A deep learning-based low-transient GABA editing spectrum high-precision reconstruction method and system

Abstract

The application discloses a low-transient GABA editing spectrum high-precision reconstruction method and system based on deep learning, relates to the technical field of magnetic resonance spectrum analysis, adopts a MEGA-PRESS sequence standard to collect multiple difference spectrum transients and carries out pretreatment; 2048 points in a preset range are intercepted to construct three-channel input; the three-channel input is input into an improved one-dimensional Vision Transformer network after pre-training to obtain a GABA signal curve; and the GABA quantification result is obtained by integrating the signal curve. The application constructs a training data set through a DCSS semi-real data generation method, so that the network focuses on GABA signal feature learning, the improved one-dimensional Vision Transformer network is adapted to a 1-dimensional spectrum diagram processing task, does not need to depend on special hardware or a non-standard sequence, can be compatible with an existing magnetic resonance scanning process, and only a small amount of transients are needed to realize GABA high-precision quantification.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance spectroscopy analysis technology, and more specifically to a method and system for high-precision reconstruction of low-transient GABA-edited spectra based on deep learning. Background Technology

[0002] Currently, magnetic resonance spectroscopy (MRS) has become an important non-invasive method for studying the concentration of metabolites in the brain, especially for detecting the major inhibitory neurotransmitter γ-aminobutyric acid (GABA), which has significant clinical and research value. However, due to the extremely low actual concentration of GABA in the brain (approximately 1 mM) and the severe overlap of its spectral peaks with those of other metabolites (such as creatine and inositol) in terms of chemical shift, traditional single-pulse MRS methods struggle to distinguish and quantify GABA signals. To address this issue, spectral editing techniques have emerged, among which Mescher-Garwood's J-differential editing sequence (MEGA-PRESS) has become one of the standard methods for quantifying GABA due to its ease of implementation on clinical MRI equipment and its ability to effectively remove overlapping main peak signals.

[0003] The MEGA-PRESS editing method targets the J-coupling of GABA (typically a 1.9 ppm C3 pulse) by alternately applying "edit ON / edit OFF" soft pulses, then differentially analyzing the two signals to highlight the 3.0 ppm peak of GABA in the difference spectrum and suppress overlapping signals from metabolites such as creatine. Researchers typically use basic spectrum fitting software (such as LCModel and Gannet) for post-processing and GABA concentration estimation in the difference spectrum. However, this traditional approach has several inherent limitations: First, low signal-to-noise ratio. Due to the low concentration of GABA itself, the signal-to-noise ratio of the MEGA-PRESS-edited spectrum is significantly lower than that of conventional MRS spectra, usually requiring numerous transient acquisitions (up to hundreds) to ensure spectral quality. This results in long scan times and significant susceptibility to motion artifacts. Second, co-editing contamination with macromolecular signals exists. In the MEGA-PRESS difference spectrum, some macromolecular signals, also affected by the editing pulses, remain significantly residual, affecting the baseline of the GABA signal region and impacting quantitative accuracy. Third, there are frequency / phase drift and differential artifacts. Traditional processing procedures are sensitive to frequency drift and phase error. These errors will produce obvious artifacts during the ON / OFF spectral difference process, further reducing the quantitative accuracy.

[0004] To address these issues, researchers have proposed various optimization schemes for MEGA-PRESS sequence design and preprocessing algorithms. For example, strategies such as using editing pulses (e.g., SLR pulses) with stronger MM signal suppression capabilities to reduce MM interference, automatically correcting frequency and phase drift, and incorporating real-time frequency adjustment at the sequence level can all improve spectral quality and quantitative consistency. However, these improvements primarily focus on spectral acquisition or preprocessing, and the core requirement of "reducing acquisition transients to achieve rapid and high-precision quantification" remains fundamentally unresolved.

