A brain perfusion parameter solving method and system based on deep learning

Abstract

The application discloses a brain perfusion parameter solving method and system based on deep learning, and belongs to the field of brain perfusion imaging processing. Based on a digital phantom method, a data set with a perfusion parameter gold standard can be simulated, which is used for quantitatively solving the error of a model and optimizing network structure parameters; an unsupervised loss function is constructed, which can solve the problem of label-free training of perfusion data; and by using a deep learning method, a mapping relationship between brain perfusion data and brain perfusion parameters can be comprehensively and implicitly learned in a data-driven manner, so that the model sensitivity is reduced, and brain perfusion parameters can be more accurately and stably solved.

Description

Technical Field

[0001] This invention relates to the field of brain perfusion imaging processing, and in particular to a method and system for calculating brain perfusion parameters based on deep learning. Background Technology

[0002] In clinical brain perfusion imaging, due to image resolution and partial volume effects, it is difficult to measure the change in contrast agent concentration over time in the nearest feeding artery of the tissue of interest (accurate arterial input equation). However, existing physical models heavily rely on the accurate measurement of the arterial input equation, which makes the calculated brain perfusion parameters highly sensitive to errors introduced by the arterial input equation.

[0003] Current methods for solving perfusion parameters based on traditional physical models rely on accurate measurement of the arterial input equation. (Formula) The current physical model used to calculate perfusion parameters requires inputs of C(t) (contrast agent concentration-time curve within the tissue of interest) and AIF(t) (arterial input equation, contrast agent concentration-time curve within the arteriole closest to the tissue of interest) to output R(t) (residual function of the tissue of interest) and calculate the perfusion parameter CBF (cerebral blood flow). Due to the model's simplicity, even slight changes in the input AIF(t) will obviously alter the calculated CBF. Furthermore, in real-world scenarios, limitations imposed by image resolution, partial volume effects, and segmentation levels make it nearly impossible to input accurate AIF(t) for the entire brain tissue. Therefore, in clinical practice, contrast agent concentration-time curves from larger feeding arteries (such as the middle cerebral artery) are typically used as the AIF(t) input to the physical model formula for CBF calculation. This substitution process inevitably introduces errors (such as contrast agent delay and diffusion between the substituted artery and the actual feeding artery), resulting in poor stability and accuracy of the calculated cerebral perfusion parameters in practical applications. Furthermore, the gold standard for brain perfusion parameters is unknown: the true brain perfusion parameters of a tissue are unknown, which makes it impossible to quantitatively evaluate the accuracy of brain perfusion parameter calculation methods, and also makes it impossible to directly learn the mapping relationship from the data-label pairs of brain perfusion data and brain perfusion parameters.

[0004] Current methods often optimize the solution by modifying traditional models. For example, they might introduce regularization terms to restrict the shape of the residual function after solution, or modify the matrix filling method of deconvolution. These methods explicitly introduce prior knowledge into the physical model formulas to reduce the degrees of freedom of the parameters to be solved. The aim of these solutions is to reduce the model's sensitivity and solve for brain perfusion parameters more accurately and stably by introducing prior knowledge and complicating the model to represent the mapping relationship between brain perfusion data and brain perfusion parameters as realistically and comprehensively as possible.

[0005] Existing methods for optimizing brain perfusion parameters are based on explicit modifications to the existing physical model. These methods include adding regularization terms and modifying the matrix filling method of deconvolution. The prior knowledge introduced by these methods when explicitly modifying the model is essentially an assumption about the mapping relationship between real brain perfusion data and brain perfusion parameters. However, the true mapping relationship is difficult to represent analytically and comprehensively. Furthermore, optimization methods that add regularization terms require adjusting parameters such as penalty factors based on the data. Parameters tuned on one set of data often do not transfer well to another set, lacking generalization ability and making them difficult to apply in real-world situations.

[0006] Troubled by the above problems, the brain perfusion parameter calculation methods currently available are unstable, and it is difficult to objectively evaluate their performance and optimize them. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for solving brain perfusion parameters based on deep learning, which enriches model knowledge through a data-driven approach, reduces model sensitivity, and solves brain perfusion parameters more accurately and stably.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] A deep learning-based method for solving brain perfusion parameters includes:

[0010] Considering the various errors introduced by the arterial input equation, a digital phantom method is used to simulate and generate a simulated perfusion dataset with the gold standard of perfusion parameters;

[0011] Construct a brain perfusion parameter calculation network and an unsupervised loss function;

[0012] Based on the simulated perfusion dataset, a brain perfusion parameter calculation network is pre-trained using an unsupervised loss function to obtain the brain perfusion parameter calculation pre-trained network. Based on the gold standard of perfusion parameters in the simulated perfusion dataset, the error of the brain perfusion parameter calculation pre-trained network is quantitatively evaluated, and the network hyperparameters are adjusted in reverse to control the network performance and obtain the optimal brain perfusion parameter calculation pre-trained network.

