Rapid magnetic resonance CEST quantitative method based on deep learning and imaging equipment

Abstract

The invention discloses a fast magnetic resonance CEST quantitative method based on deep learning and imaging equipment, and belongs to the technical field of magnetic resonance. The method comprises the following steps: acquiring magnetic resonance CEST data of an object to be imaged; a pre-trained deep neural network group is used, the pre-trained deep neural network group is composed of an encoder and a BM equation decoder, and frequency deviation of input data of the pre-trained deep neural network group is selected; calculating a reference z spectrum by using the deep neural network group; performing calculation to obtain a CEST quantitative signal; and obtaining a CEST quantitative signal diagram. According to the method, frequency selection is performed in advance, a plurality of key saturation frequency offset combinations are screened out, and training is performed by using a physical information network mode, so that on the premise of ensuring the quality of quantitative images, the CEST quantitative speed is improved, the quantity of data required for quantification is reduced, and the time required for imaging is shortened.

Description

Technical Field

[0001] This invention belongs to the field of magnetic resonance imaging, and in particular relates to a rapid chemical exchange saturation transfer imaging quantitative technique using deep neural networks. Background Technology

[0002] Chemical Exchange Saturation Transfer (CEST) imaging is a powerful magnetic resonance imaging (MRI) technique. Utilizing the continuous exchange of protons between a selectively saturated target substance and water protons, CEST imaging can detect the magnetic resonance signals generated by protons in the target substance. Therefore, it can non-invasively characterize substances at the molecular level, enabling imaging of information such as metabolites and pH. As a type of CEST imaging, Amide Proton Transfer (APT) imaging has shown great clinical potential in detecting various pathological conditions, including stroke, neurological diseases, lymphedema, and tumors.

[0003] However, the clinical application of CEST imaging technology still faces certain challenges. To obtain quantitative signals of the CEST effect, post-processing of the acquired multi-frame data is required, and different post-processing methods have their own advantages and disadvantages. In APT imaging, traditional model-free analysis methods, such as magnetization transfer ratio asymmetry (MTRasym) analysis, assume that the acquired z-spectrum is symmetrical and subtract the signal at +3.5 ppm from the signal at -3.5 ppm. However, the quantitative results of this method are easily affected by the nuclear overhauser effect (NOE) at -3.5 ppm. Other studies have proposed alternative methods, such as the three-point offset method, which subtracts the average signal value near the target frequency from the signal value at the target frequency. However, its application is mainly limited to high fields and may introduce interference from adjacent signals.

[0004] Besides simple differential post-processing methods, numerous studies have modeled data based on the Bloch-McConnell (BM) equations. Among these, the recently developed numerical fitting of extrapolated semisolid magnetization transfer reference (NEMR) method assumes the absence of CEST effects in far-frequency offset data and uses these effects to inversely solve the BM equations, generating a parameter map, which is then used to calculate the two-cell reference z-spectrum. The APT is obtained by subtracting the acquired z-spectrum from the reference z-spectrum at 3.5 ppm. # Contrast. This method has demonstrated superior performance in ischemic stroke detection. However, the clinical translation of NEMR is limited by its computationally intensive numerical fitting process and long computation time.

[0005] Furthermore, acquiring multiple frames of data across a large frequency offset range requires a long imaging time, which is a significant factor limiting the clinical application of CEST imaging technology. Therefore, regarding the aforementioned NEMR quantitative method, how to calculate from a smaller number of frequency offset signals while ensuring the accuracy of the reference z-spectrum calculation, thereby reducing the number of frequency offsets acquired in actual use and lowering the imaging time, while minimizing the computation time of NEMR post-processing, and achieving faster CEST signal acquisition and post-processing remains an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid magnetic resonance imaging (MRI) CEST quantitative method, device, medium, and equipment with short imaging time, good post-processing quantitative effect, fast calculation speed, and high clinical and commercial value, in order to solve the problems of long CEST imaging time and slow calculation speed of NEMR post-processing quantitative methods.

[0007] The specific technical solution adopted in this invention is as follows: In a first aspect, the present invention provides a fast magnetic resonance CEST quantitative method based on deep learning, comprising: S1. Acquire magnetic resonance CEST imaging signal data for the imaging object and extract the acquired z-spectrum from it. The saturation frequency shift acquired during CEST imaging needs to cover the combination of saturation frequency shifts and the target saturation frequency required for CEST effect quantification. The combination of saturation frequency shifts consists of multiple key saturation frequency shifts, which are selected and optimized in advance during the frequency selection stage. S2. For each voxel within the region of interest, the acquired z-spectrum containing only the combination of saturation frequency offsets and the B1 value are used as model inputs. A first deep neural network group, pre-trained under supervision, predicts the reference z-spectrum of the current predicted voxel to obtain a reference z-spectrum covering all saturation frequency offsets. The first deep neural network group is composed of a first encoder and a pre-trained BM equation decoder cascaded together, and both the first encoder and the BM equation decoder have fully connected network structures. The input of the first encoder is the model input of the current predicted voxel, and the output is the tissue parameters of the current predicted voxel. The input of the BM equation decoder is the B1 value of the current predicted voxel and the tissue parameters output by the first encoder, and the output is the reference z-spectrum corresponding to the current predicted voxel, which is composed of the normalized water signal intensity at all saturation frequency offsets. S3. For each voxel in the region of interest, subtract the signal value at the target saturation frequency on the acquired z-spectrum from the signal value at the target saturation frequency on the reference z-spectrum of that voxel to obtain the CEST quantitative signal of that voxel at the target saturation frequency offset. The CEST quantitative signals of all voxels in the region of interest are used to form a CEST effect quantitative map.

[0008] As a preferred embodiment of the first aspect above, in the frequency selection stage, a second deep neural network group is used to select and optimize all saturation frequency offsets to obtain the saturation frequency offset combination. The specific method is as follows: S11. Construct a second deep neural network group, which is composed of a second encoder and a pre-trained BM equation decoder cascaded together. The second encoder is obtained by cascading a batch normalization module and a third encoder, and both the third encoder and the BM equation decoder have fully connected network structures. The input of the second encoder is the acquired z-spectrum and B1 value corresponding to all saturation frequency offsets of the current predicted voxel, and the output is the tissue parameters of the current predicted voxel. The input of the BM equation decoder is the B1 value of the current predicted voxel and the tissue parameters output by the second encoder, and the output is the reference z-spectrum corresponding to the current predicted voxel, which is composed of the normalized water signal intensity at all saturation frequency offsets. S12. Fix the parameters of the BM equation decoder and perform supervised training on the second encoder in the second deep neural network group with the goal of minimizing the total loss function; the total loss function is obtained by weighting the mean square error loss term and the regularization loss term, the mean square error loss term is the mean square error between the predicted value and the true value of the reference z spectrum, and the regularization loss term is the sum of the absolute values ​​of the scaling factors corresponding to all saturation frequency offsets in the batch normalization module; S13. After training is completed, extract the scaling factor corresponding to all saturation frequency offsets in the batch normalization module, remove saturation frequency offsets with scaling factors lower than the screening threshold, and use the remaining saturation frequency offsets as key saturation frequency offsets to form the saturation frequency offset combination.

