A method for optimizing magnetic resonance scanning time based on enhanced dynamic detection probability

Abstract

The present invention discloses a method for optimizing magnetic resonance scanning time based on enhanced dynamic detection probability, by preprocessing magnetic resonance images, and then sequentially performing brain region segmentation processing, same order of magnitude processing, sliding window segmentation processing, and prediction, to obtain a first prediction sequence, and then performing spatial weighting processing on the second time series segmentation part, and splicing it into a whole brain region time series matrix, sequentially performing encoder network processing, weighted processing, decoder network processing and time series prediction to obtain a second prediction sequence, constructing a loss function through the first prediction sequence, the second prediction sequence and the first prediction sequence, and performing dynamic detection, to obtain an enhanced detection probability, and then determining the shortened length of the magnetic resonance scanning time according to the existing detection probability. The present invention shortens the scanning time of magnetic resonance by improving the dynamic detection probability, and solves the problem that some patients in the prior art cannot undergo magnetic resonance scanning for a long time.

Description

Technical Field

[0001] The present application relates to the field of brain medicine technology, and in particular to a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability. Background Art

[0002] In recent years, the non-stationarity of brain activity in the resting state has received increasing attention. Among them, studies related to the evaluation of functional connectivity (FC) in functional magnetic resonance imaging (fMRI) have shown that changes in dynamic functional connectivity (dFC) are, to some extent, neural in origin and may be related to changes in cognition or alertness, and have potential value as a lesion biomarker. In addition, studies have shown that with the increase in MRI scanning time, the probability of dFC detection will be significantly improved, the difference in dFC between the lesion group and the normal group will be amplified, and the value of dFC as a lesion diagnostic marker will be more prominent.

[0003] In addition, predicting functional magnetic resonance imaging (fMRI) time series can help shorten the scanning time of subjects, thereby reducing the pain caused by long scanning, saving time and manpower costs for hospitals, and bringing convenience to specific groups such as the elderly and mentally ill patients who are not convenient for long scanning.

[0004] However, current research mainly focuses on how to determine window-related parameters to reduce the possibility of mistaking electrophysiological noise for fluctuations in functional connectivity, or how to use the difference in dFC between the lesion group and the normal group to achieve lesion diagnosis. Few studies have focused on extending the time series to fundamentally improve the probability of dynamic detection. Summary of the invention

[0005] The present application provides a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability, so as to solve the problem in the prior art that the elderly and mentally ill patients cannot undergo magnetic resonance scanning for a long time.

[0006] In order to solve the above problems, the present application provides a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability, the method comprising the following steps:

[0007] Acquiring first original data, wherein the first original data is a resting-state functional magnetic resonance imaging image based on the cerebral cortex;

[0008] Preprocessing the first raw data to obtain a magnetic resonance image that is not affected by external conditions;

[0009] Performing brain region segmentation processing on the magnetic resonance image to obtain an average time series of voxels contained in each brain region;

[0010] Performing same-order-of-magnitude processing on the average time series to obtain a first time series with the same order of magnitude;

[0011] Performing sliding window segmentation processing on the first time series in sequence to obtain a second time series;

[0012] Performing a first prediction on the second time series to obtain a first predicted sequence;

[0013] dividing the second time series into a time series to be predicted and a time series not to be predicted, and performing spatial weighting processing on the non-prediction time series to obtain a first weighted result;

[0014] Splicing the first weighted result and the time series to be predicted into a whole brain region time series matrix;

[0015] Performing coding network processing on the whole brain area time series matrix to obtain a coding result;

[0016] Performing time-weighted processing on the encoding result to obtain a second weighted result;

[0017] Performing decoding network processing on the second weighted result and performing time series prediction to obtain a second prediction sequence;

[0018] Constructing a loss function based on the second prediction sequence and the first prediction sequence;

[0019] Performing dynamic detection according to the second time series, the first prediction series, and the second prediction series to obtain an enhanced detection probability;

[0020] A shortened length of magnetic resonance scan time is determined based on the enhanced probability of detection and the existing probability of detection.

