above steps until the second sliding window moves to the end of the original echo sequence data. Summarize all the pre-trained input samples and pre-trained output samples to obtain pre-trained samples, and pre-train the echo data reconstruction model based on the pre-trained samples.

4. The reconstruction method according to claim 3, characterized in that, The echo data reconstruction model includes: The input layer is used to receive pre-trained samples; A convolutional feature extraction module, connected after the input layer, includes a first convolutional layer, a first activation function, a second convolutional layer, and a second activation function connected in sequence. The first convolutional layer performs convolution processing on the pre-trained samples to output a first convolutional feature. The first activation function performs a nonlinear transformation on the first convolutional feature to output a first nonlinear feature. The second convolutional layer performs convolution processing on the first nonlinear feature to output a second convolutional feature. The second activation function performs a nonlinear transformation on the second convolutional feature to output a convolutional output feature. The attention mechanism module, connected after the convolutional feature extraction module, includes a global average pooling layer, a first fully connected layer, a third activation function, a second fully connected layer, a fourth activation function, and a weighting unit. The global average pooling layer performs global average pooling on the convolutional output features along the sequence dimension to generate channel description vectors. The first fully connected layer and the third activation function perform dimensionality reduction and non-linear transformation on the channel description vectors to generate intermediate features. The second fully connected layer and the fourth activation function perform dimensionality increase and normalization on the intermediate features to generate attention weights. The weighting unit performs channel-wise multiplication of the attention weights with the convolutional output features to output an enhanced feature map. A long short-term memory network layer, connected after the attention mechanism module, is used to extract temporal features from the enhanced feature map along the sequence direction and generate a temporal feature vector. The output layer, connected after the long short-term memory network layer, is used to generate a predicted echo sequence based on the temporal feature vector.

5. The reconstruction method according to claim 4, characterized in that, The first convolutional layer uses a 5×1 convolutional kernel and 16 channels, and the second convolutional layer uses a 3×1 convolutional kernel and 64 channels; The Long Short-Term Memory (LSTM) network layer contains LSTM network units with 200 hidden nodes; The output layer includes a dropout layer with a dropout rate of 0.2 and a fully connected layer with the number of neurons equal to the length of the predicted echo sequence.

6. The reconstruction method according to claim 1, characterized in that, Before extracting the target echo sequence data that meets the preset input length from the original echo sequence data, the process further includes: The original echo sequence data is then normalized.

7. The reconstruction method according to claim 1, characterized in that, Before extracting the target echo sequence data that meets the preset input length from the original echo sequence data, the process further includes: If the signal-to-noise ratio of the original echo sequence data is less than a preset threshold, the original echo sequence data to be reconstructed is subjected to denoising processing.

8. A reconstruction device for nuclear magnetic resonance echo data based on deep learning, characterized in that, include: The memory is configured to store instructions; as well as A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the deep learning-based method for reconstructing nuclear magnetic resonance echo data according to any one of claims 1 to 7.

9. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute the deep learning-based nuclear magnetic resonance echo data reconstruction method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the deep learning-based method for reconstructing nuclear magnetic resonance echo data according to any one of claims 1 to 7.

Images