[0013] The perfusion data in the real perfusion dataset is preprocessed; the preprocessing includes using a cluster-based method to automatically search for arterial input equations, simulating the situation in actual diagnosis and treatment where doctors use brain perfusion parameter calculation software to obtain arterial input equations, in order to simulate the interference of real-world conditions on the calculation performance;

[0014] Based on the preprocessed real perfusion dataset, the optimal brain perfusion parameter solution pretraining network is retrained using an unsupervised loss function to obtain the optimal perfusion parameter solution network.

[0015] After preprocessing the test perfusion data to be solved, input it into the optimal perfusion parameter solution network, and output the solved perfusion parameter map.

[0016] Optionally, considering the various errors introduced by the arterial input equation, the method of using a digital phantom to simulate and generate a simulated perfusion dataset with gold standard perfusion parameters specifically includes:

[0017] The arterial input equation is simulated using a gamma-variable function at a preset time resolution.

[0018] Based on the arterial input equation obtained from the simulation, combined with cerebral blood flow (CBF) and residual function R(t), the formula is used... Obtain the tissue contrast agent concentration-time curve C(t); where C a (t) represents the contrast agent concentration in the arteriole closest to the tissue of interest at time t. CBV represents cerebral blood volume, obtained by random sampling within the interval [0.5, 4.0]. MTT represents mean transit time, obtained by random sampling within the interval [3.4, 12]. The residual function R(t) uses four simulation curves, and the expressions for the four simulation curves are as follows:

[0019]

[0020]

[0021]

[0022]

[0023] Among them, R exponential (t) Simulated curve, R linear (t) Simulated curve, R boxshaped (t) The simulated curves are used to simulate tissue contrast agent concentration-time curves without considering errors introduced by the arterial input equation and different hemodynamic types; R eff (t) is used to simulate the contrast agent concentration-time curve of tissue with contrast agent diffusion between arterial input equations; β represents the coefficient, 1 / β corresponds to the effective average transit time from the measurement site of the arterial input equation to the tissue of interest, the larger the value of 1 / β, the more severe the contrast agent diffusion, β is randomly sampled in [0.5, 3], β is used to characterize different degrees of contrast agent diffusion;

[0024] The tissue contrast agent concentration-time curve C(t) is shifted relative to the simulated arterial input equation by a time length of dt to obtain the tissue contrast agent concentration-time curve C′(t) considering the contrast agent delay; where dt is obtained by random sampling within the interval [-1,3].

[0025] Based on the tissue contrast agent concentration-time curve C′(t) considering contrast agent delay, and according to the relationship between contrast agent concentration and signal intensity change S... CT (t)=S0+k×C′(t) or S MR (t)=S0*e -k*C′(t)*TE The signal intensity-time curve of brain perfusion imaging is determined; wherein, the brain perfusion imaging is CT or MRI, S CT (t) represents the CT signal intensity-time curve, S MR (t) represents the signal intensity-time curve of magnetic resonance imaging, S0 is the initial value of the perfusion sequence signal, k is a constant coefficient, and TE is the echo time;

[0026] Based on the signal intensity-time curve of brain perfusion imaging, and based on the prior knowledge of similar perfusion parameters in the neighborhood, the same perfusion parameters are taken in the same patch to obtain the same one-dimensional signal intensity-time curve.

[0027] The same one-dimensional signal intensity-time curve is embedded into a three-dimensional brain model according to a preset spatial arrangement, and Gaussian noise with different signal-to-noise ratio levels is added to obtain multiple four-dimensional perfusion simulation data, forming a simulation perfusion dataset with the gold standard of perfusion parameters.

[0028] Optionally, the brain perfusion parameter calculation network includes: a data input module, a first spatial feature extraction module, a first temporal feature extraction module, a second spatial feature extraction module, a second temporal feature extraction module, and an output module;

[0029] The data input module is used to obtain two-dimensional tissue with a spatial scale of 16×16 from four-dimensional perfusion data, and the process of contrast agent concentration of each voxel in the two-dimensional tissue changing over time. Assuming that the number of time points is 50, that is, the time scale is 50, then a 16×16×50 patch is input to the first spatial feature extraction module.

[0030] The first spatial feature extraction module is used to extract depth features on the spatial scale of the patch using convolutional layers, and outputs a 16×16×50 patch after feature mapping on the spatial scale.

[0031] The first-time feature extraction module is used to extract the depth features of the patch on the time scale after feature mapping on the spatial scale using the Encoder layer of Transformer, and outputs a 16×16×51 patch through two fully connected layers.

[0032] The second spatial feature extraction module is used to extract the depth features of a 16×16×51 patch in spatial scale using convolutional layers, and outputs a 16×16×51 patch with feature mapping in spatial scale.