[0021] In some embodiments, since the MRI scanning time is proportional to the probability of dynamic detection, the higher the probability of dynamic detection, the better the display effect of the lesion, so while ensuring that the MRI scanning time remains unchanged, increasing the probability of dynamic detection can effectively improve the diagnosis of the lesion, and conversely, this method can also be used to shorten the MRI scanning time. By predicting the time series of MRI, the probability of dynamic detection can be increased, and the MRI scanning time can be shortened.

[0022] Furthermore, the dynamic detection includes:

[0023] The second time series, the first prediction series and the second prediction series are subjected to discrete Fourier transformation and multiplied by uniformly distributed random phases to obtain first alternative data that eliminates dynamics;

[0024] Performing an inverse discrete Fourier transform on the first replacement data to obtain second replacement data having a time domain;

[0025] Performing windowing processing on the second replacement data and the original data, calculating the Pearson correlation coefficient of the second replacement data and the original data after the windowing processing in the same time period, and calculating the variance of the correlation coefficient of the second replacement data and the original data respectively, wherein the original data are: the second time series, the first prediction series, and the second prediction series;

[0026] constructing a null distribution according to the variance of the second replacement data, and determining whether the variance of the original data is within a preset percentile of the null distribution when the variance is zero;

[0027] If not, count the number of times the variance of the original data is not zero, and get the number of times it is not zero;

[0028] The ratio of the number of non-zero times to the number of samples in the original data is calculated to obtain the enhanced detection probability.

[0029] Further, the preprocessing of the first original data includes:

[0030] The first raw data is subjected to a first preprocessing using DPABISurf of SPM12, wherein the first preprocessing includes:

[0031] Removing the first 10 time points in the first original data to obtain the second original data;

[0032] The second raw data is sequentially subjected to time layer correction, scalp and skull removal processing, head motion correction, spatial standardization processing and smoothing processing to obtain a magnetic resonance image that is not affected by external conditions.

[0033] Furthermore, the same order of magnitude processing includes:

[0034] Performing linear normalization processing on the first time series;

[0035] The linear normalization processing formula is:

[0036]

[0037] Among them, X max is the maximum value of each input time series in the time dimension, X min The minimum value of each input time series in the time dimension.

[0038] Furthermore, the first prediction includes:

[0039] Randomly select time series with different lengths from the second time series multiple times;

[0040] The first half of the taken time series is input into the model to obtain an output result, and the second half of the taken time series is taken as the real sequence and compared with the output result to obtain the first predicted sequence.

[0041] Furthermore, the spatial weighted processing includes:

[0042] Performing a one-dimensional convolution on the non-prediction time series in the time dimension to obtain a first convolution result;

[0043] The first convolution result is weighted according to the spatial attention mechanism to obtain the first weighted result.

[0044] Furthermore, the time weighted processing includes:

[0045] Performing one-dimensional convolution on the encoding result in the spatial dimension to obtain a second convolution result;

[0046] The second convolution result is weighted according to the temporal attention mechanism to obtain a second weighted result.

[0047] Furthermore, the convolution kernel of the one-dimensional convolution is a CNN module, the number of the convolution kernels is 8, the activation function of the one-dimensional convolution is leaky_relu, the step size of the one-dimensional convolution is 1, and the spatial attention mechanism and the temporal attention mechanism are both CBAM modules.

[0048] Furthermore, the encoder network processing and the decoder network processing are both 2-layer dynamic RNN deep learning networks based on the GRU module, the loss rate of the encoder network processing is 0.5, and the activation function of the encoder network processing is tanh.

[0049] Furthermore, the encoder network processing and the decoder network processing both include a fully connected layer, the output layer dimension of the fully connected layer is 1, and the activation function of the fully connected layer is sigmoid.