[0005] With advancements in computational modeling capabilities, an increasing number of studies are exploring methods for reconstructing high-quality edited MRS spectra under low-transient acquisition conditions. The 2023 International Society for Biomedical Imaging (ISBI) hosted the GABA Edited-MRS Reconstruction Challenge, whose core objective was to evaluate the ability of machine learning-based techniques to perform high-fidelity reconstruction of low-transient (only 80 transients) edited MRS data. Participating methods included classic convolutional neural networks (CNNs), deep learning methods based on covariance matrices, and models that converted short-time Fourier transforms (STFTs) into spectrograms before inputting them into a Vision Transformer (Spectro-ViT). Some deep learning methods demonstrated comparable or superior performance to the traditional 320-transient approach in terms of SNR and linewidth, indicating the potential of deep learning for spectral recovery under low-transient conditions.

[0006] While these methods demonstrated good performance in the challenge, they still face the following limitations: First, there is a lack of consistent standards for spectral quality assessment. Some models optimize SNR and linewidth but do not truly improve the quantitative accuracy of GABA, indicating that existing evaluation metrics are insufficient in assessing true quantitative accuracy and signal information preservation. Second, the generalization ability from simulated data to real data is insufficient. Directly extending the models from simulated data to in vivo data under different acquisition conditions results in instability, limiting their application in real clinical / research environments. Finally, these deep learning methods do not directly address quantification errors. Although the reconstructed spectra may approximate high transient data visually or in local metrics, they do not offer a fundamental solution for improving the quantitative error of true GABA concentration.

[0007] Furthermore, at the data preprocessing level, some researchers have used CNN architectures to optimize frequency and phase correction to improve the stability of subsequent spectral quantification, but this still does not directly solve the quantization error problem under low transient conditions. The above methods typically rely on high-quality acquisition of large amounts of transient data or perform well under specific conditions, and lack a complete and robust solution for the technical goal of rapidly acquiring and accurately quantifying GABA under low transient conditions.

[0008] Therefore, how to provide a complete and robust solution for high-precision quantification of GABA from low transient MEGA-PRESS data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides a method and system for high-precision reconstruction of low transient GABA editing spectrum based on deep learning, in order to solve the problems existing in the background technology.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based method for high-precision reconstruction of low-transient GABA editing spectra includes: Multiple differential spectral transients were acquired using the MEGA-PRESS sequence, and standard data preprocessing was performed on the multiple differential spectral transients. All preprocessed difference spectrum transients are truncated to points within a preset range and used as three-channel inputs; The GABA signal curve is obtained by inputting the three-channel input into the pre-trained improved one-dimensional Vision Transformer network. Integrating the GABA signal curve yields the GABA quantization result.

[0011] Optionally, the use of MEGA-PRESS sequences for standard acquisition of multiple differential spectral transients specifically includes acquiring standard MEGA-PRESS GABA edit spectrum data from several volunteers or patients acquired by MRI equipment.

[0012] Optionally, the standard data preprocessing includes HSVD residual water peak removal, water signal eddy current correction, transient frequency domain alignment, and transient phase correction.

[0013] Optionally, the pre-training of the improved one-dimensional Vision Transformer network uses a training dataset constructed using the DCSS semi-real data generation method. The DCSS semi-real data generation method includes the steps of constructing a signal basis set based on real high-transient MEGA-PRESS data, generating semi-real inputs and labels through weighted superposition, and adding perturbations to simulate low-transient data characteristics. Optionally, the DCSS semi-real data generation method specifically includes the following steps: Acquire high transient data of standard MEGA-PRESS GABA edit spectrum from multiple devices, and then fit the pure GABA signal curve after preprocessing. The FID signals of the preprocessed high-quality edited spectra are normalized by dividing them by the maximum absolute value of the real part of the Fourier transform of their respective DIFF FID signals; similarly, the corresponding pure GABA signal curves are also normalized by the same factor; the FID signals are divided into ON FID and OFF FID. The normalized FID set and the corresponding pure GABA signal curve set constitute the base sets of the input and labels of the generated dataset, respectively; The base set of the input and labels of the dataset is repeatedly generated to form the training dataset.