[0033] The second temporal feature extraction module is used to extract the depth features of the patch on the temporal scale after feature mapping on the spatial scale using the Encoder layer of Transformer, and output the patch after processing on the temporal scale through two fully connected layers.

[0034] The output module is used to take the 16×16×1-dimensional data (the last time point on the time scale) in the output patch as the output delay time. Output the 16×16×50 dimensional data (the first 50 time points on the time scale) from the patch as the residual function for the solution. The brain perfusion parameters were calculated based on the delay time and residual function.

[0035] Optionally, the data input module specifically includes:

[0036] Signal intensity-time curves of tissue voxels of interest extracted from four-dimensional perfusion data;

[0037] Based on the signal intensity-time curve of each tissue voxel, the contrast agent concentration-time curve in the tissue of interest is calculated using the relationship between the contrast agent concentration and the signal intensity change.

[0038] Based on the signal intensity-time curve at the location of the automatically selected arterial input equation, and using the relationship between contrast agent concentration and signal intensity change, the tissue contrast agent concentration-time curve C at the location of the automatically selected arterial input equation is calculated. a (t);

[0039] Take the contrast agent concentration-time curve of a 16×16 two-dimensional tissue from the contrast agent concentration-time curve of the tissue of interest, and use it as a patch input to the first spatial feature extraction module.

[0040] Optionally, the brain perfusion parameters include cerebral blood volume (CBV), cerebral blood flow (CBF), time to peak contrast agent concentration (Tmax), and mean transit time (MTT).

[0041] The formula for calculating cerebral blood volume (CBV) is as follows:

[0042] The formula for calculating cerebral blood flow (CBF) is:

[0043] The formula for calculating the time Tmax for the contrast agent to reach its maximum peak value is:

[0044] The formula for calculating the Mean Passage Time (MTT) is as follows:

[0045] Optionally, the unsupervised loss function is:

[0046]

[0047] In the formula, C i (t) represents the tissue contrast agent concentration-time curve for the i-th tissue voxel input network. Let n be the residual function of the output corresponding to the i-th tissue voxel, and n be the number of tissue voxels. Let be the delay time for the i-th tissue voxel. This is the unsupervised loss function.

[0048] Optionally, the preprocessing includes: spatial calibration, data resampling, and measurement of the arterial input equation;

[0049] The spatial calibration is used to rigidly align the volume at each time point with the volume at the first time point using fsl, and to perform motion correction.

[0050] The data resampling is used to spatially resample the image to 1.5×1.5×4mm. 3 The temporal resolution resampling is 2 seconds;

[0051] The measured arterial input equation is used to automatically search for arterial input equations based on the criteria of manually selecting arterial input equations, using a clustering method.

[0052] A deep learning-based brain perfusion parameter calculation system includes:

[0053] The simulation module is used to consider various errors introduced by the arterial input equation and uses the digital phantom method to simulate and generate a simulated perfusion dataset with the gold standard of perfusion parameters.

[0054] The building blocks are used to construct the brain perfusion parameter calculation network and the unsupervised loss function;

[0055] The pre-training module is used to pre-train the brain perfusion parameter calculation network based on the simulated perfusion dataset and an unsupervised loss function to obtain the brain perfusion parameter calculation pre-trained network. Based on the gold standard of perfusion parameters of the simulated perfusion dataset, the module quantitatively evaluates the error of the brain perfusion parameter calculation pre-trained network, adjusts the network hyperparameters in reverse, controls the network performance, and obtains the optimal brain perfusion parameter calculation pre-trained network.

[0056] The preprocessing module is used to preprocess the perfusion data in the real perfusion dataset. The preprocessing adopts a clustering-based method to automatically search for the arterial input equation, simulating the situation in actual diagnosis and treatment where doctors use brain perfusion parameter calculation software to obtain the arterial input equation, so as to simulate the interference of the real situation on the calculation performance.

[0057] The training module is used to retrain the pre-trained network for solving the optimal brain perfusion parameters based on the pre-processed real perfusion dataset and an unsupervised loss function, so as to obtain the optimal perfusion parameter solving network.

[0058] The solution module is used to preprocess the test perfusion data to be solved, input the optimal perfusion parameter solution network, and output the solved perfusion parameter map.

[0059] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0060] This invention discloses a method and system for calculating brain perfusion parameters based on deep learning. The digital phantom method can simulate datasets with gold-standard perfusion parameters to quantify the error of the calculation model and optimize network structure parameters. An unsupervised loss function is constructed to address the problem of unlabeled training on perfusion data. Furthermore, using deep learning methods, a data-driven approach, the mapping relationship between brain perfusion data and brain perfusion parameters can be comprehensively and implicitly learned, reducing model sensitivity and enabling more accurate and stable calculation of brain perfusion parameters. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 A flowchart illustrating a deep learning-based brain perfusion parameter calculation method provided in an embodiment of the present invention;

[0063] Figure 2 A schematic diagram illustrating the principle of the brain perfusion parameter calculation method based on deep learning provided in this embodiment of the invention;

[0064] Figure 3 This is a framework diagram of the brain perfusion parameter calculation network provided in an embodiment of the present invention. Detailed Implementation

[0065] 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.