[0050] The present invention obtains the resting state functional magnetic resonance imaging image of the human cerebral cortex, performs preprocessing so that it is not affected by external conditions, and sequentially performs brain region segmentation processing, same order of magnitude processing, sliding window segmentation processing, and first prediction to obtain a first prediction sequence, and then performs spatial weighted processing on the non-to-be-predicted time series segmented by the second time series, and splices it with the to-be-predicted time series to form a whole brain region time series matrix, and sequentially performs encoder network processing, weighted processing, decoder network processing and time series prediction on the whole brain region time series matrix to obtain a second prediction sequence, and constructs a loss function through the first prediction sequence, the second prediction sequence and the first prediction sequence and performs dynamic detection to obtain an enhanced detection probability, and then determines the shortened length of the magnetic resonance scanning time according to the existing detection probability. The present invention predicts the time series of magnetic resonance and improves the dynamic detection probability, thereby shortening the scanning time of magnetic resonance, and solves the problem that the elderly and mentally ill patients in the prior art cannot perform magnetic resonance scanning for a long time. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0052] Figure 1 A schematic diagram of a process for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability in the present application;

[0053] Figure 2 A schematic diagram of the preprocessing process in a method for optimizing magnetic resonance scanning time based on enhanced dynamic detection probability in the present application;

[0054] Figure 3 A schematic diagram of the process of dynamic detection in a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability in the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art of the technology described in the embodiments of the present invention.

[0057] Functional magnetic resonance imaging is a new method for studying human brain function. It is non-invasive and has high temporal and spatial resolution. It is applied in many fields of neuroscience in clarifying the functional connections between higher-level neurophysiological and neuropsychological activities and cortices, intraoperative navigation to maximize the resection of functional cortical lesions and reduce surgical complications, understanding the degree of differentiation and prognosis of brain tumors, and revealing the pathophysiological changes of abnormal cortical function in neurological and psychiatric lesions.

[0058] However, during the MRI process, patients are required to reduce their activities as much as possible, which makes the MRI detection of patients who cannot remain still for a long time inaccurate, which is bound to cause greater pain to patients and increase unnecessary time and manpower costs for medical staff. Therefore, based on how to reduce the scanning time of MRI without affecting the MRI detection effect, the present application proposes the following implementation method.

[0059] refer to Figure 1 , is a flow chart of a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability in the present application;

[0060] Depend on Figure 1 It can be seen that the present application provides a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability, and the method comprises the following steps S100-S1400:

[0061] S100, acquiring first original data, where the first original data is a resting-state functional magnetic resonance imaging image based on the cerebral cortex;

[0062] Specifically, in this embodiment, functional magnetic resonance imaging of the human brain is used as the research object, and the purpose of the resting state is to minimize a series of problems such as image instability and unclearness caused by patient movement. It should be noted that the method provided in this application is aimed at the cerebral cortex, but it is not only applicable to the cerebral cortex. If there are other needs, for example: for a series of lesion diagnosis that can be completed by functional magnetic resonance imaging of the cerebellum, pituitary gland, etc., the method provided in this application can be used.

[0063] S200, preprocessing the first raw data to obtain a magnetic resonance image that is not affected by external conditions;

[0064] Specifically, in this embodiment, although the functional magnetic resonance images of the human brain in the resting state are provided in the above step S100, there are still a series of factors that affect the operation of the present method. For example, psychiatric patients can refrain from making large movements during magnetic resonance imaging, but small movements of the brain are difficult to avoid. Therefore, the image is preprocessed before extracting the time series to reduce or avoid the influence of such factors so as to obtain more accurate information.

[0065] S300, performing brain region segmentation processing on the magnetic resonance image to obtain an average time series of voxels contained in each brain region;

[0066] Specifically, in this embodiment, the method provided by the present application reduces the scanning time of magnetic resonance imaging by improving the probability of dynamic detection. The method for improving the probability of dynamic detection is achieved by predicting the time series. Therefore, the extraction of the time series is particularly important. If all brain regions of the entire brain are studied as a whole, there will inevitably be many details that cannot be taken care of. Therefore, the entire brain is divided into 360 brain regions, and the time series is extracted for each brain region, and the average time series of the voxels contained in each brain region is obtained. The number of time points is T, and the dimension of the time series is Tx360.

