CN113838105B - Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning - Google Patents
Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning Download PDFInfo
- Publication number
- CN113838105B CN113838105B CN202111107046.XA CN202111107046A CN113838105B CN 113838105 B CN113838105 B CN 113838105B CN 202111107046 A CN202111107046 A CN 202111107046A CN 113838105 B CN113838105 B CN 113838105B
- Authority
- CN
- China
- Prior art keywords
- model
- microcirculation
- diffusion
- value
- values
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000004089 microcirculation Effects 0.000 title claims abstract description 86
- 238000000034 method Methods 0.000 title claims abstract description 84
- 238000009792 diffusion process Methods 0.000 title claims abstract description 77
- 238000013135 deep learning Methods 0.000 title claims abstract description 56
- 239000006185 dispersion Substances 0.000 claims abstract description 53
- 238000012549 training Methods 0.000 claims abstract description 38
- 230000033001 locomotion Effects 0.000 claims abstract description 12
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 claims abstract description 11
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 25
- 238000010606 normalization Methods 0.000 claims description 22
- 230000006870 function Effects 0.000 claims description 19
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 239000008280 blood Substances 0.000 claims description 11
- 210000004369 blood Anatomy 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 11
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 230000007246 mechanism Effects 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 10
- 238000010276 construction Methods 0.000 claims description 8
- 125000004122 cyclic group Chemical group 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000006798 recombination Effects 0.000 claims description 7
- 238000005215 recombination Methods 0.000 claims description 7
- 238000003860 storage Methods 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 7
- 238000012935 Averaging Methods 0.000 claims description 6
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 230000004962 physiological condition Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 10
- 210000002826 placenta Anatomy 0.000 description 9
- 238000012360 testing method Methods 0.000 description 9
- 239000013598 vector Substances 0.000 description 6
- 238000003384 imaging method Methods 0.000 description 5
- 230000007423 decrease Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 238000013398 bayesian method Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000002598 diffusion tensor imaging Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002595 magnetic resonance imaging Methods 0.000 description 2
- 230000008447 perception Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 206010028980 Neoplasm Diseases 0.000 description 1
- 210000001015 abdomen Anatomy 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000006073 displacement reaction Methods 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000001605 fetal effect Effects 0.000 description 1
- 210000003754 fetus Anatomy 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000003475 lamination Methods 0.000 description 1
- 230000007786 learning performance Effects 0.000 description 1
- 230000008774 maternal effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007427 paired t-test Methods 0.000 description 1
- 231100000915 pathological change Toxicity 0.000 description 1
- 230000036285 pathological change Effects 0.000 description 1
- 230000010412 perfusion Effects 0.000 description 1
- 230000035935 pregnancy Effects 0.000 description 1
- 230000000241 respiratory effect Effects 0.000 description 1
- 230000029058 respiratory gaseous exchange Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004904 shortening Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012418 validation experiment Methods 0.000 description 1
- 230000036266 weeks of gestation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Multimedia (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning. Firstly, collecting multi-b value data corresponding to a diffusion microcirculation model, and carrying out iterative registration on diffusion weighting data of each b value to eliminate motion artifacts; then adopting a Bayesian estimation method to fit voxel data in the region of interest, and taking the obtained model parameters as gold standards of training data; and then designing an encoder based on a transducer and an SCDNN decoder based on a dispersion microcirculation model, combining the encoder and the SCDNN decoder to obtain a deep learning network structure based on model driving, and finally training the network by using gold standard data to obtain a model which can be used for estimating dispersion microcirculation model parameters. The method can obtain image information with similar quality under the condition of less acquisition time, has higher accuracy and precision, has better estimation effect than other dispersion microcirculation model estimation methods, and has better model interpretation.
Description
Technical Field
The application relates to the field of magnetic resonance imaging optimization and data model fitting, in particular to parameter estimation method optimization and device of a dispersion microcirculation model and a storage medium.
Background
Diffusion magnetic resonance imaging (dMRI) is an important medical imaging tool for noninvasively detecting tissue microstructure based on limited diffusion of water molecules in biological tissue. The Apparent Diffusion Coefficient (ADC) is commonly calculated using an index, which is sensitive to pathological changes such as stroke, tumor, etc., but not to microstructure characteristics. Advanced biophysical models have now been developed to characterize specific microstructure properties such as Diffusion Tensor Imaging (DTI), diffusion Kurtosis Imaging (DKI), intra-voxel incoherent motion model (IVIM), and other atrioventricular models.
Accurate estimation of model parameters of microstructure characteristics is of great importance for diagnosis. However, most dwri models consist of multiple mathematically complex and highly nonlinear parts. The models are fitted by using traditional optimization techniques such as least square method, and estimation errors are easy to generate. Furthermore, from a data acquisition perspective, advanced dwri models require acquisition of multiple b values and diffusion directions in q-space, which is time consuming and susceptible to motion artifacts.
The deep learning technology opens up a new approach for dMRI model fitting. The q-space learning method is a multi-layer perceptron (MLP) based method that uses a subset of q-space data to estimate DTI parameters. However, it is independent of the biophysical model and therefore difficult to interpret. The introduction of domain knowledge as a priori information in deep neural networks is considered an effective way to improve network performance and interpretability. Furthermore, the MLP framework has limited ability to capture rich information in images. Convolutional networks are designed for image feature extraction, but their application is limited due to the fixed perception fields. And adding a self-attention mechanism adapting to the dynamic perception field in the q-space deep learning method, so as to improve learning performance and form a transducer structure.