[0066] The purpose of this invention is to provide a method and system for solving brain perfusion parameters based on deep learning, which enriches model knowledge through a data-driven approach, reduces model sensitivity, and solves brain perfusion parameters more accurately and stably.

[0067] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] This invention proposes a deep learning-based method for solving brain perfusion parameters. By employing a data-driven approach that covers as many error-introducing mechanisms as possible, it implicitly and comprehensively supplements the mapping relationship between brain perfusion data and brain perfusion parameters, reducing the model's sensitivity to errors introduced by the arterial input equation. A digital volumetric simulation dataset with the gold standard for perfusion parameters is used to quantify the model error and optimize the model structure, resulting in a brain perfusion parameter solution model with superior accuracy and stability.

[0069] like Figure 1 and Figure 2 As shown in the figure, an embodiment of the present invention provides a brain perfusion parameter calculation method based on deep learning, which includes the following steps:

[0070] Step S1: Considering the various errors introduced by the arterial input equation, a digital phantom method is used to simulate and generate a simulated perfusion dataset with the gold standard of perfusion parameters.

[0071] First, a simulation dataset is generated based on the digital phantom method. This dataset covers as many errors as possible that might be introduced into the arterial input equation under real-world conditions. Specifically, given perfusion parameters that conform to physiological conditions (such as cerebral blood volume (CBV), cerebral blood flow (CBF), and mean transit time (MTT), a simulation dataset with the gold standard for perfusion parameters is generated by combining it with a physical model. Furthermore, errors that might occur under real-world conditions are introduced into the simulation as much as possible, providing a generalized dataset for deep learning and model performance evaluation. Specifically, the simulation dataset includes the following process:

[0072] S1.1) The arterial input equation (AIF, C) is simulated using a gamma-variable function at the required time resolution. a (t)), as in formula (4.1).

[0073]

[0074] Where t0 is the arrival time of the contrast agent, and parameters C0, a, and b are all AIF curve morphology parameters. In this invention, t0 = 12s, C0 = 1.0, a = 3.0, and b = 1.5s are used.

[0075] S1.2) Based on the physical model, i.e. formula (4.2), input CBF, AIF and residual function R(t) to obtain the tissue contrast agent concentration curve C(t).

[0076]

[0077] in CBV was obtained by random sampling within the interval [0.5, 4.0], and MTT was obtained by random sampling within the interval [3.4, 12]. Different perfusion parameters can simulate tissues with different perfusion levels.

[0078] Residual function R(t): R(t) depends on the local vascular structure and is an unknown physiological curve. This invention uses four curves to simulate R(t), as shown in formulas (4.3)-(4.6). Among them, (4.3)-(4.5) are input into formula (4.2) to simulate C(t) of different hemodynamic types without considering the errors that may be introduced by the arterial input equation. Formula (4.6) is input into formula (4.2) to simulate C(t) with contrast agent diffusion between arterial input equations, let For the measured AIF (such as the contrast agent concentration-time curve in the middle artery typically input into the model), h * Let (t) be the blood vessel distribution function (describing the probability distribution at each transmission time), then with h * (t) Convolution can represent the diffusion of the contrast agent, used to represent the true AIF (C) of the tissue of interest. a (t)). As shown in Equation (4.7), C(t) with contrast agent diffusion can also be expressed as the measured C(t). With effective residual function The convolution, where R (eff) Simulated by formula (4.6), where 1 / β corresponds to the effective "mean transit time" from the AIF measurement site to the tissue of interest; a larger value indicates more severe diffusion. In this invention, β is randomly sampled within [0.5, 3] to simulate different degrees of contrast agent diffusion.

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] Then, the C(t) curve is shifted relative to AIF by a time length of dt, which is obtained by random sampling within the interval [-1,3] to simulate different degrees of contrast agent delay.

[0085] S1.3) Based on the relationship between contrast agent concentration and signal intensity changes in CT (Computed Tomography) or MRI (Magnetic Resonance Imaging), such as Equation (4.8) or Equation (4.9), simulate the signal intensity-time curve of CT or MRI.

[0086] S CT (t)=S0+k×C(t) (4.8)

[0087] S MR (t)=S0*e -k*C(t)*TE (4.9)

[0088] Where S0 is the initial value of the perfusion sequence signal, k is a constant coefficient, and TE is the echo time.