[0067] S400, performing same order of magnitude processing on the average time series to obtain a first time series of the same order of magnitude;

[0068] Specifically, in this embodiment, the average time series obtained will inevitably have some values ​​that are too high or too low, and too high or too low will have a great impact on the subsequent calculations. Therefore, before performing time series prediction, it is necessary to perform the same order of magnitude processing to eliminate the values ​​that are too large and too small. The dimension of the processed time series is Tx360.

[0069] S500, performing sliding window segmentation processing on the first time series in sequence to obtain a second time series;

[0070] Specifically, in this embodiment, the sliding window technology is a common method for evaluating dynamic functional connectivity patterns at this stage, and most of the current dynamic detection probability enhancement methods based on sliding windows are aimed at selecting a suitable value for common parameters such as window length, sliding window shape, and correlation coefficient, and cannot fundamentally improve the dynamic detection probability by changing the properties of the time series. The window length is seq_len, and the time series dimension becomes (T-seq_len+1) x seq_len x 360.

[0071] S600, performing a first prediction on the second time series to obtain a first predicted sequence;

[0072] Specifically, in this embodiment, a time series of batch_size (the number of data samples captured for one training) is randomly selected from the second time series as the input for one model training, and the dimension of the time series input into the model for training each time is batch_size x seq_len x 360. For the second half of the sequence input into the model, a sequence of length pre_len is taken as the true sequence to be compared with the prediction result after the first half is input into the model. The dimension of the second time series is batch_size x (seq_len–pre_len) x 360.

[0073] S700, dividing the second time series into a time series to be predicted and a time series not to be predicted;

[0074] Specifically, in this embodiment, there are 359 time series to be predicted and 1 time series not to be predicted.

[0075] S701, performing spatial weighting processing on the non-prediction time series to obtain a first weighted result, and splicing the first weighted result and the time series to be predicted into a whole brain region time series matrix;

[0076] Specifically, in this embodiment, the method is to predict based on the time series of the entire brain area, so no time series can be missing, and the time series not to be predicted is spatially weighted to obtain spatial information, and is spliced ​​with the time series to be predicted. The time series after splicing has the time series of the entire brain area, and is converted into a matrix.

[0077] S800, performing coding network processing on the whole brain area time series matrix to obtain a coding result;

[0078] S900, performing time weighting processing on the encoding result to obtain a second weighted result;

[0079] S1000-S1100, performing decoding network processing on the second weighted result and performing time series prediction to obtain a second prediction sequence;

[0080] Specifically, in this embodiment, the time points of the pre_len length of the rear time series in the decoding result are taken as the prediction results.

[0081] S1200, constructing a loss function according to the second prediction sequence and the first prediction sequence;

[0082] Specifically, in this embodiment, the loss function constructed is:

[0083] L(X,Y)=λ adv L adv (Y')+λ p L p (Y,Y')+λ dpl L dpl (Y,Y') (1-1)

[0084] L adv (Y') = L sce (D(Y'),1) (1-2)

[0085] L sce (A,B)=-∑B i log(sigmoid(A i ))+(1-B i )log(1-sigmoid(A i )) (1-3)

[0086] L dpl (Y,Y')=|sign(Y' T+1 -Y T )-sign(Y T+1 -Y T )| (1-4)

[0087] L P (Y,Y') = |Y'-Y| (1-5);

[0088] Among them, is the overall distribution error L obtained by the GAN network adv (Y'), L p (Y,Y') is the MAE error between the true sequence and the predicted sequence, L dpl (Y,Y') is the directional error between the true sequence and the predicted sequence, Y' is the predicted sequence, Y is the true sequence, D(Y') is the probability that the discriminator network in the GAN network judges that the predicted result is Y', and sign is the sign function.

[0089] The parameters of each layer of the model are as follows:

[0090] Table 1. Details of the model architecture

[0091]

[0092]

[0093] Among them, the loss function is used to train the model while performing the subsequent operations.