Disclosure of Invention
In order to overcome the defects of long time, large influence of motion and poor interpretation of a deep learning method of the prior acquisition technology, the invention provides a dispersion microcirculation model driving parameter estimation method based on deep learning.
In order to achieve the above purpose, the invention is realized by adopting the following technical scheme:
in a first aspect, the present invention provides a method for estimating driving parameters of a diffusion microcirculation model based on deep learning, which is used for estimating model parameters of the diffusion microcirculation model, and includes the following steps:
s1, performing cyclic registration among multiple b values on a magnetic resonance diffusion weighted image acquired under the multiple b values, and dividing an interested region on the registered image;
s2, taking the signal values of voxels in the image interested region under the multiple b values obtained in the S1 as fitting data, and performing parameter fitting on the dispersion microcirculation model by adopting a Bayes estimation method to obtain model parameters of the dispersion microcirculation model which are regarded as gold standards;
s3, taking part of b values as b value combinations on the basis of the magnetic resonance diffusion weighted image acquired in the S1, taking diffusion magnetic resonance signals corresponding to the b values in the b value combinations as input, taking model parameters obtained by fitting in the corresponding voxels in the S2 as truth labels, and constructing training data of default b values; training a deep learning network driven by a diffusion microcirculation model by using training data of a default b value, and obtaining a parameter estimation model after training;
The deep learning network consists of an encoder part and a Sparse Coding Deep Neural Network (SCDNN) decoder based on a diffusion microcirculation model; in the encoder part, inputting image blocks taking each voxel to be estimated as a center under different b values in training data, firstly encoding the image blocks, then sequentially passing through a plurality of transform encoders, and taking the output of the last transform encoder as the input of a decoder; in the decoder, the signals passing through the sparse coding depth neural network iteration unit are overlapped with the output signals which are not iterated, and the model parameter estimation values corresponding to the voxels to be estimated are output after repeated iteration;
s4, acquiring a magnetic resonance diffusion weighted image of the object to be estimated, acquired under each b value in the b value combination, inputting the magnetic resonance diffusion weighted image into a parameter estimation model after registration and region-of-interest division, and estimating to obtain model parameters of a diffusion microcirculation model.
Based on this approach, the steps may further provide the following preferred implementations. It should be noted that the technical features of the preferred embodiments can be combined with each other without any conflict. These preferred embodiments can be realized by other means capable of achieving the same technical effects, and are not limited.
Preferably, the implementation method of S1 is as follows:
for each tested individual, respectively obtaining a plurality of magnetic resonance Diffusion Weighted Images (DWI) corresponding to b values, and correcting motion artifacts among the b-value images through cyclic registration; during cyclic registration, firstly, averaging the images of all b values to obtain an average template, then comparing and registering all the images of the b values with the average template through rigid transformation and affine transformation, then averaging the images obtained after registration again to obtain a new average template, and then comparing and registering again after obtaining the new average template, and repeating the steps for 6-10 times to obtain a registration result of diffusion weighted images under each b value; and finally dividing the region of interest on the basis of the registration result.
Preferably, in the step S2, the dispersion microcirculation model is as follows:
wherein: s is S b For the signal value at the value b, S 0 For signal values without diffusion weighting, f is the volume fraction of the microcirculation, D is the diffusion coefficient of the tissue water molecules, D * Is pseudo-dispersion coefficient, f, D of water molecules in microcirculation blood * Are all model parameters to be fitted.
Preferably, in the step S3, each sample in the training data of the default b value contains a sequence data combination of a voxel x and a model parameter truth value label corresponding to the voxel x, the sequence data combination of the voxel x includes sequence data corresponding to different b values in the b value combination, and the sequence data corresponding to each b value is represented by a normalized diffusion weighting signal S of each voxel in a graphic block centered on the voxel x in the magnetic resonance diffusion weighted image corresponding to the b value b /S 0 Sequentially forming; the number of b values in the b value combination is preferably not less than 3, and more preferably 3 to 5.
Preferably, in the step S3, the encoder section is formed by connecting a full-connection layer and a plurality of transducers, the sequence data in the input samples are encoded by the full-connection layer and then sequentially pass through the plurality of transducers, the output of the former transducer is used as the input of the next transducer, and the output of the last transducer is used as the output of the encoder section for inputting to the decoder; in each transducer encoder, the original input of the transducer encoder is input into a multi-head attention mechanism layer after being normalized by a layer, the output of the multi-head attention mechanism layer and the original input which is not normalized by the layer are overlapped through residual connection to obtain a new input, and then the new input sequentially passes through the layer normalization layer and the two full connection layers, and the obtained output is added with the new input to obtain the final output of the transducer encoder.
Preferably, the construction method of the SCDNN decoder based on the dispersion microcirculation model comprises the following steps:
The SCDNN decoder is a decoder driven by a dispersion microcirculation model, and the constructed decoder is obtained by expanding the dispersion microcirculation model by using an iterative algorithm. Wherein the dispersion microcirculation model is as follows:
wherein f is the perfusion fraction, D is the water molecule diffusion coefficient, D * Is the pseudo dispersion coefficient of water molecules.