[0089] S1.4) Based on the prior knowledge of similar perfusion parameters in the neighborhood, the same perfusion parameters are taken in the same patch (CT: 32×32, MRI: 16×16) to obtain the same one-dimensional signal intensity-time curve. These curves are then embedded into the three-dimensional brain model according to a certain spatial arrangement. Gaussian noise with different signal-to-noise ratios is added to obtain four-dimensional perfusion simulation data.

[0090] Based on the above process, the simulation dataset of this invention covers the main sources of error in current methods for reducing the performance of perfusion parameter calculation, such as contrast agent delay, diffusion and spatiotemporal scale noise, providing a generalized dataset for reducing model sensitivity through data-driven approaches.

[0091] Step S2: Construct the brain perfusion parameter calculation network and the unsupervised loss function.

[0092] This invention, taking into account the superior performance of Transformer in time series analysis and Convolutional Neural Networks (CNNs) in spatial feature extraction, proposes the following... Figure 3The brain perfusion parameter calculation network framework shown includes an input block, a spatial features extraction block, a temporal features extraction block, and an output block.

[0093] Data Input Module: Extracts the signal intensity-time curve (S(t)) of each tissue voxel from the four-dimensional perfusion images. Based on the relationship between contrast agent concentration and signal intensity change, such as formula (4.8) or formula (4.9), calculates the tissue contrast agent concentration-time curve (C(t)) to obtain the four-dimensional C(t). Input the signal intensity-time curve (S) of the automatically selected AIF location. AIF Similarly, based on the relationship between contrast agent concentration and signal intensity change, AIF(C) is calculated. a (t)). A three-dimensional C(t) with a spatial scale of 16×16 (two spatial dimensions plus one time dimension) is taken as the patch input into the network.

[0094] Spatial feature extraction module: It uses convolutional layers to extract deep features at the spatial scale of the input patch, without processing them at the temporal scale, and outputs a patch with the same feature mapping dimension.

[0095] Temporal feature extraction module: The Encoder layer of Transformer is used to extract deep features on the time scale, and the output is mapped to the patch after processing on the time scale through two fully connected layers.

[0096] Output module: The last dimension of the time scale in the output 16×16×51 patch, i.e., the 16×16×1 dimension data, is used as the output delay time. The remaining 16×16×50 dimensional data is used as the residual function for the solution. and Jointly participate in backpropagation. The brain perfusion parameters CBV, CBF, and Tmax are input into formulas (4.10)-(4.12) to calculate the brain perfusion parameters, and then the brain perfusion parameter MTT is calculated according to formula (4.13) to obtain the brain perfusion parameters calculated by this method. Parameter maps represent parameter mappings.

[0097]

[0098]

[0099]

[0100]

[0101] Unsupervised loss function: Since the true perfusion level of the tissue is unknown, supervised learning can only be applied to labeled simulation data, which limits the potential of the algorithm. Therefore, this invention combines a physical model to construct an unsupervised loss function as shown in formula (4.14).

[0102]

[0103] Where C i (t) represents C(t) of the input network for the i-th voxel. a (t) represents the AIF used in the input model to solve for the perfusion parameters. Let C be the residual function of the output corresponding to the i-th voxel. Based on the physical model, C... a (t) and delay Duration Convolution can estimate Through calculation estimation With input C t The mean square error between (t) can measure the computational performance of the model.

[0104] This invention avoids the problem of not having a gold standard for training with brain perfusion parameters by constructing an unsupervised loss function in conjunction with a physical model.

[0105] Step S3: Based on the simulated perfusion dataset, a brain perfusion parameter calculation network is pre-trained using an unsupervised loss function to obtain the brain perfusion parameter calculation pre-trained network. Based on the gold standard of perfusion parameters in the simulated perfusion dataset, the error of the brain perfusion parameter calculation pre-trained network is quantitatively evaluated, and the network hyperparameters are adjusted in reverse to control the network performance and obtain the optimal brain perfusion parameter calculation pre-trained network.

[0106] Input the simulated perfusion dataset into, for example Figure 3 In the brain perfusion parameter calculation network, the output solution residual function and delay time The network converges by iterating backward using the unsupervised loss function shown in Equation (4.14).

[0107] This invention pre-trains a network based on simulation data. The simulation test set is input into the network to obtain the calculated perfusion parameters. The correlation and variance of the relative error between these parameters and the gold standard of the test set are calculated. The higher the correlation and the smaller the variance, the more accurate and stable the network solution is. Based on this standard, the network parameters (number of convolutional layers, number of Transformer Encoder layers, number of neurons in fully connected layers, number of cascaded spatial and temporal feature extraction blocks (the number of feature extraction blocks can be increased or decreased according to the performance of the pre-trained network) are adjusted in reverse to control the performance of the pre-trained network and obtain the optimal pre-trained network.