[0094] S1300, inputting the second time series, the first prediction series, and the second prediction series into a loss function to perform dynamic detection, and obtaining an enhanced detection probability;

[0095] Specifically, in this embodiment, there are originally 800 time series, and on the basis of the predicted time series, a second predicted sequence is predicted, totaling 200, and the dynamic detection probability is increased from the original 46.15% to 60%. It should be noted that the first predicted sequence is not the final time series prediction result. The first predicted sequence is used as an intermediate process value in this application, and the second predicted sequence is the final predicted time series.

[0096] S1400, determining a shortened length of the magnetic resonance scanning time according to the enhanced detection probability and the existing detection probability.

[0097] Specifically, in this embodiment, the dynamic detection probability is greatly improved. Therefore, according to the extent of the improvement of the dynamic detection probability, the time of the magnetic resonance scanning can be shortened without changing the original magnetic resonance diagnosis effect.

[0098] refer to Figure 3 , is a schematic diagram of a flow chart of dynamic detection in a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability in the present application;

[0099] Depend on Figure 3 It can be seen that the dynamic detection includes:

[0100] S1301-S1303, performing discrete Fourier transform on the second time series, the first prediction series, and the second prediction series and multiplying them by uniformly distributed random phases to obtain first substitute data that eliminates dynamics;

[0101] S1304, performing an inverse discrete Fourier transform on the first replacement data to obtain second replacement data in the time domain;

[0102] Specifically, in this embodiment, the alternative data is first constructed by phase randomization method, the time series is first discrete Fourier transformed, and then multiplied by a uniformly distributed random phase, thereby eliminating the dynamics of the time series, and then inverse Fourier transform is performed to obtain the second replacement data with a time domain, thereby constructing the non-dynamic alternative data. Repeat the above process to establish 2000 alternative data samples for each brain region of the subject.

[0103] S1305-S1307, performing windowing processing on the second replacement data and the original data, calculating the Pearson correlation coefficient of the windowed second replacement data and the original data in the same time period, and calculating the variance of the correlation coefficient of the second replacement data and the original data respectively, wherein the original data are: the second time series, the first prediction series, and the second prediction series;

[0104] S1308-S1310, constructing a null distribution according to the variance of the second replacement data, and determining whether the variance of the original data is within a preset percentile of the null distribution when the variance is zero;

[0105] If not, count the number of times the variance of the original data is not zero to get the number of times it is not zero;

[0106] Specifically, in this embodiment, if the variance is not zero, it means that dynamics is detected. Therefore, by counting the number of times the variance is not zero, the number of samples that detect dynamics can be obtained; the part that determines whether the variance is zero uses the original data as a parameter.

[0107] S1311, calculate the ratio of the number of non-zero times to the number of samples in the original data to obtain an enhanced detection probability.

[0108] Specifically, in this embodiment, the second substitute data and the original data are windowed, and the Pearson correlation coefficient is calculated for the windows of the two brain regions corresponding to the same time period, and then the variance of the correlation coefficient is calculated for each second substitute data and the original data.

[0109] Then, assuming that the variance is 0, the variance of all the second alternative data of each subject constructs a null distribution belonging to the subject. If the variance of the original data falls outside the 95th percentile of the null distribution, the null hypothesis of variance can be realized, and it is considered that the subject has detected dynamics. Finally, the ratio of the number of non-zero times to the number of samples in the original data is calculated to obtain the enhanced detection probability; it should be noted that the variance of zero described in this embodiment is not only equal to zero, but can also be understood as a range of variance values ​​close to zero.

[0110] refer to Figure 3 , is a schematic diagram of a flow chart of dynamic detection in a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability in the present application;

[0111] Depend on Figure 2 It can be seen that the preprocessing of the first original data includes:

[0112] The first raw data is subjected to a first preprocessing using DPABISurf of SPM12, wherein the first preprocessing includes:

[0113] S201, removing the first 10 time points in the first original data to obtain second original data;

[0114] Specifically, in this embodiment, the SPM12 is a dynamic causal model, and the DPABISurf can perform cortical-based data preprocessing on functional magnetic resonance imaging data with one click, and calculate structural indicators such as cortical thickness, area, curvature and subcortical nucleus volume, calculate functional indicators such as cortical-based low-frequency amplitude (ALFF / fALFF), local consistency (ReHo), degree centrality and functional connectivity (FC) map, provide cortical-based image statistics and display functions, and can run independently of MATLAB.