[0009] As a preferred embodiment of the first aspect above, in the training sample set for supervised training of the second deep neural network group, each training sample includes a acquired z-spectrum, the B1 value corresponding to the acquired z-spectrum, and a reference z-spectrum calculated using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method. Furthermore, the acquired z-spectrum input to the second deep neural network group must include all the saturation frequency offsets required for calculation by the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method.

[0010] As a preferred embodiment of the first aspect above, the BM equation decoder is pre-trained using a simulation-generated CEST dataset, and the method for obtaining the simulation-generated CEST dataset is as follows: A two-cell Bloch-McConnell mathematical model, including a water molecule cell and a semi-solid magnetization transfer cell, was established. The tissue parameters and B1 values ​​were sampled within their respective ranges to obtain combinations of sampled values. Each combination of sampled values ​​was input into the two-cell Bloch-McConnell mathematical model. The reference z-spectrum for that combination of sampled values ​​was obtained using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR). This resulted in training samples with the sampled combinations of tissue parameters and B1 values ​​as model input and the reference z-spectrum as the ground truth label, which were then added to the CEST dataset.

[0011] As a preferred embodiment of the first aspect, when supervising the training of the first deep neural network group, the parameters of the BM equation decoder are fixed, and the mean square error between the predicted value and the true value of the reference z spectrum is used as the loss function to optimize the parameters of the first encoder by minimizing the loss function.

[0012] As a preferred embodiment of the first aspect above, in the training sample set for supervised training of the first deep neural network group, each training sample includes a acquired z-spectrum, a B1 value corresponding to the acquired z-spectrum, and a reference z-spectrum calculated using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method, and the acquired z-spectrum input to the first deep neural network group only contains the saturation frequency offset combination.

[0013] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the deep learning-based fast magnetic resonance CEST quantitative method as described in any of the solutions of the first aspect above.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables the implementation of the deep learning-based fast magnetic resonance CEST quantitative method as described in any of the solutions of the first aspect above.

[0015] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the deep learning-based fast magnetic resonance CEST quantitative method as described in any of the first aspects above.

[0016] Fifthly, the present invention provides a magnetic resonance imaging device, which includes a magnetic resonance scanner and a control unit; The magnetic resonance scanner is used to scan the imaging object and acquire magnetic resonance CEST imaging signal data of the imaging object under all CEST saturation frequency offsets. The control unit stores a computer program, which, when executed, reads the magnetic resonance CEST imaging signal data acquired by the magnetic resonance scanner and generates a quantitative map of the CEST effect in the region of interest according to the deep learning-based fast magnetic resonance CEST quantification method described in any of the first aspects above.

[0017] Compared with the prior art, the present invention has the following advantages: (1) This invention combines deep learning algorithms and NEMR post-processing quantitative algorithms. On the one hand, by using neural network calculation methods, compared with NEMR post-processing quantitative methods, it can significantly improve the quantitative calculation speed while ensuring calculation accuracy. On the other hand, this invention can be directly used for post-processing quantitative calculation of multi-frame CEST data collected clinically, and has good clinical and commercial application value.

[0018] (2) This invention fully guarantees the reliability of post-processing quantitative results. This invention is designed according to the implementation process of the NEMR post-processing quantitative algorithm, trains encoder and BM equation decoder models and forms a deep neural network group, inputs the collected data into the encoder to obtain tissue parameters, and then inputs the tissue parameters into the BM equation decoder to obtain the reference z-spectrum, and uses the calculation error of the reference z-spectrum as a loss function for constraint. Physical information from the BM equation is introduced in the calculation process. Therefore, compared with the existing methods, the reference z-spectrum calculation results of this invention can achieve more reliable CEST quantification.

[0019] (3) This invention proposes an input signal frequency selection method based on deep learning technology, which can automatically select the optimal input signal frequency according to the adjustment of different regularization coefficients. While ensuring the accuracy of calculation, it reduces the number of frequency offsets used in subsequent processing. In addition, the obtained optimal input signal frequency can also provide guidance for the selection of frequency offset during imaging scanning, thereby achieving the purpose of reducing imaging time. Attached Figure Description

[0020] Figure 1 The flowchart shows the steps of the fast magnetic resonance CEST quantitative method based on deep learning. Figure 2 This is a flowchart illustrating the overall usage of frequency selection and CEST quantification in this invention. Figure 3 Methods for constructing datasets for network training; Figure 4 This diagram illustrates the network structure, operating principle, and a frequency selection result of the frequency selection part of this invention. (a) shows the network structure of the frequency selection part and how the frequency selection encoder processes the input data. (b) shows, from left to right, a frequency selection result with a regularization coefficient of 0.003, and the relationship between the number of selected frequency points and the regularization coefficient. Figure 5This example shows the post-processing quantitative results for a patient with ischemic stroke. The first row, from left to right, shows the quantitative results of the NEMR method (denoted as NEMR), the quantitative results of the proposed method (denoted as Proposed), the error map of the proposed quantitative results relative to the NEMR method (denoted as Error Map), and the corresponding normalized root mean square error (nRMSE). The second row shows the reference z-spectrum results calculated by the NEMR method and the proposed method from the patient's representative voxel data, from left to right, representing representative voxels from the contralateral white matter, gray matter, and lesion region. The specific locations of the representative voxels used are marked in the NEMR method quantitative results map in the first row and correspond with the same symbols. The quantitative index in the figure is the mean absolute error (MAE). Figure 6 This figure shows the post-processing quantitative results for four ischemic stroke patients in the example. From left to right, each patient represents a different patient. The first row shows the quantitative results using the NEMR method (denoted as NEMR); the second row shows the quantitative results using the proposed method; and the third row shows the error map (denoted as ErrorMap) between the proposed and NEMR quantitative results. The first two rows indicate the computation time used for the entire graph by the NEMR method and the proposed method, respectively. The third row indicates the normalized root mean square error (nRMSE) between the quantitative results of the two methods. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0022] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature.

[0023] like Figure 1As shown, as a preferred implementation of the present invention, a rapid magnetic resonance imaging (MRI) CEST quantitative method based on deep learning is provided, comprising steps S1 to S3. This method can be used for post-processing and quantitative analysis of clinically acquired CEST data. In this method, multiple key saturation frequency shifts are pre-optimized and screened from all saturation frequency shifts, and input into a deep neural network to predict a reference z-spectrum containing all saturation frequency shifts, thereby calculating the CEST quantitative signal. This method can improve the speed of CEST quantitative analysis while ensuring the quality of the quantitative image, and reduce the amount of data required for quantitative analysis and the imaging time required in practical applications. The specific implementation of steps S1 to S3 is described in detail below.