The original model is subjected to linearization treatment by a method for carrying out dictionary on signals obtained by the dispersion microcirculation model. The following is shown:
where y is the observed signal and,is composed of the dispersion coefficient D of discretized tissue water molecules and the pseudo-dispersion coefficient D of water molecules in microcirculation blood * The dictionary vector is formed, and x is a coefficient vector of a dictionary formed by discretized f. By constructing the linearized model, an objective function as follows is constructed to solve:
where β is its regularization coefficient. The solution process of the objective function is expanded by using a hard threshold iterative algorithm, and the expansion process is as follows:
x n+1 ＝H M (Wy+Sx n )
wherein the method comprises the steps ofH M Is a nonlinear operator. When the absolute value of the input value is smaller than the set threshold value, the absolute value is 0, and when the absolute value is larger than or equal to the threshold value, the absolute value is kept unchanged. According to the above process, the convolution layer and the batch can be combined The normalization layer, threshold layer, and convolution layer constructions form a decoder.
Therefore, the specific model structure of the decoder in S3 is as follows: the decoder receives the output of the encoder part as input, firstly, the input sequentially passes through a convolution layer, a batch normalization layer, a threshold layer and the convolution layer to obtain an output signal, then the output signal is input into a continuous iteration unit, and the iteration unit is designed according to a dictionary construction method, so that the construction of a dictionary is realized through an iteration process; the input signal in each iteration unit sequentially passes through a threshold layer and a convolution layer threshold layer to form a nonlinear operator, the convolution layer combines the data in the channel, and then the signal output by the convolution layer is overlapped with the original input signal of the iteration unit through residual connection to be used as the output signal after passing through the iteration unit and used as the input signal of the next iteration unit; the nth iteration unit realizes the general expression of the hard threshold iteration algorithm in the dispersion microcirculation model through the processing of signals: x is x n+1 ＝H M (Wy+Sx n ) Wherein the weights areIs composed of the dispersion coefficient D of discretized tissue water molecules and the pseudo-dispersion coefficient D of water molecules in microcirculation blood * Dictionary vectors composed,/- >I is an identity matrix, H M Is a nonlinear operator, y is an observed signal, x n And x n+1 Representing the input signal and the output signal of the nth iteration unit respectively; repeating the iteration to obtain a final output signal after passing through all iteration units, inputting the final output signal into a normalization layer to obtain a volume fraction f, and limiting the signal to 0-1 after normalization to meet the actual physiological condition; dividing the output of the normalization layer into two equal-length parts, respectively normalizing again, inputting two different convolution layers for signal combination, outputting the result of recombination by the dispersion coefficient D of the discretized tissue water molecules by one convolution layer, and rolling the other convolution layerThe lamination output is formed by discretizing the pseudo-diffusion coefficient D of water molecules in the microcirculation blood * Results of the recombination.
Preferably, in the step S3, the deep learning network performs network parameter optimization by minimizing a total loss function for each training batch, where the total loss function is a sum of loss function values of all input samples in each training batch, and an absolute error between an output model parameter estimation result and a true value label of the loss function of each input sample is as follows:
Wherein [ f, D * ]Is the gold standard value of three parameters in the model parameters,and estimating three parameters in the model parameters output by the deep learning network.
Preferably, in the step S4, the magnetic resonance diffusion weighted image of the object to be estimated is registered and the region of interest is divided according to the same method as that of the samples in the training data.
In a second aspect, the invention provides a parameter estimation device of a diffusion microcirculation model driven deep learning network, which comprises a memory and a processor;
the memory is used for storing a computer program;
the processor is configured to implement the deep learning based dispersion microcirculation model driving parameter estimation method according to any one of the first aspects when executing the computer program.
In a third aspect, the present invention provides a computer readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the method for estimating a diffusion microcirculation model driving parameter based on deep learning according to any one of the first aspects is implemented.
Compared with the prior art, the invention has the following beneficial effects:
the invention provides a diffusion microcirculation model driving parameter estimation method based on deep learning, which can achieve the parameter fitting effect of multiple b values only by a small number of b value data and has the effect of shortening the acquisition time in actual operation. And because voxel incoherent motion imaging is a complex and highly nonlinear model mathematically, estimation errors are easy to generate when the model is fitted by using traditional optimization techniques such as a least square method and the like, and erroneous judgment is caused. Meanwhile, as the traditional optimization algorithm is an optimization method based on iteration, a long time is needed for iteration. The traditional q-space learning method is a black box model and has no interpretation. According to the invention, in combination with a traditional diffusion magnetic resonance model fitting method and a deep learning method, an iterative solution process of a diffusion microcirculation model is converted into a deep learning network. The method can effectively reduce the time required for solving the characteristic parameters while collecting the b-value data. Compared with a deep learning method of q space, the method not only has great improvement on fitting accuracy, but also has model interpretation.
Drawings
FIG. 1 is a schematic diagram of a model-based driving deep learning network.
Fig. 2 shows a comparison of the results of the present invention with other networks, where (a) shows a comparison of the results of the encoder of the present invention with the results of the convolutional encoder, and (b) shows a comparison of the results of the decoder of the present invention with the results of the q-space decoder.