[0108] Step S4: Preprocess the perfusion data in the real perfusion dataset; the preprocessing includes using a clustering-based method to automatically search for arterial input equations, simulating the situation where doctors use brain perfusion parameter calculation software to obtain arterial input equations during actual diagnosis and treatment, in order to simulate the interference of real-world conditions on the calculation performance.

[0109] Before inputting the data into the network, preprocessing of the perfusion data is required. Preprocessing includes three parts: spatial calibration, data resampling, and AIF measurement. For spatial calibration: Due to the long acquisition time of perfusion data, the patient may move. This invention uses freeform flow (FSL) to rigidly align the volume at each time point with the volume at the first time point for motion correction. For data resampling: The image is spatially resampled to 1.5 × 1.5 × 4 mm. 3 The temporal resolution resampling is 2 seconds. These two preprocessing steps reduce unnecessary changes that the network needs to handle, speed up analysis, and improve the accuracy and repeatability of network parameter calculation. For measuring AIF: This invention uses a clustering method to automatically search for AIF based on the criteria for manually selecting AIF (earlier contrast agent arrival time, steeper peak of the contrast agent concentration-time curve, earlier peak arrival, narrower and higher peak, larger area under the curve, etc.), and inputs it into the network for calculating perfusion parameters. This invention uses a cluster-based automatic search method for AIF measurement instead of manual selection by professional doctors to simulate the situation of acute stroke (an important application scenario for brain perfusion parameter calculation) where "time is brain." Doctors usually do not have time to carefully calibrate AIF and often use the cluster-based automatic search method in existing brain perfusion parameter calculation software to measure AIF, which inevitably introduces errors into AIF and poses challenges to the calculation method. This invention reduces the sensitivity of the calculation method to AIF errors, reduces the impact of inaccurate AIF measurement in real-world situations, and contributes to the diagnosis and treatment of acute stroke.

[0110] Step S5: Based on the preprocessed real perfusion dataset, the optimal brain perfusion parameter calculation pre-training network is retrained using an unsupervised loss function to obtain the optimal perfusion parameter calculation network.

[0111] Since the spatiotemporal noise distribution simulated by simulation data has limitations, this invention captures the real noise distribution from real data to supplement the mapping relationship between real perfusion data and perfusion parameters.

[0112] The preprocessed data is input into a network pre-trained with simulation data, and the network is trained unsupervised until convergence to obtain the infusion parameter solution network.

[0113] Step S6: After preprocessing the test perfusion data to be solved, input it into the optimal perfusion parameter solution network and output the solved perfusion parameter map.

[0114] The signal intensity-time curve S(t) for each voxel of interest can be obtained from the perfusion data to be solved. When brain perfusion imaging is a CT image, the signal intensity-time curve S(t) is S CT (t). When brain perfusion imaging is performed using magnetic resonance imaging, the signal intensity-time curve S(t) is S MR (t).

[0115] This invention uses deep learning methods and a data-driven approach to comprehensively and implicitly learn the mapping relationship between brain perfusion data and brain perfusion parameters, reducing model sensitivity. Even when there are difficult-to-quantify errors in the measured arterial input equation, it can still solve brain perfusion parameters more accurately and stably.

[0116] This invention uses a digital phantom method to simulate a dataset with the gold standard of brain perfusion parameters and proposes an objective quantitative method to calculate model errors.

[0117] This invention proposes a reliable and effective method for training brain perfusion parameter calculation networks by constructing an unsupervised loss function based on a physical model when the gold standard for perfusion parameters is unknown.

[0118] The advantages of this invention are as follows:

[0119] 1. This invention uses deep learning to learn the mapping relationship between perfusion data and perfusion parameters from diverse simulation data and real data, which can achieve a more accurate and stable solution than the solution method based on physical models.

[0120] 2. This method, based on digital phantoms, can simulate datasets with perfusion parameters that meet the gold standard, quantifying the errors in the computational model and optimizing network structure parameters. It also supplements the residual function used in previous simulations by introducing a blood vessel distribution function to simulate contrast agent diffusion, a significant source of error affecting the performance of the computational method. This allows for the generation of more realistic datasets that better test the model's robustness against errors.

[0121] 3. This method is based on the infusion solution physical model and constructs an unsupervised loss function, which can solve the problem of unlabeled training on infused data.