[0024] S1. Acquire magnetic resonance CEST imaging signal data for the imaging object and extract the acquired z-spectrum from it. The saturation frequency shift acquired during CEST imaging must cover the combination of saturation frequency shifts and the target saturation frequency required for CEST effect quantification.

[0025] It is important to note that in step S1, the signal data for all CEST saturation frequency offsets of the target object can be acquired either offline or online. For offline acquisition, the signal data can simply be retrieved from the storage device storing it; for online acquisition, a magnetic resonance imaging (MRI) device is needed to perform a CEST imaging scan on the target object to obtain the signal data.

[0026] It should be noted that the aforementioned imaging object can be any object that can be imaged by magnetic resonance CEST, such as a patient's brain.

[0027] It should be noted that the above-mentioned combination of saturated frequency offsets consists of multiple key saturated frequency offsets. Before executing step S1, a frequency selection stage is required to filter key saturated frequency offsets from all saturated frequency offsets. This combination of saturated frequency offsets can be obtained by pre-selecting and optimizing all saturated frequency offsets during the frequency selection stage. All saturated frequency offsets need to cover all saturated frequency offset inputs used in the NEMR post-processing method.

[0028] S2. For each voxel in the region of interest, the acquired z-spectrum containing only the combination of the saturation frequency offsets and the B1 value are used as model inputs. The first deep neural network group, which has been pre-trained under supervision, predicts the reference z-spectrum of the current prediction voxel to obtain a reference z-spectrum covering all saturation frequency offsets.

[0029] It should be noted that the region of interest in this invention refers to the area within the voxel domain of magnetic resonance CEST imaging that requires CEST quantification. This region can be selected according to the actual CEST quantification needs and is not limited thereto. The data processing flow executed in the first deep neural network group described above is performed on a per-voxel basis.

[0030] Before detailing the deep neural network group used in this invention, a brief introduction to the basic theory of the numerical fitting of extrapolated semisolid magnetization transfer reference (NEMR) method (hereinafter referred to as the NEMR post-processing method) will be given to facilitate understanding of the data processing flow of the deep neural network group in this invention, as follows: The NEMR post-processing method used in this invention is mainly based on the two-cell BM equation, which includes a water cell and a semisolid magnetization transfer (MT) cell: in Indicates a pool, Indicates the MT pool; They represent respectively Chihe pool Components of the magnetization vector in three directions; It is the amplitude of a saturated radio frequency (RF) pulse; and These are the radio frequency offsets relative to the water tank and the MT tank, respectively; From Pool to The exchange rate of the pool, and From Pool to The exchange rate of the pool; yes The longitudinal relaxation rate of the pool, yes The transverse relaxation rate of the pool, yes The longitudinal relaxation rate of the pool, yes The transverse relaxation rate of the pool, , , , They are respectively Longitudinal relaxation time of the pool, The lateral relaxation time of the pool, Longitudinal relaxation time of the pool, The lateral relaxation time of the pool; and They are Chihe The equilibrium magnetization vector of the pool, and satisfies the relation .

[0031] Due to the MT pool Very short, therefore assume and It can quickly reach a steady state, so equations (4) and (5) can be rewritten as equations (7) and (8): Therefore, we have the following equation (9): Redefining the saturation pulse caused The longitudinal magnetization vector loss rate of the pool is: Then according to equation (9), we have .

[0032] By rearranging equations (7) to (10) above and combining them with the original two-pool BM equations, we can obtain: The system of equations (11)-(14) above can be expressed in matrix form: in, , , Therefore, equation (15) can be solved by numerical iteration: in , ; It is a time variable; It is an increment of time; Represents the matrix index.

[0033] The above NEMR post-processing method assumes that the signal at far frequency offsets does not exhibit the CEST effect, and based on the modified BM equation, uses a nonlinear least squares fitting method to obtain the above... Chihe The cell organization parameters are obtained and then input into the modified BM equation, which extrapolates the signal at the far frequency offset to all frequency offsets, resulting in a two-cell reference z-spectrum containing only the effects of water and the MT cell, without the CEST effect. Finally, the CEST effect is extracted by subtracting the acquired z-spectrum from the obtained reference z-spectrum. Therefore, based on the basic theory and processing flow of the above NEMR post-processing method, a deep neural network group can be established and optimized.

[0034] The following describes in detail the specific network structure, internal data processing, and network training process of the first deep neural network group used in step S2 above.

[0035] See Figure 2 As shown, the aforementioned first deep neural network group includes a first encoder and a BM equation decoder, with the BM equation decoder cascaded after the first encoder. The deep neural network structure of the first encoder is a fully connected network, used to obtain tissue parameters from the z-spectrum and B1 values ​​composed of key saturation frequency offsets. Its model input is the acquired z-spectrum and B1 values ​​corresponding only to the saturation frequency offset combination in the current prediction voxel, and the output is the tissue parameters of the current prediction voxel. The tissue parameters include several key parameters that can be used to predict the reference z-spectrum. In the embodiments of this invention, the tissue parameters include... longitudinal relaxation time of the pool , lateral relaxation time of the pool , longitudinal relaxation time of the pool , The equilibrium magnetization vector of the pool ,from Pool to Pool exchange rate There are a total of 5 parameters. The deep neural network structure of the BM equation decoder described above is also a fully connected network, used to obtain the reference z-spectrum composed of all saturation frequency offsets from the B1 value and tissue parameters. Its input is the B1 value of the current predicted voxel and the tissue parameters output by the first encoder, and its output is the reference z-spectrum corresponding to the current predicted voxel, composed of the normalized water signal intensity at all saturation frequency offsets. Therefore, the above first encoder and BM equation decoder can be expressed by the following formula: in, The tissue parameters predicted by the first encoder; The mapping relationship represented by the first encoder; The z-spectrum signal of the first encoder is used as input, where It is a combination of saturated frequency offsets, which contains multiple key saturated frequency offsets obtained by pre-selecting and optimizing all saturated frequency offsets during the frequency selection stage. The value represents the intensity of the corresponding radio frequency pulse saturation power; These are all the learnable parameters in the first encoder; The reference z-spectrum is composed of all saturation frequency offsets predicted by the BM equation decoder, and the reference z-spectrum is composed of the normalized water signal intensity corresponding to all saturation frequency offsets. This represents the mapping relationship expressed by the BM equation decoder; These are all the learnable parameters in the BM equation decoder.