[0014] Optionally, the normalized FID set and the corresponding pure GABA signal curve set respectively constitute the base set of the generated dataset input and label, specifically including: Label generation: Randomly select 1 to 4 curves from the set of pure GABA signal curves, and then randomly weight and superimpose them to obtain labels; Input generation: Extract the FIDs corresponding to the 1-4 pure GABA signal curves selected above from the FID set, perform the same weighted superposition, and obtain the semi-real high-quality edit spectrum FID; Two perturbations were added independently to the semi-real high-quality edited spectrum FID: additive white Gaussian noise of fixed intensity, zero-order phase distortion, and frequency shift were added independently to the ON FID and OFF FID to obtain two pairs of semi-real noisy ON / OFF FID. Then, ON-OFF was used to obtain two semi-real noisy DIFF FIDs. After Fourier transform, the real part was taken to obtain two semi-real noisy MEGA-PRESS DIFF spectra. The average of two semi-real noisy MEGA-PRESS DIFF spectra is used as channel 1, and the two semi-real noisy MEGA-PRESS DIFF spectra are used as channels 2 and 3, respectively, to obtain three-channel input data.

[0015] Optionally, the improved one-dimensional Vision Transformer network includes: Input layer: Receives three channels of input data; Dimension rearrangement and convolutional layer: After rearranging the dimensions of the input data to (b,e,k) format, feature extraction is performed through a 16×16 one-dimensional convolution with a stride of 16. Position encoding and Transformer encoder: After adding non-trainable position encoding, input 8 cascaded Transformer encoder modules; Deconvolution decoder: Remove the classification header and input all tokens into a decoder consisting of four one-dimensional deconvolution layers. The number of channels in the four one-dimensional deconvolution layers are 128→64→32→16, and the stride is 2. After each layer, BN processing and GELU activation are performed in sequence. Output layer: After one-dimensional adaptive average pooling, dimensional rearrangement and fully connected layer, the output is a GABA signal curve.

[0016] A high-precision reconstruction system for low-transient GABA editing spectrum based on deep learning, comprising: The data acquisition and processing module uses the MEGA-PRESS sequence to acquire multiple differential spectral transients using standard methods, and performs standard data preprocessing on the multiple differential spectral transients. The input module extracts points within a preset range from all preprocessed difference spectrum transients and uses them as three-channel inputs. The curve output module inputs the three-channel input into the pre-trained improved one-dimensional VisionTransformer network to obtain the GABA signal curve. The integration module integrates the GABA signal curve to obtain the GABA quantization result.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for high-precision reconstruction of low transient GABA editing spectrum based on deep learning, which has the following beneficial effects: 1. Existing technologies lack the ability to accurately quantify GABA from MEGA-PRESS GABA edit spectra with low transient counts; all require a large number of transients for reliable quantification. This invention, however, can quantify GABA with high precision using only a small number of transients.

[0018] 2. This invention is universal and can be used with all MRI equipment. Furthermore, the technical solution described in this invention is based on standard MEGA-PRESS edit spectrum data processing, does not rely on special hardware, additional radiofrequency pulse design, or non-standard acquisition sequences, and is directly compatible with existing clinical or research MRI scanning procedures.

[0019] 3. This invention uses a semi-real dataset generation method based on a dual-component signal set, DCSS, to enable deep learning networks to bypass the reconstruction of the remaining signal components of MEGA-PRESS and focus on feature learning of the target signal GABA. Furthermore, it improves the Vision Transformer structure to make it more suitable for 1D spectrogram processing tasks, ultimately enabling high-precision quantification of GABA from low transient data. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 A schematic diagram of the DCSS semi-realistic data generation method provided by the present invention; Figure 3 This is a schematic diagram of the improved one-dimensional Vision Transformer network structure provided by the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention discloses a high-precision reconstruction method for low-transient GABA editing spectra based on deep learning, comprising: Multiple differential spectral transients were acquired using the MEGA-PRESS sequence, and standard data preprocessing was performed on the multiple differential spectral transients. All preprocessed difference spectrum transients are truncated to 2048 points within a preset range and used as three-channel inputs; The GABA signal curve is obtained by inputting the three-channel input into the pre-trained improved one-dimensional Vision Transformer network. Integrating the GABA signal curve yields the GABA quantization result.