[0122] 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 systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0123] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for calculating brain perfusion parameters based on deep learning, characterized in that, include: Considering the various errors introduced by the arterial input equation, a digital phantom method is used to simulate and generate a simulated perfusion dataset with the gold standard of perfusion parameters; Construct a brain perfusion parameter calculation network and an unsupervised loss function; Based on the simulated perfusion dataset, a brain perfusion parameter calculation network is pre-trained using an unsupervised loss function to obtain the brain perfusion parameter calculation pre-trained network. Based on the gold standard of perfusion parameters in the simulated perfusion dataset, the error of the brain perfusion parameter calculation pre-trained network is quantitatively evaluated, and the network hyperparameters are adjusted in reverse to control the network performance and obtain the optimal brain perfusion parameter calculation pre-trained network. The perfusion data in the real perfusion dataset is preprocessed; the preprocessing includes using a cluster-based method to automatically search for arterial input equations, simulating the situation in actual diagnosis and treatment where doctors use brain perfusion parameter calculation software to obtain arterial input equations, in order to simulate the interference of real-world conditions on the calculation performance; Based on the preprocessed real perfusion dataset, the optimal brain perfusion parameter solution pretraining network is retrained using an unsupervised loss function to obtain the optimal perfusion parameter solution network. After preprocessing the test perfusion data to be solved, input it into the optimal perfusion parameter solution network and output the solved perfusion parameter map; The brain perfusion parameter calculation network includes: a data input module, a first spatial feature extraction module, a first temporal feature extraction module, a second spatial feature extraction module, a second temporal feature extraction module, and an output module; The data input module is used to obtain data from four-dimensional perfusion data. The process of contrast agent concentration changing over time in a two-dimensional tissue at a spatial scale, and for each voxel within that tissue, assuming 50 time points (i.e., 50 time scales), then the input... The patch is applied to the first spatial feature extraction module; The first spatial feature extraction module is used to extract depth features at the patch spatial scale using convolutional layers, and outputs... The patch after feature mapping at the spatial scale; The first-time feature extraction module is used to extract features from the Transformer's Encoder layer. The deep features of the patch, mapped at the spatial scale, are then mapped and output through two fully connected layers. The patch; The second spatial feature extraction module is used to extract features using convolutional layers. The patch's depth features at the spatial scale, output The patch after feature mapping at the spatial scale; The second temporal feature extraction module is used to extract 16 features from the Transformer's Encoder layer. The deep features of the patch after feature mapping on the spatial scale are mapped to the output patch after processing on the temporal scale through two fully connected layers. The output module is used to output the last time point on the time scale of the output patch. Delay time for dimensional data as output Output the remaining parts of the patch. Dimensional data as the residual function of the solution The brain perfusion parameters were calculated based on the delay time and residual function.

2. The brain perfusion parameter calculation method based on deep learning according to claim 1, characterized in that, Considering the various errors introduced by the arterial input equation, a digital phantom method is used to simulate and generate a simulated perfusion dataset with gold standard perfusion parameters, specifically including: The arterial input equation is simulated using a gamma-variable function at a preset time resolution. Based on the arterial input equation obtained from the simulation, combined with cerebral blood flow... CBF and residual function Obtain tissue contrast agent concentration-time curves ;in, for t Contrast agent concentration in the arteriole closest to the tissue of interest at any given time. , CBV For cerebral blood volume, CBV From interval Obtained by internal random sampling. MTT The average transit time, MTT From interval Obtained by internal random sampling, residual function Four simulation curves are used, and the expressions for the four simulation curves are as follows: in, , Simulated curves The simulation curves are used to simulate the concentration-time curves of tissue contrast agents without considering the errors introduced by the arterial input equation and the different hemodynamic types. Tissue contrast agent concentration-time curves used to simulate contrast agent diffusion between arterial input equations; This corresponds to the effective average transit time from the measurement site of the arterial input equation to the tissue of interest. A higher value indicates more severe contrast agent diffusion. exist Internal random sampling, Used to characterize different degrees of contrast agent diffusion; Tissue contrast agent concentration-time curve By shifting the simulated arterial input equation by the time dt, a tissue contrast agent concentration-time curve considering contrast agent delay is obtained. ; where dt is the interval Obtained by internal random sampling; Based on tissue contrast agent concentration-time curves that take contrast agent delay into account Based on the relationship between contrast agent concentration and signal intensity change or The signal intensity-time curve of brain perfusion imaging is determined; wherein, the brain perfusion imaging is CT or MRI. The signal intensity-time curve of CT scan. This is the signal intensity-time curve of magnetic resonance imaging. The initial value of the perfusion sequence signal, The constant coefficient, Echo time; Based on the signal intensity-time curve of brain perfusion imaging, and based on the prior knowledge of similar perfusion parameters in the neighborhood, the same perfusion parameters are taken in the same patch to obtain the same one-dimensional signal intensity-time curve. The same one-dimensional signal intensity-time curve is embedded into a three-dimensional brain model according to a preset spatial arrangement, and Gaussian noise with different signal-to-noise ratio levels is added to obtain multiple four-dimensional perfusion simulation data, forming a simulation perfusion dataset with the gold standard of perfusion parameters.