[0115] The purpose of removing the first 10 time points in the first original data is to balance the signal and reduce or eliminate the impact caused by signal fluctuations.

[0116] S202-S207, performing time layer correction, scalp and skull stripping, head motion correction, space standardization and smoothing processing on the second original data in sequence to obtain a magnetic resonance image that is not affected by external conditions.

[0117] Specifically, in this embodiment, the time layer correction is to ensure the uniformity of the scanning time of each layer within a scanning cycle through mathematical methods such as interpolation to meet the requirements of subsequent processing; the head movement correction is to evaluate the subject's head movement and adjust the image misalignment at different times caused by it. It can also be understood that the influence of the subject's head movement is eliminated through the head movement correction.

[0118] Preferably, the same order of magnitude processing includes:

[0119] Performing linear normalization processing on the first time series;

[0120] The linear normalization processing formula is:

[0121]

[0122] Among them, X max is the maximum value of each input time series in the time dimension, X min The minimum value of each input time series in the time dimension.

[0123] Specifically, in this embodiment, the method realizes proportional scaling of the original data, where X norm is the normalized data, X is the original data, and X max , X min are the maximum and minimum values ​​of the original data set, respectively.

[0124] Preferably, the first prediction includes:

[0125] Randomly select time series with different lengths from the second time series multiple times;

[0126] The first half of the taken time series is input into the model to obtain the output result, and the second half of the taken time series is taken as the real series and compared with the output result to obtain the first predicted series;

[0127] The second time series is the first half of the taken time series.

[0128] Preferably, the spatial weighted processing includes:

[0129] Performing a one-dimensional convolution on the non-prediction time series in the time dimension to obtain a first convolution result;

[0130] The first convolution result is weighted according to the spatial attention mechanism to obtain the first weighted result.

[0131] Specifically, in this embodiment, the convolutional neural network is a type of feedforward neural network that includes convolution calculations and has a deep structure, and is one of the representative algorithms for deep learning. The convolutional neural network has representational learning capabilities and can classify input information in a translation-invariant manner according to its hierarchical structure, so it is also called a "translation-invariant artificial neural network." Among them, the one-dimensional convolution is a convolution applicable to time series, and the one-dimensional convolution moves in the same direction.

[0132] Among them, the spatial attention mechanism is one of the attention mechanisms. The so-called attention mechanism is a mechanism that focuses on local information. For example, when observing a certain image area in an image, the attention area often changes as the observation task changes. The attention mechanism is to find the most useful information, rather than all areas in the image contributing equally to the task. Only task-related areas need to be concerned, such as the main body of the classification task. The spatial attention mechanism is to find the most important parts of the network for processing. In the spatial dimension, time series at different locations influence each other, and the mutual influence is highly dynamic. Therefore, we use the attention mechanism to adaptively capture the dynamic correlation between nodes in the spatial dimension.

[0133] Preferably, the time weighted processing includes:

[0134] Performing one-dimensional convolution on the encoding result in the spatial dimension to obtain a second convolution result;

[0135] The second convolution result is weighted according to the temporal attention mechanism to obtain a second weighted result.

[0136] Specifically, in this embodiment, in the time dimension, there is a correlation between time series in different time periods, and the correlation is different in different situations. Similarly, we use the attention mechanism to adaptively assign different weights to the data. Through the above spatial attention mechanism and temporal attention mechanism, the time series integrates time information, spatial information, and global distribution information, further improving the relevance of dynamic detection.

[0137] Preferably, the convolution kernel of the one-dimensional convolution is a CNN module, the number of the convolution kernels is 8, the activation function of the one-dimensional convolution is leaky_relu, the step size of the one-dimensional convolution is 1, and the spatial attention mechanism and the temporal attention mechanism are both CBAM modules.