[0036] In the first deep neural network group, following the first encoder, the BM equation is introduced in the form of a neural network. This means that a pre-trained neural network achieves the parameter mapping function that originally required the BM equation. This pre-trained neural network is the aforementioned BM equation decoder. The BM equation decoder based on a deep neural network effectively solves the computational bottleneck in training the encoder using a deep neural network group. Traditional BM equations require multiple numerical iterations to solve the z-spectrum. Applying traditional BM equations to the encoder training of a deep neural network group leads to an excessively long gradient backpropagation chain, significantly increasing training time. This method, however, directly learns the mapping relationship through a neural network, directly mapping the tissue parameters of a voxel. and By inputting the values ​​into the BM equation decoder, a reference z-spectrum containing all saturation frequency shifts can be obtained, thereby significantly improving training efficiency.

[0037] The pre-training of the BM equation decoder mentioned above relies on a training dataset constructed from large-scale samples. However, in real-world scenarios, the true values ​​of most tissue parameters are unavailable, and collecting a large-scale combination of multiple tissue parameter values ​​is also impossible. Therefore, as another preferred embodiment of the present invention, the BM equation decoder is trained using a simulation-generated CEST dataset. The method for obtaining the simulation-generated CEST dataset is as follows: Establish a water molecule pool ( (Pool), semi-solid magnetization transfer cell ( The model uses a two-pool Bloch-McConnell mathematical model (also known as the BM equation), which includes the tissue parameter and the B1 value. Then, it samples the tissue parameter and the B1 value within the range of their values ​​to obtain the sampled value combination of the tissue parameter and the B1 value. Each sampled value combination is input into the two-pool Bloch-McConnell mathematical model, and the reference z-spectrum under the sampled value combination is obtained according to the NEMR post-processing method. Thus, the training samples with the sampled value combination of the tissue parameter and the B1 value as the model input and the reference z-spectrum as the ground truth label are obtained and added to the CEST dataset.

[0038] It should be noted that due to tissue parameters There are 5 in total, therefore each group of sampled values ​​obtained is a combination of... The sample actually includes 6 parameters. During sampling, the range of values ​​for the tissue parameters and B1 value is determined by the range of values ​​for each parameter itself, which can be determined based on literature values ​​or empirical values. Sampling methods can include uniform sampling, random sampling, Latin hypercube sampling, etc. The final combination of all sampled values ​​can be denoted as... ,in Used to count samples of combinations of sampled values. A set is obtained. Then, fixing the scan parameters in the two-pool Bloch-McConnell mathematical model, Each set of sampled values ​​is input into a two-pool Bloch-McConnell mathematical model to obtain the target reference z-spectrum. Finally, the set is... Each set of sampled values ​​corresponds one-to-one with the obtained reference z-spectrum, resulting in a CEST dataset containing various combinations of tissue parameters, B1 values, and the target reference z-spectrum for a single sample.

[0039] A fully connected network can be trained using the CEST dataset. The input to the fully connected network is designed as a set of organization parameters and B1 values, and the output is designed as a reference z-spectrum containing all saturation frequency shifts. The optimal number of layers in the fully connected network is four. During training, the mean squared error (MSE) between the true and predicted values ​​of the reference z-spectrum can be used as the loss function. The fully connected network is trained and, once trained, can be used as the BM equation decoder. Thus, by pre-training the BM equation decoder, the weights of the pre-trained BM equation decoder can be fixed and combined into subsequent first and second deep neural network groups to further train the encoder.

[0040] For training the first deep neural network group, the first encoder in the form of a fully connected network can be concatenated with the pre-trained BM equation decoder, and then supervised training can be performed on the training sample set. The supervised training freezes the learnable parameters of the BM equation decoder so that they do not participate in the parameter update in the backpropagation stage, while the learnable parameters in the first encoder are unfrozen and participate in the parameter update in the backpropagation stage, thereby achieving the optimization of the first encoder.

[0041] The training sample set for supervised training of the first deep neural network group can be constructed using the acquired and computed CEST data samples. Specifically, each training sample in the training sample set includes an acquired z-spectrum, the B1 value corresponding to the acquired z-spectrum, and a reference z-spectrum computed using the NEMR post-processing method. The saturation frequency offset of the acquired z-spectrum input to the first deep neural network group only includes the combination of the saturation frequency offsets. Other non-critical saturation frequencies, except for the critical saturation frequencies, do not need to be input into the first deep neural network group.

[0042] like Figure 3 The diagram illustrates the specific construction process of a single training sample in the aforementioned training sample set. The specific steps are as follows: 1) Acquire signal data at all CEST saturation frequency offsets collected by the CEST magnetic resonance imaging program, and process the data to extract the z-spectrum of each voxel within the region of interest. To distinguish it from the reference z-spectrum, this extracted z-spectrum is the acquired z-spectrum. The saturation frequency of the acquired z-spectrum includes all saturation frequency offsets, that is, it includes not only the critical saturation frequency but also other non-critical saturation frequencies. The acquired z-spectrum is composed of the normalized water signal intensity corresponding to all saturation frequency offsets.

[0043] 2) For each voxel within the region of interest, the B1 value of that voxel is read from the B1 image acquired by CEST magnetic resonance imaging. Then, based on the B1 value and the acquired z-spectrum, the tissue parameters are obtained by performing least-squares fitting on the BM equation using traditional NEMR post-processing methods. The fitted tissue parameters are then substituted into the BM equation to calculate the reference z-spectrum corresponding to that voxel. Similarly, the saturation frequencies of this reference z-spectrum include all saturation frequency shifts, i.e., it includes not only the critical saturation frequencies but also other non-critical saturation frequencies. The reference z-spectrum is composed of the normalized water signal intensity corresponding to all saturation frequency shifts.

[0044] 3) Pair the acquired z-spectrum, reference z-spectrum, and B1 value of each voxel to obtain a training sample.

[0045] However, it should be noted that although the z-spectrum collected in the training sample contains all saturation frequency offsets, when inputting the first deep neural network group as training data, only the z-spectrum data corresponding to the key saturation frequencies in the saturation frequency offset combination needs to be included. The z-spectrum data corresponding to other non-key saturation frequencies do not need to be input into the first deep neural network group.

[0046] As mentioned earlier, the above combination of saturation frequency offsets consists of multiple key saturation frequency offsets. See also... Figure 2 As shown, this combination of saturated frequency offsets needs to be pre-selected and optimized from all saturated frequency offsets during the frequency selection stage. Once the combination of saturated frequency offsets is determined, only the acquired z-spectrum data corresponding to these key saturated frequency offsets are needed during the inference stage to predict the reference z-spectrum containing all saturated frequency offsets. Therefore, by reducing the number of saturated frequency offset points input to the first encoder, the scanning time required in the actual CEST imaging process can be reduced.