FIG. 3 shows the relative error of dispersion microcirculation model parameter estimation at signal-to-noise ratio of 10-70.
FIG. 4 is a multi-center comparison of the deep learning network of the present invention.
Detailed Description
The following method according to the present invention will demonstrate specific technical effects thereof with reference to examples so that those skilled in the art can better understand the spirit of the present invention.
The invention combines a physical mechanism based on a dMRI microcirculation model (namely an IVIM model) with a deep learning method to obtain a better effect.
In a preferred embodiment of the present invention, taking placenta as an example, there is provided a diffusion microcirculation model driving parameter estimation method based on deep learning, the method is used for estimating model parameters of a diffusion microcirculation model, and specifically includes the following steps:
step one: multiple b-value magnetic resonance Diffusion Weighted Images (DWIs) of placenta are acquired over a single-shot diffusion weighted EPI sequence on a human magnetic resonance system. The number of b values required to be covered in the initially acquired multi-b value DWI can be adjusted according to the actual situation. In this example, using human placenta data as an example, the acquisition of diffusion-weighted EPI sequences by single excitation on a 1.5T GE magnetic resonance system comprises b values of 0, 10, 20, 50, 80, 100, 150, 200, 300, and 500s/mm 2 DWI of 10 b-values, the subjects were scanned for 29 normal pregnant women with gestational weeks 13-37.
Because the b value that imaging needs to be acquired is more, the time for image acquisition is longer, and intrauterine imaging is easy to be subjected to respiratory motion of the abdomen of the pregnant woman and irregular motion images of the fetus. In order to ensure the consistency of the image space of all different b values, so that the calculation of each voxel parameter is more accurate, registration is firstly required to be carried out on the DWI acquired by each tested individual under the multiple b values, and the displacement among the diffusion weighted images of each b value caused by the parent respiration and fetal movement is removed, so that the tissue structures in the images of different b values are consistent in space, and the influence of motion artifact is eliminated. In the invention, the motion artifact correction is realized by adopting cyclic registration, and the specific method comprises the following steps: the method comprises the steps of firstly, averaging DWI of each b value to obtain an average template, then comparing and registering the DWI of all b values with the average template through rigid transformation and affine transformation, then averaging the DWI of all b values obtained after registration again to obtain a new average template, and comparing and registering again after obtaining the new average template, and repeating the steps for 6-10 times to obtain a registration result of diffusion weighted images under each b value. Since the DWI of the human placenta also contains other tissue images besides the placenta, after the registration of the images is completed, the region of interest, i.e., the placenta region in the images, is finally divided on the basis of the registration result.
Of course, in other embodiments, the specific region of interest needs to be selected according to the actual study needs, which is not limited.
Step two: since the DWI of the human placenta needs to be labeled with a true value in advance when it is used for training in the deep learning network, the current accepted accurate calculation method is required to be used as the label value. In this embodiment, the signal value of the voxel in the region of interest of the image under the multiple b values obtained in S1 is used as fitting data, and a bayesian estimation method is adopted to perform parameter fitting on the dispersion microcirculation model, so as to obtain model parameters of the dispersion microcirculation model regarded as a gold standard.
Wherein, the dispersion microcirculation model is as follows:
wherein: s is S b For the signal value at the value b, S 0 For signal values without diffusion weighting, f is the volume fraction of the microcirculation, D is the diffusion coefficient of the tissue water molecules, D * Is pseudo-dispersion coefficient, f, D of water molecules in microcirculation blood * All are three model parameters to be fitted.
When fitting this step, the fitting data is the signal values of voxels in the region of interest of DWI at 10 b values, and since fitting is performed on a single voxel, the fitting independent variable and dependent variable are also data in the corresponding single voxel, respectively.
After obtaining the truth value label corresponding to each voxel in the interested area of the DWI, the deep learning method can be adopted subsequently, and the characteristic parameters are obtained through training of the fitted data, and the specific practice is described below.
Step three: based on the magnetic resonance diffusion weighted image acquired in the step S1, taking part of b values as b value combinations, combining the image interested area corresponding to each b value in the b value combinations with the model parameters obtained by fitting in the step S2 as truth labels, and constructing training data of default b values; training a deep learning network driven by a diffusion microcirculation model by using training data of a default b value, and obtaining a parameter estimation model after training.
In the present invention, although all b values are needed when the foregoing bayesian estimation method fits the model parameter gold standard, too many b values are acquired, which results in too long acquisition time, so we want b values needed by the deep learning network to be smaller than b values needed by the bayesian estimation method, that is, less than 10 b values in the present embodiment. Therefore, in this embodiment, only m b values are taken out from 10 b values and recorded as b value combinations, and the deep learning network can realize dispersion microcirculation model parameter estimation of each voxel in the region of interest only by DWI data corresponding to each b value in the b value combinations. However, the number m of b values in the b value combination cannot be too small, otherwise, the estimation accuracy will be poor, and m is preferably 3-5 in consideration of accuracy and acquisition time.