[0024] Specifically, based on the goal of achieving rapid acquisition by high-precision quantitative analysis of GABA from low transient MEGA-PRESS GABA edit spectra, the process of this invention is as follows: Figure 1 As shown, the main steps include: 1. Using an MRI machine with a MEGA-PRESS sequence, 10 standard difference spectrum transients are acquired, followed by standard data preprocessing (including HSVD residual water peak removal, water signal eddy current correction, and frequency-phase correction). 2. 2048 points within the 0-5ppm range of all preprocessed transients are extracted to construct a three-channel input for a deep learning network (channel 1 is the average of the ten transients, and channels 2 and 3 are the averages of the first 5 and last 5 transients, respectively). 3. The constructed three-channel difference spectrum is input into a pre-trained improved one-dimensional Vision Transformer network to obtain a GABA signal curve (range 2.79-3.445ppm). Integrating the signal curve yields a high-precision GABA quantization result.

[0025] To achieve the above objectives, this invention proposes a semi-realistic data generation method for low transient data of MEGA-PRESS GABA edited spectrum, which generates a large amount of simulated data to train the network. This method is called the DCSS (Dual-component semi-synthetic) data generation method, and the specific steps are as follows: Obtaining the input signal set and the signal set for generating the label. To construct a large amount of semi-realistic data, it is necessary to first obtain a set of real signals used to generate the data. First, acquire standard MEGA-PRESS GABA edited spectra (TR / TE=2000 / 68ms) data from several volunteers or patients using MRI equipment (e.g., 5 devices, 7 data points per device, totaling 35 data points). These data points should have a sufficiently high number of scans (e.g., NSA=320) to ensure adequate data quality.

[0026] Afterwards, data preprocessing can be performed using open-source software such as Gannet based on MATLAB, including HSVD residual water peak removal, water signal eddy current correction, transient frequency domain alignment, and transient phase correction.

[0027] After data preprocessing, quantization software such as Gannet is used to fit all the data to obtain their respective pure GABA curves (range 2.79~3.445ppm, 128 points, the number of points can be changed according to the actual application). It should be noted that the "pure GABA curve" here is a Gaussian single peak and does not contain any other signal components.

[0028] Semi-real dataset generation strategy based on two-component signal sets

[0029] This invention generates a large amount of simulated semi-realistic training by utilizing real high transient MEGA-PRESS GABA edit spectrum and pure GABA signal curves obtained by corresponding quantization fitting.

[0030] In the construction of the training set, the FID signals (divided into ONFID and OFF FID) of several preprocessed high-quality edited spectra are first normalized by dividing them by the maximum absolute value of the real part of the Fourier transform of their respective DIFF FID (ON FID – OFF FID). Similarly, the corresponding pure GABA signal curves are also normalized by the same factor.

[0031] The normalized FID set and the corresponding pure GABA signal curve set constitute the base set of the input and label of the generated dataset, respectively.

[0032] like Figure 2As shown, the steps of the DCSS data generation method to generate a semi-realistic dataset are as follows: Label generation: Randomly select 1 to 4 lines from the set of pure GABA signal curves, and then randomly weight and superimpose them (the weight sum is 1) to obtain the labels.

[0033] Generate input: Extract the FIDs corresponding to the 1-4 pure GABA signal curves selected above from the FID set, and perform the same weighted superposition (the weights are consistent with the weights when generating the labels) to obtain the semi-real high-quality edit spectrum FIDs.