[0047] Theoretically, a series of existing data dimensionality reduction or importance filtering methods can be used to select and optimize all saturated frequency offsets during the frequency selection stage, filtering out key saturated frequency offsets to form the aforementioned saturated frequency offset combination. However, in another preferred embodiment of the present invention, a second deep neural network group corresponding to the first deep neural network group can be introduced during the frequency selection stage to select and optimize all saturated frequency offsets to obtain the aforementioned saturated frequency offset combination. The specific structure of the second deep neural network and the frequency offset selection optimization method are described in detail below. The specific method is as follows: S11. Construct a second deep neural network group, which is formed by cascading a second encoder and a pre-trained BM equation decoder. The second encoder is obtained by cascading a batch normalization module and a third encoder. The network structures of the third encoder and the BM equation decoder are both fully connected networks.

[0048] like Figure 4As shown in (a), the input to the second encoder is the acquired z-spectrum and B1 value corresponding to all saturation frequency offsets of the current predicted voxel, and the output is the tissue parameters of the current predicted voxel; while the input to the BM equation decoder is the B1 value of the current predicted voxel and the tissue parameters output by the second encoder, and the output is the reference z-spectrum corresponding to the current predicted voxel, composed of the normalized water signal intensity at all saturation frequency offsets. Therefore, the main difference between the second deep neural network group and the first deep neural network group is that an additional batch normalization module needs to be introduced into the first encoder. The rest of the network structure is basically the same, but due to the difference in the number of input saturation frequency offsets, the input layer dimensions of the first and third encoders will also differ. The learnable parameters of the BM equation decoders in both deep neural network groups are fixed to the optimal parameters obtained during the pre-training phase, and both only need to optimize the encoder part during training. The introduced batch normalization module is a single-layer BatchNorm layer, which is used to select a smaller number of saturation frequency offset inputs from the basic saturation frequency offset inputs.

[0049] S12. Based on the constructed second deep neural network group, with the parameters of the BM equation decoder fixed, supervised training is performed on the second encoder in the second deep neural network group with the objective of minimizing the total loss function. Here, the total loss function is obtained by weighting a mean squared error loss term and a regularization loss term. The mean squared error loss term is the mean squared error between the predicted and true values ​​of the reference z-spectrum, while the regularization loss term is the sum of the absolute values ​​of the scaling factors corresponding to all saturation frequency offsets in the batch normalization module.

[0050] To facilitate understanding, specific formulas are introduced below to explain the frequency selection principle of the second deep neural network group incorporating the batch normalization module. In the batch normalization module, the BatchNorm layer normalizes, linearly scales, and shifts the data for each saturation frequency offset in the input. Since the input to the second deep neural network group consists of all saturation frequency offsets and the B1 value, assuming the total number of saturation frequency offsets is... The batch normalization module actually includes Input For each training batch of the second deep neural network group, BatchNorm will process each original input... Perform the following normalization, linear scaling, and translation processing: In the formula: , Represents the B1 value. ~ represent The saturation frequency offset of the original acquisition; Enter the batch data as input The average value; Enter the batch data as input Standard deviation; and These are the scaling factor and translation factor of BatchNorm, respectively, both of which are learnable parameters.

[0051] During training of the second deep neural network group, the BM equation decoder is frozen to keep its weights fixed, and batch training is performed only on the second encoder. Furthermore, to select for all saturation frequency shifts of the input, the total loss function of this invention employs a composite loss function specifically designed for frequency selection, obtained by weighting a mean squared error loss term and a regularization loss term: in This represents the set of all learnable parameters of the second encoder in the second deep neural network group. Including the second deep neural network group All learnable parameters outside the third encoder, i.e., all learnable parameters in the third encoder. This represents the learnable parameters of the batch normalization module, i.e. One scaling factor; This is the index of the training samples in the current batch. This represents the total number of training samples in the current batch. This represents the output of the second deep neural network group. Group reference z spectrum, Indicates the first Truth labels for the group reference z-spectrum; To The regularization coefficients used for regularization are hyperparameters that can be optimized and adjusted.

[0052] It should be noted that the composite loss function shown in equation (21) above... In the first part, the mean squared error loss term is the loss function used for training the first deep neural network group. ,and It is based on which all were introduced The regularization coefficient is obtained as a regularization term. The second deep neural network group is trained using the composite loss function shown in equation (21), and the regularization coefficient can be continuously adjusted. The size of the regularization coefficient is determined and trained separately. When using a smaller regularization coefficient... At that time, the second deep neural network group adjusts the scaling parameters. The punishment was minor, therefore the final result was... Most of the values ​​are not close to 0; when using a large regularization coefficient At that time, the second deep neural network group adjusts the scaling parameters. The penalty is large, maximizing the penalty scaling parameter while ensuring the accuracy of the network prediction results. Therefore, the final result is There will be many values ​​close to 0. The scaling parameter for the batch normalization layer is selected based on the frequency distribution at the end of training. The value of the scaling factor. Including near-zero saturation frequency offsets in the set ,gather The saturation frequency input signal The non-critical saturation frequency offsets, which are considered to contribute little to the predictions of the second deep neural network group, can be removed; the remaining scaling factors... Saturation frequency offset that is not close to 0 This is considered a key saturation frequency offset that significantly contributes to the predictions of the second deep neural network group. In embodiments of the present invention, a filtering threshold can be set. To judge Does the corresponding scaling factor satisfy...? .

[0053] S13. After training is completed, extract the scaling factor corresponding to all saturation frequency offsets in the batch normalization module, remove saturation frequency offsets with scaling factors lower than the screening threshold, and use the remaining saturation frequency offsets as key saturation frequency offsets to form the saturation frequency offset combination.

[0054] Therefore, the above set of non-critical saturation frequencies was eliminated. It can be represented as The combination of saturation frequency shifts formed by the remaining key saturation frequency shifts can be expressed as: Filtering threshold This is a hyperparameter, and a positive value close to 0 needs to be selected as the threshold. The larger the value, the more saturation frequency shifts are filtered out, and vice versa.

[0055] Similarly, in the training sample set for supervised training of the second deep neural network group, each training sample includes the acquired z-spectrum, the corresponding B1 value, and the reference z-spectrum calculated using the NEMR post-processing method. However, unlike the training of the first deep neural network group, the acquired z-spectrum input to the second deep neural network group must include all the saturation frequency offsets required for calculation by the NEMR post-processing method. That is, assuming the total number of original acquired saturation frequency offsets is... Then the z-spectrum collected by the second deep neural network group needs to cover these... A saturation frequency offset, but the z-spectrum acquired by the input to the first deep neural network group does not need to cover this. For each saturation frequency offset, only the combination of saturation frequency offsets needs to be covered. That's fine. However, the training sample set for training the second deep neural network group can actually be the same as the training sample set for training the first deep neural network group. The only difference is that the saturation frequency offset in the acquired z-spectrum needs to be adjusted when training different deep neural network groups.