Thus, each sample in the training data of the default b value is constructed for a voxel in the region of interest, so the deep learning network is input and output voxel by voxel to achieve training. For any one of the templates corresponding to the voxels x, the template comprises a sequence data combination of the voxels x and a model parameter truth value label corresponding to the voxels x, the sequence data combination of the voxels x comprises sequence data corresponding to different b values in the b value combination, and the sequence data corresponding to each b value is formed by a normalized diffusion weighting signal S of each voxel in a graph block centering on the voxel x in the magnetic resonance diffusion weighting image corresponding to the b value b /S 0 Sequentially. In the actual processing process, the DWI corresponding to each b value can be scanned in a sliding way by using an n x n extraction frame, the central voxel of each extraction frame is the voxel x corresponding to the sample, and the normalized dispersion weighting signals S of n x n voxels b /S 0 And sequentially combining the two-dimensional sequence data with the length of n. The one-dimensional sequence data of the same voxel x obtained in DWI with different b values constitutes the sequence data combination in the sample, and the voxels obtained in the second step are combined And (3) the model parameter truth value label corresponding to the x can form a complete sample with the label. In this embodiment, n is preferably 3.
The model-driven deep learning network designed by the invention consists of an encoder part and a Sparse Coding Deep Neural Network (SCDNN) decoder based on a diffusion microcirculation model. The encoder part is composed of a layer normalization layer, a multi-head attention mechanism layer, a full connection layer, a residual connection layer and the like, and a plurality of transform encoders are contained in the encoder part. In the encoder part, an image block taking each voxel to be estimated as a center under different b values in training data is input, the image block is encoded, then sequentially passes through a plurality of transducer encoders, and the output of the last transducer encoder is used as the input of a decoder. The decoder is a model-driven decoder, and the dispersion microcirculation model in the second step is unfolded through a multi-layer neural network by using an iterative algorithm to obtain the constructed decoder. In the decoder, the signals passing through the sparse coding depth neural network iteration unit are overlapped with the output signals which are not iterated, and the model parameter estimated values corresponding to the voxels to be estimated are output after repeated iteration.
The decoder is driven based on a dispersion microcirculation model, and the original model is subjected to linearization treatment by a method for carrying out dictionary treatment on signals obtained by the dispersion microcirculation model, wherein the linearization treatment is as follows:
where y is the observed signal and,is formed by discretizing D and D * The dictionary vector is formed, and x is a coefficient vector of a dictionary formed by discretized f. By constructing the linearized model, an objective function as follows is constructed to solve:
where β is its regularization coefficient. The solution process of the objective function is expanded by using a hard threshold iterative algorithm, and the expansion process is as follows:
x n+1 ＝H M (Wy+Sx n )
wherein the method comprises the steps ofH M Is a nonlinear operator. When the absolute value of the input value is smaller than the set threshold value, the absolute value is 0, and when the absolute value is larger than or equal to the threshold value, the absolute value is kept unchanged. According to the above procedure, a decoder can be formed by combining the convolutional layer, the batch normalization layer, the threshold layer, and the convolutional layer.
The specific structure of the encoder portion and decoder in the deep learning network is described in detail below with reference to FIG. 1
As shown in fig. 1 c, the encoder section is formed by connecting a full connection layer for encoding and a plurality of transducers (the number of transducers is represented by L), and the decoder section is connected to the encoder section.
In the encoder section, the sequence data combination in the input samples is encoded by the full-concatenated layer, and then sequentially passes through the plurality of transducers, the encoding result of the full-concatenated layer is input to the first transducer encoder, and the other transducers except the first transducer encoder use the output of the previous transducer encoder as the input of the next transducer encoder, and the output of the last transducer encoder as the output of the encoder section for input to the decoder. As shown in fig. 1 (a), the structure of a single transducer encoder is shown, each transducer encoder has the same structure and comprises a layer normalization layer, a multi-head attention mechanism layer, a residual connection layer and a full connection layer, the original input of the transducer encoder is input into the multi-head attention mechanism layer after being normalized by the layer, and the output of the multi-head attention mechanism layer and the original input which is not normalized by the layer are overlapped through the residual connection to obtain a new input; the new input is then passed through the layer normalization layer and the two fully connected layers in sequence, and the resulting output is added to the new input to yield the final output of the transducer encoder.