[0138] Specifically, in this embodiment, the full name of CNN is convolutional neural network. A neural network is a mathematical model or computational model that imitates the structure and function of a biological neural network (the central nervous system of an animal, especially the brain). A neural network is composed of a large number of artificial neurons, and different networks are constructed according to different connection methods. CNN is one of them, and there are also GAN (generative adversarial network), RNN (recurrent neural network), etc. Neural networks can have simple decision-making and simple judgment abilities similar to humans, and can give better results in image and speech recognition. The CNN structure can be divided into three layers:

[0139] 1. Convolutional layer: its main function is to extract features.

[0140] 2. Pooling layer: Its main function is to downsample without damaging the recognition results.

[0141] 3. Fully connected layer, its main function is classification.

[0142] The CBAM module is a simple and effective attention module for feedforward convolutional neural networks. Given an intermediate feature map, the CBAM module sequentially infers the attention map along two independent dimensions (channel and space), and then multiplies the attention map with the input feature map for adaptive feature optimization. Since CBAM is a lightweight general module, the overhead of the module can be ignored and it can be seamlessly integrated into any CNN architecture, and it can be trained end-to-end with the base CNN.

[0143] Preferably, the encoder network processing and the decoder network processing are both 2-layer dynamic RNN deep learning networks based on the GRU module, the loss rate of the encoder network processing is 0.5, and the activation function of the encoder network processing is tanh.

[0144] Specifically, in this embodiment, the GRU module is also called the gated recurrent unit, which is a gating mechanism in the recurrent neural network; RNN is used to process time series data. In the traditional neural network model, it is from the input layer to the hidden layer and then to the output layer. The layers are fully connected, and the nodes between each layer are disconnected. However, this ordinary neural network is powerless for many problems. For example, if you want to predict what the next word in a sentence is, you generally need to use the previous word, because the previous and next words in a sentence are not independent. The reason why RNN is called a recurrent neural network is that the current output of a sequence is also related to the previous output. The specific manifestation is that the network will remember the previous information and apply it to the calculation of the current output, that is, the nodes between the hidden layers are no longer disconnected but connected, and the input of the hidden layer includes not only the output of the input layer but also the output of the hidden layer at the previous moment. In theory, RNN can process time series data of any length.

[0145] Preferably, the encoder network processing and the decoder network processing also include a fully connected layer, the output layer dimension of the fully connected layer is 1, and the activation function of the fully connected layer is sigmoid.

[0146] Specifically, in this embodiment, the function of the fully connected layer is mainly to perform classification. The features obtained by the convolution and pooling layers are classified in the fully connected layer. The fully connected layer is a fully connected neural network. The weight of each neuron feedback is different according to the weight, and the classification result is finally obtained by adjusting the weight and the network.

[0147] The present invention proposes a method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability, which is characterized by taking the whole brain sequence as input, using a generative adversarial network to establish a loss function for the overall distribution of the prediction results, and weighting the sequence in the time dimension and the space dimension through the time attention mechanism and the space attention mechanism, so as to achieve high-precision brain function sequence prediction of the specified brain area, and apply the prediction results to improve the dynamic detection probability of the time series. Finally, 200 time points are predicted from 800 time points, the sequence length changes from 800 to 1000, and the dynamic detection probability changes from 46.15% to 60%, which is an increase of 30%. It is a successful application of deep learning in the dynamic enhancement of brain function sequences, and the shortened length of magnetic resonance scanning time can be determined according to the increase in the dynamic detection probability, so as to reduce the pain of patients who cannot scan magnetic resonance for a long time and the time and manpower cost of medical staff. In the future, this method can be applied to tasks that rely on dynamics for disease diagnosis. At the same time, due to the similarity between other modal data and resting functional magnetic resonance imaging image data, this method can also be extended to related research on other modal data in the future.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability, characterized in that: The following steps are involved: Acquiring first original data, wherein the first original data is a resting-state functional magnetic resonance imaging image based on the cerebral cortex; Preprocessing the first raw data to obtain a magnetic resonance image that is not affected by external conditions; Performing brain region segmentation processing on the magnetic resonance image to obtain an average time series of voxels contained in each brain region; Performing same-order-of-magnitude processing on the average time series to obtain a first time series with the same order of magnitude; Performing sliding window segmentation processing on the first time series in sequence to obtain a second time series; Performing a first prediction on the second time series to obtain a first predicted sequence; dividing the second time series into a time series to be predicted and a time series not to be predicted, and performing spatial weighting processing on the time series not to be predicted to obtain a first weighted result; Splicing the first weighted result and the time series to be predicted into a whole brain region time series matrix; Performing coding network processing on the whole brain area time series matrix to obtain a coding result; Performing time-weighted processing on the encoding result to obtain a second weighted result; Performing decoding network processing on the second weighted result and performing time series prediction to obtain a second prediction sequence; Constructing a loss function based on the second prediction sequence and the first prediction sequence; Performing dynamic detection according to the second time series, the first prediction series, and the second prediction series to obtain an enhanced detection probability; A shortened length of magnetic resonance scan time is determined based on the enhanced probability of detection and the existing probability of detection.

2. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The dynamic detection includes: The second time series, the first prediction series and the second prediction series are subjected to discrete Fourier transformation and multiplied by uniformly distributed random phases to obtain first alternative data that eliminates dynamics; Performing an inverse discrete Fourier transform on the first replacement data to obtain second replacement data having a time domain; Performing windowing processing on the second replacement data and the original data, calculating the Pearson correlation coefficient of the second replacement data and the original data after the windowing processing in the same time period, and calculating the variance of the correlation coefficient of the second replacement data and the original data respectively, wherein the original data are: the second time series, the first prediction series, and the second prediction series; constructing a null distribution according to the variance of the second replacement data, and determining whether the variance of the original data is within a preset percentile of the null distribution when the variance is zero; If not, count the number of times the variance of the original data is not zero, and get the number of times it is not zero; The ratio of the number of non-zero times to the number of samples in the original data is calculated to obtain the enhanced detection probability.

3. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The preprocessing of the first original data comprises: The first raw data is subjected to a first preprocessing using DPABISurf of SPM12, wherein the first preprocessing includes: Removing the first 10 time points in the first original data to obtain the second original data; The second raw data is sequentially subjected to time layer correction, scalp and skull stripping, head motion correction, spatial standardization and smoothing to obtain a magnetic resonance image that is not affected by external conditions.

4. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The same order of magnitude processing includes: Performing linear normalization processing on the first time series; The linear normalization processing formula is: Among them, X max is the maximum value of each input time series in the time dimension, X min The minimum value of each input time series in the time dimension.

5. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The first forecast includes: Randomly select time series with different lengths from the second time series multiple times; The first half of the taken time series is input into the model to obtain the output result, and the second half of the taken time series is taken as the real series and compared with the output result to obtain the first predicted series; The second time series is the first half of the taken time series.

6. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The spatial weighting process includes: Performing a one-dimensional convolution on the non-to-be-predicted time series in the time dimension to obtain a first convolution result; The first convolution result is weighted according to the spatial attention mechanism to obtain the first weighted result.

7. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 6, characterized in that: The time weighted processing includes: Performing one-dimensional convolution on the encoding result in the spatial dimension to obtain a second convolution result; The second convolution result is weighted according to the temporal attention mechanism to obtain a second weighted result.

8. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 7, characterized in that: The convolution kernel of the one-dimensional convolution is a CNN module, the number of the convolution kernels is 8, the activation function of the one-dimensional convolution is leaky_relu, the step size of the one-dimensional convolution is 1, and the spatial attention mechanism and the temporal attention mechanism are both CBAM modules.

9. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 1, characterized in that: The encoding network processing and the decoding network processing are both 2-layer dynamic RNN deep learning networks based on the GRU module. The loss rate of the encoding network processing is 0.5, and the activation function of the encoding network processing is tanh.

10. The method for optimizing magnetic resonance scanning time based on enhancing dynamic detection probability according to claim 9, characterized in that: The encoding network processing and the decoding network processing also include a fully connected layer, the output layer dimension of the fully connected layer is 1, and the activation function of the fully connected layer is sigmoid.

Images