[0056] The training and inference application process of the first deep neural network combined with the second deep neural network group is as follows: Figure 2 As shown, the base frequency points represent all saturation frequency offsets, while the selected frequency points represent the optimized key saturation frequency offsets. This is achieved by training the first deep neural network group to obtain the optimal set of network parameters. Then, the actual CEST effect can be quantified through the first deep neural network group containing the optimal parameters. The specific method will be explained in step S3.

[0057] S3. For each voxel in the region of interest, subtract the signal value at the target saturation frequency on the reference z-spectrum (output by the first deep neural network group in S2) of the voxel from the signal value at the target saturation frequency on the acquired z-spectrum obtained in S1 to obtain the CEST quantitative signal of the voxel at the target saturation frequency offset. The CEST quantitative signals of all voxels in the region of interest are used to form the CEST effect quantitative map.

[0058] It should be noted that the aforementioned target saturation frequency offset refers to the saturation frequency offset required for CEST quantification. Taking the APT signal at 3.5 ppm as an example, this quantitative signal is obtained by subtracting the signal values ​​of the reference z-spectrum and the acquired z-spectrum at a saturation frequency offset of 3.5 ppm. Therefore, its corresponding target saturation frequency offset is 3.5 ppm. Thus, for each voxel, the reference z-spectrum output by the first deep neural network group in S2 contains the same number of saturation frequency offsets as the acquired z-spectrum obtained in S1. The reference z-spectrum obtained in S3 can be directly subtracted from the acquired z-spectrum obtained in S1 to extract a differential signal spectral line used to extract the CEST quantitative signal. The APT signal at 3.5 ppm in this differential signal spectral line is the voxel's CEST quantitative signal. in The APT signal at 3.5 ppm is used for quantitative analysis. The reference z-spectrum is the output of the first deep neural network group. The z-spectrum acquired by CEST imaging.

[0059] Of course, the above differential signal spectrum can be used to extract CEST quantitative signals under multiple different target saturation frequency offsets, but this differential signal spectrum is not necessarily required. If only the quantitative signal of a certain target saturation frequency offset needs to be extracted, the signal value of the target saturation frequency offset can be extracted directly from the reference z-spectrum and the acquired z-spectrum and then subtracted.

[0060] After extracting the CEST quantitative signal on each voxel within the region of interest of the imaging object, the target CEST effect quantitative map of the imaging object can be obtained.

[0061] The following describes the specific technical effects of the methods described in S1 to S3 above, combined with specific embodiments, so that those skilled in the art can better understand the essence of the present invention.

[0062] Example The deep learning-based fast magnetic resonance CEST quantitative method described in S1-S3 above will be applied to a specific embodiment to demonstrate its technical effects. The specific framework and process described in S1-S3 above are as previously stated, and will not be repeated in full in this embodiment. The following focuses on demonstrating the specific implementation details and technical effects of each step.

[0063] 1. Data Preparation 1.1 Simulation of Training Data for BM Equation Decoder A two-cell Bloch-McConnell mathematical model, including a water molecule cell and a semi-solid magnetization transfer cell, was established. The combinations of tissue parameters were iterated within a pre-defined range, while the B1 value was randomly selected within the same pre-defined range, resulting in multiple sets of tissue parameters and B1 values. By fixing the scan parameters in the BM equations, Each set of tissue parameters and B1 value is input into the BM equation to obtain the target reference z-spectrum. Finally, the set is... By corresponding one-to-one with the obtained reference z-spectrum, a dataset containing various combinations of tissue parameters, B1 values, and target reference z-spectrums is obtained.

[0064] Following the steps described above, this embodiment generated a total of 20,000,000 training samples, each containing corresponding tissue parameters, B1 and reference z-spectral data.

[0065] 1.2 MRI Data Acquisition In order to train a first deep neural network group using key saturation frequency offset as input, test the generalization performance of the trained first deep neural network group, and evaluate the effectiveness of frequency selection and network training, this embodiment acquires CEST data through MR experiments and divides it into training and validation data and test data according to the imaging object.

[0066] Specifically, the brains of 35 ischemic stroke patients were scanned using a 3T Siemens scanner (MAGNETOM Skyra; Siemens Healthcare, Erlangen, Germany). The source images were acquired using DWI sequences, with a total of 20 slices and b-values ​​of 0 and 1000 s / mm. 2 The scanner automatically calculates and generates the corresponding ADC map. T2-weighted liquid attenuation inversion recovery (FLAIR) images were acquired at the same locations as the DWI images. The CEST sequence consists of four modules: frequency stabilization, saturation, fat suppression, and readout. The scanner's instantaneous B0 drift is tracked and corrected in real time by the frequency stabilization module. The saturation module contains ten 100 ms Gaussian pulses, separated by a pulse gradient of 10 mT / m for 5 ms. The readout module uses a fast spin echo (TSE) sequence with the following parameters: TR / TE = 3000 / 7.2 ms, FOV = 211 × 185.5 mm. 2 The matrix size is 96×84, the turbine coefficient is 84, and the layer thickness is 5 mm.

[0067] To normalize the signal, a single-layer fast spin echo (CEST) image was acquired at the far-off resonant frequency (>1500 ppm), and the remaining 62 saturation frequency offsets (in ppm) were acquired sequentially as follows: [15.65, 0, 0.25, -0.25, 0.5, -0.5, 0.75, -0.75, 1.0, -1.0, 1.5, -1.5, 2.0, -2.0, 2.5, -2.5, 2.5, -2.5, -2.5, 3.0, -3.0, 3.0, -3.0, 3.25, -3.25, 3.25, -3.25, 3.5, -3. ... -3.5, 3.75, -3.75, 3.75, -3.75, 4.0, -4.0, 4.0, -4.0, 4.5, -4.5, 5.0, -5.0, 6.0, -6.0, 10, 20, 30, 40, 50, 60, 70, 80).

[0068] CEST data at saturation power is The images were acquired under the specified conditions, with a scan time of 3.2 minutes. The field of view (FOV) and matrix size of the B0 and B1 images were the same as those of the CEST images. Specifically, the B0 image was acquired using a dual-echo gradient echo sequence (TE=4.92 and 9.84 ms), while the B1 image was generated using the preprocessing RF method provided by the manufacturer.

[0069] 1.3 Construction of Training Dataset for Deep Neural Networks To perform frequency selection and train the second deep neural network group, it is necessary to construct training, validation, and test sets using data acquired by MR.