And as shown in fig. 1 (b), a structure of a decoder is shown, the decoder receiving an output of the encoder section as an input. In the decoder, the output of the encoder is firstly taken as the input, the output signal is obtained after the encoder part sequentially passes through the convolution layer, the batch normalization layer, the threshold layer and the convolution layer, and then the output signal is input into a continuous iteration unit which is designed according to the dictionary construction method, so that the dictionary construction is realized through the iteration process. The input of the first iteration unit is the output signal of the convolution layer, and the input of the subsequent iteration unit is the output of the previous iteration unit. Each iteration unit is composed of a threshold layer, a convolution layer and residual connection, after an input signal in each iteration unit sequentially passes through the threshold layer and the convolution layer, the threshold layer is a nonlinear operator, the convolution layer combines data in a channel, and then a signal output by the convolution layer is overlapped with an original input signal of the iteration unit through the residual connection to be used as an output signal after passing through the iteration unit and used as an input signal of a next iteration unit. The above layer structure of each iteration unit is used for realizing the general term expression of the hard threshold iteration algorithm in the dispersion microcirculation model through processing signals, and the general term expression realized by any nth iteration unit is as follows: x is x n+1 ＝H M (Wy+Sx n ) Wherein the weights areIs a dictionary vector formed by discretized diffusion coefficient D of tissue water molecules and pseudo diffusion coefficient D of water molecules in microcirculation blood, +.>I is an identity matrix, H M Is a nonlinear operator, y is an observed signal, x n And x n+1 Input and output signals respectively representing an nth iteration unitNumber (x). And repeating the iteration to obtain a final output signal after passing through all iteration units. Based on the final output signal, three different parameters of the model parameters can be calculated, but the method of calculating the three different parameters is slightly different: for the volume fraction f, the final output signal is input into a normalization layer to obtain the volume fraction f, and the signal can be limited to 0-1 after normalization to meet the actual physiological condition; however, for the other two parameters, the output from the normalization layer is divided into two equal parts according to the length, and then normalized again and input into two different convolution layers, wherein the two convolution layers output the diffusion coefficient D of water molecules of tissue and the pseudo-diffusion coefficient D of water molecules in the microcirculation blood * The result of discretization signal recombination, that is to say, two equal-length parts enter two network branches respectively, one part of output signal line passes through a normalization layer and then passes through a convolution layer in the first network branch to obtain the result of discretization tissue water molecule diffusion coefficient D recombination, and the other part of output signal line passes through a normalization layer and then passes through the convolution layer in the second network branch to obtain the discretization microcirculation blood water molecule pseudo diffusion coefficient D * Results of the recombination.
It should be noted that, the threshold layer in the present invention plays a role similar to an activation function, and when the input value of the threshold layer is greater than a preset threshold, the threshold layer is set to 0 and output, and when the input value of the threshold layer is greater than or equal to the preset threshold, the input value is kept unchanged and output.
The deep learning network can perform batch training by utilizing the labeled training data, and can be used as a parameter estimation model after the network precision meets the requirement. During training, each training batch (batch) performs network parameter optimization by minimizing a total loss function, wherein the total loss function is the sum of loss function values of all input samples in each training batch, and the absolute error of the output model parameter estimation result and the truth value label of the loss function of each input sample is as follows:
wherein [ f, D * ]Is the gold standard value of three parameters in the model parameters,three parameter estimation values of model parameters output by the deep learning network, [ f, D * ]And->And calculating absolute errors in a one-to-one correspondence mode, and adding to obtain the loss function value of the sample.
Step four: in practical application, the DWI acquired by the object to be estimated under each b value in the b value combination can be acquired according to the same method as the first step, and after registration and region division of interest which are the same as those of the samples in the training data, the DWI is input into a parameter estimation model, and model parameters of a dispersion microcirculation model are estimated and obtained. In the parameter estimation model, model parameter estimation is performed on a pixel-by-pixel basis, and each pixel forms an image block of n×n with itself as a center point, thereby forming an input of a network.
Therefore, the invention can obtain three characteristic parameters of f, D and D of each voxel in the placenta image by using the parameter estimation model only by adopting 3-5 DWI with b values, which can greatly reduce the sampling times required for obtaining the multi-b-value DWI of the pregnant woman, shorten the time required for acquisition and not influence the reconstruction effect of the characteristic parameters.
The technical effects of the method based on steps one to four of the method described above are shown in combination with examples so that those skilled in the art can better understand the spirit of the present invention.
Examples
The diffusion microcirculation model driving parameter estimation method based on deep learning is tested in the tested data of 29 pregnancy periods in 13-37 weeks, and the basic flow of the specific steps is as described above. Magnetic resonance scanning was performed on a universal electrical sign HDXT 1.5T magnetic resonance scanner using diffusion weightingPlanar echo imaging sequence acquires placenta images from maternal sagittal: echo Time (TE)/repetition Time (TR) =76/3000 ms, field of view (FOV) =320×320mm, in-plane resolution of 1.25×1.25mm, layer thickness 4mm, total 15 layers, total 10 b-value data acquired at 0, 10, 20, 50, 80, 100, 150, 200, 300 and 500s/mm respectively 2 . The normal pregnant woman samples of 29 pregnant women in 13-37 weeks of gestation are divided into a training set, a verification set and a test set according to the ratio of 6:2:2. Wherein the training set has 15 samples (324016 voxels), the verification set has 4 samples (11628 voxels), and the test set has 5 samples (131867 voxels).
Meanwhile, in order to contrast and show the technical effects of the invention, the implementation compares the results obtained by the neural network in the test set with the gold standard obtained by fitting and the results obtained by other methods, and simultaneously performs anti-noise capability test and multi-center generalization capability test under different noise conditions. The experimental results are shown in fig. 2 to 4:
FIG. 2 (a) shows a q-Space learning decoder (q-Space decoder) and a sparse coded depth neural network (SCDNN decoder) according to the present invention, with the input of 5 b-values (20, 50, 150, 300, 500s/mm 2 ) The comparison result at that time. The q-space learning decoder is designed according to the method of Golkov, and consists of three fully connected layers and a nonlinear active ReLU. The result in fig. 2 (a) shows that the model parameters estimated from the SCDNN decoder have a higher correlation with the true values than the q-space learning decoder. The SCDNN decoder has lower estimation errors than the q-space learning decoder, especially for Δf and Δd plots.