3. The brain perfusion parameter calculation method based on deep learning according to claim 1, characterized in that, The data input module specifically includes: Signal intensity-time curves of tissue voxels of interest extracted from four-dimensional perfusion data; Based on the signal intensity-time curve of each tissue voxel, the contrast agent concentration-time curve in the tissue of interest is calculated using the relationship between contrast agent concentration and signal intensity change. Based on the signal intensity-time curve at the location of the automatically selected arterial input equation, the tissue contrast agent concentration-time curve at that location is calculated using the relationship between contrast agent concentration and signal intensity change. ; Take from the contrast agent concentration-time curve within the tissue of interest The contrast agent concentration-time curve of the two-dimensional tissue is used as a patch input to the first spatial feature extraction module.

4. The brain perfusion parameter calculation method based on deep learning according to claim 1, characterized in that, The cerebral perfusion parameters include cerebral blood volume. CBV Cerebral blood flow CBF Time to peak contrast agent and average transit time MTT ; The cerebral blood volume CBV The solution formula is: ; The cerebral blood flow CBF The solution formula is: ; The time it takes for the contrast agent to reach its maximum peak value The solution formula is: ; The average passage time MTT The solution formula is: .

5. The brain perfusion parameter calculation method based on deep learning according to claim 3, characterized in that, The unsupervised loss function is: In the formula, For the first Tissue contrast agent concentration-time curves for input networks of individual tissue voxels. For the first The residual function corresponding to each tissue voxel. The number of tissue voxels, For the first The delay time of an individual tissue voxel This is the unsupervised loss function.

6. The brain perfusion parameter calculation method based on deep learning according to claim 1, characterized in that, The preprocessing includes: spatial calibration, data resampling, and measurement of arterial input equations; The spatial calibration is used to rigidly align the volume at each time point with the volume at the first time point using fsl, and to perform motion correction. The data resampling is used to spatially resample the image to... The temporal resolution resampling is 2 seconds; The measured arterial input equation is used to automatically search for arterial input equations based on the criteria of manually selecting arterial input equations, using a clustering method.

7. A brain perfusion parameter calculation system based on deep learning, characterized in that, include: The simulation module is used to consider various errors introduced by the arterial input equation and uses the digital phantom method to simulate and generate a simulated perfusion dataset with the gold standard of perfusion parameters. The building blocks are used to construct the brain perfusion parameter calculation network and the unsupervised loss function; The pre-training module is used to pre-train the brain perfusion parameter calculation network based on the simulated perfusion dataset and an unsupervised loss function to obtain the brain perfusion parameter calculation pre-trained network. Based on the gold standard of perfusion parameters of the simulated perfusion dataset, the module quantitatively evaluates the error of the brain perfusion parameter calculation pre-trained network, adjusts the network hyperparameters in reverse, controls the network performance, and obtains the optimal brain perfusion parameter calculation pre-trained network. The preprocessing module is used to preprocess the perfusion data in the real perfusion dataset. The preprocessing adopts a clustering-based method to automatically search for the arterial input equation, simulating the situation in actual diagnosis and treatment where doctors use brain perfusion parameter calculation software to obtain the arterial input equation, so as to simulate the interference of the real situation on the calculation performance. The training module is used to retrain the pre-trained network for solving the optimal brain perfusion parameters based on the pre-processed real perfusion dataset and an unsupervised loss function, so as to obtain the optimal perfusion parameter solving network. The solution module is used to preprocess the test perfusion data to be solved, input the optimal perfusion parameter solution network, and output the solved perfusion parameter map. The brain perfusion parameter calculation network includes: a data input module, a first spatial feature extraction module, a first temporal feature extraction module, a second spatial feature extraction module, a second temporal feature extraction module, and an output module; The data input module is used to obtain data from four-dimensional perfusion data. The process of contrast agent concentration changing over time in a two-dimensional tissue at a spatial scale, and for each voxel within that tissue, assuming 50 time points (i.e., 50 time scales), then the input... The patch is applied to the first spatial feature extraction module; The first spatial feature extraction module is used to extract depth features at the patch spatial scale using convolutional layers, and outputs... The patch after feature mapping at the spatial scale; The first-time feature extraction module is used to extract features from the Transformer's Encoder layer. The deep features of the patch, mapped at the spatial scale, are then mapped and output through two fully connected layers. The patch; The second spatial feature extraction module is used to extract features using convolutional layers. The patch's depth features at the spatial scale, output The patch after feature mapping at the spatial scale; The second temporal feature extraction module is used to extract features from the Transformer's Encoder layer. The deep features of the patch after feature mapping on the spatial scale are mapped to the output patch after processing on the temporal scale through two fully connected layers. The output module is used to output the last time point on the time scale of the output patch. Delay time for dimensional data as output Output the remaining parts of the patch. Dimensional data as the residual function of the solution The brain perfusion parameters were calculated based on the delay time and residual function.

Images