[0070] In this embodiment, the MR acquisition data is preprocessed. First, regions of interest (ROIs) are delineated for all patients' data. The data within the ROIs are then post-processed using the NEMR method to obtain the reference z-spectrum for each voxel. Next, data from five patients are randomly selected as the test set, with the remaining patient data used for training and validation. In the training and validation phase, a total of 307,746 data sets from 30 patients are used. Each data set is obtained from one voxel and includes one acquired z-spectrum, one reference z-spectrum, and the corresponding B1 value. 90% of the data is randomly selected for the training set, and the remaining 10% is used for the validation set.

[0071] Thus, each independent data sample includes a set of acquired z-spectrums, a set of reference z-spectrums, and corresponding B1 values. The data and B1 values ​​at partial frequency shifts of the acquired z-spectrums will be used as inputs for frequency selection and training of the deep neural network group, while the reference z-spectrums will be used as training labels (Ground Truth).

[0072] 2. Model building and training All the deep neural networks mentioned below are built using the deep learning framework PyTorch.

[0073] 2.1 Construction and Training of the BM Equation Decoder To obtain the BM equation decoder, a four-layer fully connected neural network model was constructed and trained using the simulated CEST dataset. Training employed the Adam optimizer, minimizing the mean squared error (MSE) between the output reference z-spectrum and the ground truth label of the reference z-spectrum in the dataset. Training was performed for 500 iterations on a single NVIDIA RTX 3090 Ti GPU. A validation set was used to check the model's performance during training. Upon completion, the BM equation decoder was obtained and used to construct a first deep neural network group combined with a second deep neural network group.

[0074] 2.2 Construction and Training of the Second Deep Neural Network Group A fully connected neural network model with four layers is constructed, and a BatchNorm module is added before this fully connected neural network model to form the aforementioned second encoder. The trained BM equation decoder is then concatenated after this second encoder to form the second deep neural network group. The parameters of the BM equation decoder are fixed, and supervised training is performed using the training dataset. The Adam optimizer is used to train the second encoder by minimizing the loss function shown in equation (21). Training is performed on a single NVIDIA RTX 3090 Ti GPU for 2000 iterations. During training, a validation set is used to check the model training status. Simultaneously, the regularization coefficient is adjusted... The size of the BatchNorm module after statistical training. The size of the key saturation frequency offset is used to obtain the key saturation frequency offset optimization results of multiple selected different saturation frequency offset numbers and different combinations of saturation frequency offsets, and these results are used as key saturation frequency offsets for training the first deep neural network group.

[0075] 2.3 Construction and Training of the First Deep Neural Network Group A fully connected neural network model was constructed as the first encoder, followed by a pre-trained BM equation decoder to form the first deep neural network group. Training was performed using a collected training dataset for the deep neural network group, where the parameters of the BM equation decoder were fixed, and only the first encoder could be updated via backpropagation. Training employed the Adam optimizer, minimizing the mean squared error (MSE) between the output reference z-spectrum and the ground truth label of the reference z-spectrum in the dataset. The input frequencies of the z-spectrum data used for training were adjusted based on the scaling factor optimization results of the BatchNorm module of the second deep neural network group, eliminating saturation offset frequencies with scaling factors less than or close to a threshold of 0. The first deep neural network group was trained for 2000 iterations on a single NVIDIA RTX 3090 Ti GPU. A validation set was used to check the model's training progress during the process.

[0076] 3. Data post-processing After the first deep neural network group is trained, it can be used for post-processing analysis of the collected data. Specifically, in this embodiment, the collected z-spectrum and B1 value of the corresponding frequency of the patients in the test set are used as the input of the first deep neural network group, and the output of the first deep neural network group is the corresponding reference z-spectrum. The collected z-spectrum is subtracted from the reference z-spectrum to obtain the CEST quantitative result. Specifically, in this embodiment, for example, equation (22) is used to subtract the reference z-spectrum signal value at 3.5 ppm from the collected z-spectrum signal value to obtain the APT. #The quantitative results were obtained. Simultaneously, to compare and evaluate the performance of the quantitative method of this invention, the NEMR algorithm was used to post-process the patient data in the test set to obtain APT. # The graph is then compared with the post-processed results of the first deep neural network group. As a special case in this embodiment, regularization coefficients are primarily used here. The input is compared with a first deep neural network group whose key saturation frequency offset is 6.

[0077] 4. Results Analysis Figure 4 (b) The regularization coefficients in the quantitative method of the present invention are shown from left to right. The graph shows the frequency selection results when the value is 0.003, the relationship between the number of selected frequency points and the regularization coefficient.

[0078] Figure 5 The present invention demonstrates the quantitative method of the present invention for quantifying APT in a patient with ischemic stroke using CEST. # The results show that the quantitative method of this invention differs little from the NEMR algorithm. The reference z-spectrums calculated from representative voxel data by both methods also agree well, demonstrating the computational accuracy of the quantitative method of this invention.

[0079] Figure 6 The quantitative method of the present invention was demonstrated to quantify APT in four additional ischemic stroke patients in the test set using CEST. # Results and corresponding quantification times. The experimental results show that, except for some areas of scalp and cerebrospinal fluid that were not completely removed, the quantitative results of the present invention and the NEMR method are very similar. Furthermore, the quantification time indicates that the quantitative method of the present invention can significantly improve the quantification speed, greatly enhancing its potential for clinical application.

[0080] It should be noted that this embodiment only shows the regularization coefficient. The quantitative results of a deep neural network group with 6 input frequencies on the test data are shown. Other deep neural network groups with different input frequencies can also achieve similar quantitative results as demonstrated in this embodiment.

[0081] It should be noted that the method steps shown in S1 to S3 above can essentially be implemented in the form of computer programs or software functional modules.

[0082] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the deep learning-based fast magnetic resonance CEST quantitative method provided in the above embodiments, which includes a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the deep learning-based fast magnetic resonance CEST quantitative method as described above when executing the computer program.

[0083] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the deep learning-based fast magnetic resonance CEST quantitative method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, implements the deep learning-based fast magnetic resonance CEST quantitative method as described above.

[0084] Similarly, based on the same inventive concept, this invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the deep learning-based fast magnetic resonance CEST quantitative method as described above.

[0085] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media can also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs. Of course, with the widespread use of cloud servers, the aforementioned software programs can also be hosted on cloud platforms to provide corresponding services; therefore, computer-readable storage media are not limited to the form of local hardware.

[0086] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0087] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the device and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0088] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0089] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a magnetic resonance imaging device corresponding to the deep learning-based fast magnetic resonance CEST quantitative method provided in the above embodiments, which includes a magnetic resonance scanner and a control unit. Wherein: Magnetic resonance scanners are used to scan the object being imaged, acquiring data at all CEST saturation frequency offsets of the object. The control unit stores a computer program that, when executed, reads the magnetic resonance CEST imaging signal data acquired by the magnetic resonance scanner and generates a quantitative map of the CEST effect within the region of interest according to the deep learning-based fast magnetic resonance CEST quantification method described above.