FIG. 2 (b) shows a diffusion microcirculation model driven parameter estimation method encoder (transducer encoder) and a convolutional encoder (Conv 2D encoder) based on deep learning, with the values at 5 b inputs (20, 50, 150, 300, 500s/mm 2 ) The comparison result at that time. The convolutional encoder consists of a convolutional layer, a batch normalization layer and a modified linear unit activation function. The correlation plot and Δf, Δd, and Δd plots in fig. 2 (b) indicate that the transducer encoder provides a more accurate dispersion microcirculation model parameter estimation.
Fig. 3 is an experiment of the result of estimating the parameters of the influence of the signal-to-noise ratio of the image. In images with signal to noise ratio levels from 10 to 70 we evaluate the relative error (percentage of error to true value) at different signal to noise ratio levels. Fig. 3 (a) shows that the relative error of f gradually decreases with increasing signal-to-noise ratio and stabilizes at signal-to-noise ratios greater than 40. In contrast, the estimation of D is relatively insensitive to signal-to-noise ratio (fig. 3 (b)). The relative error of D is not greatly changed along with the signal to noise ratio, and is stable when the signal to noise ratio is more than 40, and the result shows that the dispersion microcirculation model driving parameter estimation method based on deep learning has stronger robustness to noise, and the optimal estimation precision can be ensured when the signal to noise ratio is more than 40.
Fig. 4 is a multi-center validation experiment. Previous training, validation and testing was performed using data acquired on a 1.5T universal electrical sign HDXT scanner, where we tested network performance using data acquired on a 3.0T universal electrical 750W of another hospital using the same acquisition protocol. The new test data included 2 normal subjects (37194 voxels). The results show that the estimation error of the multi-center data is slightly increased, but still sufficient for parameter estimation.
Meanwhile, the invention also compares error performances of the dispersion microcirculation model driving parameter estimation method based on deep learning under the conditions of different numbers of b values and different combinations of b values, as shown in table 1. We selected a subset of 5 b values of 20, 50, 150, 300, 500s/mm respectively 2 . Here we tested the other four combinations listed in table 1, with the combination with the best performance highlighted in bold for each model parameter. In general, 20, 50, 150, 300, 500s/mm 2 The b value combination of (a) realizes the optimal balance of the parameter estimation precision of all three dispersion microcirculation models. While we tested the effect of different numbers of b values. Table 1 shows that the mean square error decreases significantly as the b value increases from 3 to 5, but increases from 5 to 7, the scan time is 1.4 times that of 5 b values, but the increase in error decrease is limited.
Table 1: the dispersion microcirculation model parameters estimated using DWI data of different b-value combinations are compared to the Mean Square Error (MSE) of gold standard.
In addition, the invention also makes comparison with the traditional algorithm and other deep learning algorithms. Table 2 compares four different algorithms, including two traditional optimization methods: least Squares (LSQ) and Bayesian methods (Bayesian), and two learning-based methods: q-space deep learning and SCDNN (without transducer). Compared with the least square method and the Bayesian method, the learning-based method has higher estimation accuracy. Of the learning-based methods, the model-driven method has better performance than q-space learning without prior information, and the deep learning-based diffusion microcirculation model-driven parameter estimation Method (MESSC) has the best performance.
Table 2: comparing the five methods using reduced b-values (20, 50, 150, 300, 500s/mm 2 ) Mean Square Error (MSE) of the parameters of the dispersive microcirculation model was estimated, p < 0.05, p < 0.01, p < 0.001 for each algorithm paired with METSC.
Meanwhile, the present invention also compares the influence of different structures (q-Space decoder/SCDNN decoder, transducer encoder/convolutional encoder, block image (original image is divided into several blocks according to overlapping) input/non-block image (original image is not operated on), position coding (1-D position coding information obtained by learning)/non-position coding (no position information)) on the microcirculation parameter estimation on all test data, as shown in table 3. This will indicate that the existing structure has the best performance.
Table 3: on five test data (about 131867 voxels), the Mean Square Error (MSE) of the METSC was encoded using different encoders, decoders, input forms and positions. Paired t-test was performed on four sets of comparisons, p < 0.05, p < 0.01, p < 0.001, respectively.
In addition, in other embodiments, a diffusion microcirculation model driving parameter estimation device based on deep learning can also be provided, which comprises a memory and a processor;
The memory is used for storing a computer program;
the processor is used for realizing the dispersion microcirculation model driving parameter estimation method based on the deep learning when executing the computer program.
In addition, in other embodiments, a computer readable storage medium may be provided, where a computer program is stored on the storage medium, and when the computer program is executed by a processor, the foregoing method for estimating driving parameters of a diffusion microcirculation model based on deep learning is implemented.
It should be noted that the Memory may include random access Memory (Random Access Memory, RAM) or may include Non-Volatile Memory (NVM), such as at least one magnetic disk Memory. The processor may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; but also digital signal processors (Digital Signal Processing, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate arrays (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Of course, the apparatus should also have necessary components to implement the program operation, such as a power supply, a communication bus, and the like.
It should be noted that the above-mentioned embodiment is only a preferred embodiment of the present invention, but is not limited thereto. Various changes and modifications may be made by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present invention. Therefore, all the technical schemes obtained by adopting the equivalent substitution or equivalent transformation are within the protection scope of the invention.