[0034] Two separate perturbations are added to the semi-real high-quality edited spectrum FID: additive white Gaussian noise of a certain intensity, zero-order phase distortion, and frequency shift are added independently to the ON FID and OFF FID. This results in two pairs of semi-real noisy ON / OFF FIDs, which are then converted to ON-OFF to obtain two semi-real noisy DIFF FIDs. After Fourier transform, the real parts are taken to obtain two semi-real noisy MEGA-PRESS DIFF spectra.

[0035] Finally, the mean of these two semi-real noisy MEGA-PRESS DIFF spectra is used as channel 1, and the two semi-real noisy MEGA-PRESS DIFF spectra are used as channels 2 and 3 respectively, to obtain a three-channel input.

[0036] This invention addresses the challenge of reconstructing quantitative GABA from low transient MEGA-PRESS GABA editing spectra by proposing a semi-real dataset generation method based on a two-component signal set. This method enables deep learning models to accurately reconstruct GABA during learning, avoiding the influence of other signals such as lipids on deep learning network training and quantitative GABA. Furthermore, it can generate a large amount of semi-real data from limited real data, thereby improving the network's generalization ability and reconstruction effect.

[0037] The input channels are all in the range of 0-5 ppm, with 2048 points each (the number of points can be changed according to the actual application, as long as the FID zero-padding does not exceed three times the original length). The applied perturbation strength is adjusted according to the actual application to make the final input match the real data actually provided to the network. A large amount of semi-real data (e.g., 60,000 input-label pairs) is generated through the above process to train the network.

[0038] Meanwhile, in order to make deep learning networks more adaptable to target application scenarios, the improved VisionTransformer network of this invention can directly process the input three-channel 1D edit spectrogram. It innovatively removes the classification head originally used for classification tasks and uses all token inputs to a set of 1D deconvolution upsampling layers to decode the tokens and obtain the network output.

[0039] like Figure 3 As shown, the improved Vision Transformer network structure for spectral processing is as follows: The improved one-dimensional Vision Transformer network includes: Input layer: Receives three channels of input data; Dimension rearrangement and convolutional layer: After extracting features from the input data through a 16×16 one-dimensional convolution with a stride of 16, the feature dimensions are rearranged from (b,e,k) to (b,k,e). Position encoding and Transformer encoder: After adding non-trainable position encoding, input 8 cascaded Transformer encoder modules; Deconvolution decoder: Remove the classification header and input all tokens into a decoder consisting of four one-dimensional deconvolution layers. The number of channels in the four one-dimensional deconvolution layers are 128→64→32→16, and the stride is 2. After each layer, BN processing and GELU activation are performed in sequence. Output layer: After one-dimensional adaptive average pooling, dimensional rearrangement and fully connected layer, the output is a GABA signal curve.

[0040] A high-precision reconstruction system for low-transient GABA editing spectrum based on deep learning, comprising: The data acquisition and processing module uses the MEGA-PRESS sequence to acquire multiple differential spectral transients using standard methods, and performs standard data preprocessing on the multiple differential spectral transients. The input module extracts 2048 points from all preprocessed difference spectrum transients within a preset range and uses them as three-channel inputs. The curve output module inputs the three-channel input into the pre-trained improved one-dimensional VisionTransformer network to obtain the GABA signal curve. The integration module integrates the GABA signal curve to obtain the GABA quantization result.

[0041] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning, characterized in that, include: Multiple differential spectral transients were acquired using the MEGA-PRESS sequence, and standard data preprocessing was performed on the multiple differential spectral transients. All preprocessed difference spectrum transients are truncated to points within a preset range and used as three-channel inputs; The GABA signal curve is obtained by inputting the three-channel input into the pre-trained improved one-dimensional Vision Transformer network. Integrating the GABA signal curve yields the GABA quantization result.

2. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 1, characterized in that, The method of using MEGA-PRESS sequences to acquire multiple differential spectral transients specifically includes acquiring standard MEGA-PRESS GABA edit spectrum data from several volunteers or patients acquired by MRI equipment.

3. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 1, characterized in that, The standard data preprocessing includes HSVD residual water peak removal, water signal eddy current correction, transient frequency domain alignment, and transient phase correction.

4. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 1, characterized in that, The improved one-dimensional Vision Transformer network is pre-trained using a training dataset constructed using the DCSS semi-real data generation method. The DCSS semi-real data generation method includes the steps of constructing a signal base set based on real high transient MEGA-PRESS data, generating semi-real inputs and labels by weighted superposition, and adding perturbations to simulate the characteristics of low transient data.

5. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 4, characterized in that, The DCSS semi-realistic data generation method specifically includes the following steps: Acquire high transient data of standard MEGA-PRESS GABA edit spectrum from multiple devices, and then fit the pure GABA signal curve after preprocessing. The FID signals of the preprocessed high-quality edited spectra are normalized by dividing them by the maximum absolute value of the real part of the Fourier transform of their respective DIFF FID signals; similarly, the corresponding pure GABA signal curves are also normalized by the same factor; the FID signals are divided into ON FID and OFF FID. The normalized FID set and the corresponding pure GABA signal curve set constitute the base sets of the input and labels of the generated dataset, respectively; The base set of the input and labels of the dataset is repeatedly generated to form the training dataset.

6. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 5, characterized in that, The normalized FID set and the corresponding pure GABA signal curve set constitute the base sets of the generated dataset input and labels, respectively, specifically including: Label generation: Randomly select 1 to 4 curves from the set of pure GABA signal curves, and then randomly weight and superimpose them to obtain labels; Input generation: Extract the FIDs corresponding to the 1-4 pure GABA signal curves selected above from the FID set, perform the same weighted superposition, and obtain the semi-real high-quality edit spectrum FID; Two perturbations were added independently to the semi-real high-quality edited spectrum FID: additive white Gaussian noise of fixed intensity, zero-order phase distortion, and frequency shift were added independently to the ON FID and OFF FID to obtain two pairs of semi-real noisy ON / OFF FID. Then, ON-OFF was used to obtain two semi-real noisy DIFF FIDs. After Fourier transform, the real part was taken to obtain two semi-real noisy MEGA-PRESS DIFF spectra. The average of two semi-real noisy MEGA-PRESS DIFF spectra is used as channel 1, and the two semi-real noisy MEGA-PRESS DIFF spectra are used as channels 2 and 3, respectively, to obtain three-channel input data.

7. The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning according to claim 1, characterized in that, The improved one-dimensional Vision Transformer network includes: Input layer: Receives three channels of input data; Dimension rearrangement and convolutional layer: After rearranging the dimensions of the input data to (b,e,k) format, feature extraction is performed through a 16×16 one-dimensional convolution with a stride of 16. Position encoding and Transformer encoder: After adding non-trainable position encoding, input 8 cascaded Transformer encoder modules; Deconvolution decoder: Remove the classification header and input all tokens into a decoder consisting of four one-dimensional deconvolution layers. The number of channels in the four one-dimensional deconvolution layers are 128→64→32→16, and the stride is 2. After each layer, BN processing and GELU activation are performed in sequence. Output layer: After one-dimensional adaptive average pooling, dimensional rearrangement and fully connected layer, the output is a GABA signal curve.

8. A high-precision reconstruction system for low-transient GABA editing spectrum based on deep learning, characterized in that, The method for high-precision reconstruction of low-transient GABA editing spectrum based on deep learning, as described in any one of claims 1-7, includes: The data acquisition and processing module uses the MEGA-PRESS sequence to acquire multiple differential spectral transients using standard methods, and performs standard data preprocessing on the multiple differential spectral transients. The input module extracts points within a preset range from all preprocessed difference spectrum transients and uses them as three-channel inputs. The curve output module inputs the three-channel input into the pre-trained improved one-dimensional Vision Transformer network to obtain the GABA signal curve. The integration module integrates the GABA signal curve to obtain the GABA quantization result.

Images