[0090] It should be noted that the magnetic resonance imaging (MRI) device can be any MRI scanner capable of scanning data of the object being imaged. Its structure is existing technology, and mature commercial products can be used; the specific model is not limited. In addition to storing the aforementioned computer program, the control unit of the MRI device should also contain the imaging sequences and other software programs necessary for CEST imaging.

[0091] Of course, the aforementioned control unit can be an independent control unit or a control unit integrated into the magnetic resonance scanner. That is, the aforementioned deep learning-based fast magnetic resonance CEST quantitative method can be integrated into the control unit of the magnetic resonance imaging device in the form of a data processing program, so that the results can be directly output by the magnetic resonance scanner without the need for an additional control unit.

[0092] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A fast quantitative method for magnetic resonance CEST based on deep learning, characterized in that, include: S1. Acquire magnetic resonance CEST imaging signal data for the imaging object and extract the acquired z-spectrum from it. The saturation frequency shift acquired during CEST imaging needs to cover the combination of saturation frequency shifts and the target saturation frequency required for CEST effect quantification. The combination of saturation frequency shifts consists of multiple key saturation frequency shifts, which are selected and optimized in advance during the frequency selection stage. S2. For each voxel within the region of interest, the acquired z-spectrum containing only the combination of saturation frequency offsets and the B1 value are used as model inputs. A first deep neural network group, pre-trained under supervision, predicts the reference z-spectrum of the current predicted voxel to obtain a reference z-spectrum covering all saturation frequency offsets. The first deep neural network group is composed of a first encoder and a pre-trained BM equation decoder cascaded together, and both the first encoder and the BM equation decoder have fully connected network structures. The input of the first encoder is the model input of the current predicted voxel, and the output is the tissue parameters of the current predicted voxel. The input of the BM equation decoder is the B1 value of the current predicted voxel and the tissue parameters output by the first encoder, and the output is the reference z-spectrum corresponding to the current predicted voxel, which is composed of the normalized water signal intensity at all saturation frequency offsets. S3. For each voxel in the region of interest, subtract the signal value at the target saturation frequency on the acquired z-spectrum from the signal value at the target saturation frequency on the reference z-spectrum of that voxel to obtain the CEST quantitative signal of that voxel at the target saturation frequency offset. The CEST quantitative signals of all voxels in the region of interest are used to form a CEST effect quantitative map.

2. The rapid magnetic resonance CEST quantitative method based on deep learning as described in claim 1, characterized in that, In the frequency selection stage, a second deep neural network group is used to select and optimize all saturation frequency offsets to obtain the saturation frequency offset combination. The specific method is as follows: S11. Construct a second deep neural network group, which is composed of a second encoder and a pre-trained BM equation decoder cascaded together. The second encoder is obtained by cascading a batch normalization module and a third encoder, and both the third encoder and the BM equation decoder have fully connected network structures. The input of the second encoder is the acquired z-spectrum and B1 value corresponding to all saturation frequency offsets of the current predicted voxel, and the output is the tissue parameters of the current predicted voxel. The input of the BM equation decoder is the B1 value of the current predicted voxel and the tissue parameters output by the second encoder, and the output is the reference z-spectrum corresponding to the current predicted voxel, which is composed of the normalized water signal intensity at all saturation frequency offsets. S12. Fix the parameters of the BM equation decoder and perform supervised training on the second encoder in the second deep neural network group with the goal of minimizing the total loss function; the total loss function is obtained by weighting the mean square error loss term and the regularization loss term, the mean square error loss term is the mean square error between the predicted value and the true value of the reference z spectrum, and the regularization loss term is the sum of the absolute values ​​of the scaling factors corresponding to all saturation frequency offsets in the batch normalization module; S13. After training is completed, extract the scaling factor corresponding to all saturation frequency offsets in the batch normalization module, remove saturation frequency offsets with scaling factors lower than the screening threshold, and use the remaining saturation frequency offsets as key saturation frequency offsets to form the saturation frequency offset combination.

3. The deep learning-based fast magnetic resonance CEST quantitative method as described in claim 2, characterized in that, In the training sample set for supervised training of the second deep neural network group, each training sample includes the acquired z-spectrum, the B1 value corresponding to the acquired z-spectrum, and the reference z-spectrum calculated using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method. The acquired z-spectrum input to the second deep neural network group must include all the saturation frequency offsets required for calculation by the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method.

4. The deep learning-based fast magnetic resonance CEST quantitative method as described in claim 1 or 2, characterized in that, The BM equation decoder is pre-trained using the CEST dataset generated by simulation. The method for obtaining the CEST dataset generated by simulation is as follows: A two-cell Bloch-McConnell mathematical model, including a water molecule cell and a semi-solid magnetization transfer cell, was established. The tissue parameters and B1 values ​​were sampled within their respective ranges to obtain combinations of sampled values. Each combination of sampled values ​​was input into the two-cell Bloch-McConnell mathematical model. The reference z-spectrum for that combination of sampled values ​​was obtained using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR). This resulted in training samples with the sampled combinations of tissue parameters and B1 values ​​as model input and the reference z-spectrum as the ground truth label, which were then added to the CEST dataset.

5. The rapid magnetic resonance CEST quantitative method based on deep learning as described in claim 1, characterized in that, During supervised training of the first deep neural network group, the parameters of the BM equation decoder are fixed, and the mean square error between the predicted and true values ​​of the reference z-spectrum is used as the loss function. The parameters of the first encoder are optimized by minimizing the loss function.

6. The rapid magnetic resonance CEST quantitative method based on deep learning as described in claim 5, characterized in that, In the training sample set for supervised training of the first deep neural network group, each training sample includes the acquired z-spectrum, the B1 value corresponding to the acquired z-spectrum, and the reference z-spectrum calculated using the numerical fitting of extrapolated semisolid magnetization transfer reference signals (NEMR) method, and the acquired z-spectrum input to the first deep neural network group only contains the saturation frequency offset combination.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they can implement the fast magnetic resonance CEST quantitative method based on deep learning as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the deep learning-based fast magnetic resonance CEST quantitative method as described in any one of claims 1 to 6.

9. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the deep learning-based fast magnetic resonance CEST quantitative method as described in any one of claims 1 to 6 when executing the computer program.

10. A magnetic resonance imaging device, characterized in that, Includes a magnetic resonance scanner and a control unit; The magnetic resonance scanner is used to scan the imaging object and acquire magnetic resonance CEST imaging signal data of the imaging object under all CEST saturation frequency offsets. The control unit stores a computer program, which, when executed, reads the magnetic resonance CEST imaging signal data acquired by the magnetic resonance scanner and generates a quantitative map of the CEST effect within the region of interest according to the deep learning-based fast magnetic resonance CEST quantification method as described in any one of claims 1 to 6.

Images