Claims (8)
1. The dispersion microcirculation model driving parameter estimation method based on deep learning is used for estimating model parameters of a dispersion microcirculation model and is characterized by comprising the following steps of:
s1, performing cyclic registration among multiple b values on a magnetic resonance diffusion weighted image acquired under the multiple b values, and dividing an interested region on the registered image;
s2, taking the signal values of voxels in the image interested region under the multiple b values obtained in the S1 as fitting data, and performing parameter fitting on the dispersion microcirculation model by adopting a Bayes estimation method to obtain model parameters of the dispersion microcirculation model which are regarded as gold standards;
s3, based on the magnetic resonance diffusion weighted image acquired in the S1, taking part of b values as b value combinations, combining the image interested area corresponding to each b value in the b value combinations with model parameters obtained by fitting in the S2 as truth labels, and constructing training data of default b values; training a deep learning network driven by a diffusion microcirculation model by using training data of a default b value, and obtaining a parameter estimation model after training;
The deep learning network consists of an encoder part and a Sparse Coding Deep Neural Network (SCDNN) decoder based on a diffusion microcirculation model; in the encoder part, inputting an image block taking each voxel to be estimated as a center under different b values in training data, firstly encoding the image block, then sequentially passing through a plurality of transform encoders, and taking the output of the last transform encoder as the input of a decoder; in the decoder, the signals passing through the sparse coding depth neural network iteration unit are overlapped with the output signals which are not iterated, and the model parameter estimation values corresponding to the voxels to be estimated are output after repeated iteration;
s4, acquiring a magnetic resonance diffusion weighted image of the object to be estimated, acquired under each b value in the b value combination, inputting the image into a parameter estimation model after registration and region division of interest, and estimating to obtain model parameters of a diffusion microcirculation model;
in the step S3, the encoder part is formed by connecting a full connection layer and a plurality of transducers, the sequence data combination in the input samples is encoded by the full connection layer and then sequentially passes through the plurality of transducers, the output of the former transducer is used as the input of the next transducer, and the output of the last transducer is used as the output of the encoder part for inputting to a decoder; in each transducer encoder, the original input of the transducer encoder is firstly normalized by a layer and then is input into a multi-head attention mechanism layer, then the output of the multi-head attention mechanism layer and the original input which is not normalized by the layer are overlapped through residual connection to obtain a new input, and then the new input sequentially passes through the layer normalization layer and the two full connection layers, and then the obtained output is added with the new input to obtain the final output of the transducer encoder;
In the step S3, the decoder receives the output of the encoder part as input, firstly, the input sequentially passes through a convolution layer, a batch normalization layer, a threshold layer and the convolution layer to obtain an output signal, then the output signal is input into a continuous iteration unit, and the iteration unit is designed according to a dictionary construction method, so that the construction of a dictionary is realized through an iteration process; the input signal in each iteration unit sequentially passes through a threshold layer and a convolution layer, the threshold layer is a nonlinear operator, the convolution layer combines the data in the channel, and then the signal output by the convolution layer is overlapped with the original input signal of the iteration unit through residual connection to be used as the output signal after passing through the iteration unit and used as the input signal of the next iteration unit; the nth iteration unit realizes the general expression of the hard threshold iteration algorithm in the dispersion microcirculation model through the processing of signals: x is x n+1 ＝H M (Wy+Sx n ) Wherein the weights areIs composed of the dispersion coefficient D of discretized tissue water molecules and the pseudo-dispersion coefficient D of water molecules in microcirculation blood * Dictionary formed by the methodVector (S)>I is an identity matrix, H M Is a nonlinear operator, y is an observed signal, x n And x n+1 Representing the input signal and the output signal of the nth iteration unit respectively; repeating the iteration to obtain a final output signal after passing through all iteration units, inputting the final output signal into a normalization layer to obtain a volume fraction f, and limiting the signal to 0-1 after normalization to meet the actual physiological condition; dividing the output of the normalization layer into two equal-length parts, respectively normalizing again, inputting two different convolution layers for signal combination, wherein one convolution layer outputs the result of recombination by the dispersion coefficient D of discretized tissue water molecules, and the other convolution layer outputs the pseudo-dispersion coefficient D of discretized microcirculation blood water molecules * Results of the recombination.
2. The diffusion microcirculation model driving parameter estimation method based on deep learning according to claim 1, wherein the implementation method of S1 is as follows:
for each tested individual, respectively obtaining a plurality of magnetic resonance Diffusion Weighted Images (DWI) corresponding to b values, and correcting motion artifacts among the b-value images through cyclic registration; during cyclic registration, firstly, averaging the images of all b values to obtain an average template, then comparing and registering all the images of the b values with the average template through rigid transformation and affine transformation, then averaging the images obtained after registration again to obtain a new average template, and then comparing and registering again after obtaining the new average template, and repeating the steps for 6-10 times to obtain a registration result of diffusion weighted images under each b value; and finally dividing the region of interest on the basis of the registration result.
3. The method for estimating driving parameters of a diffusion microcirculation model based on deep learning according to claim 1, wherein in S2, the diffusion microcirculation model is as follows:
wherein: s is S b For the signal value at the value b, S 0 For signal values without diffusion weighting, f is the volume fraction of the microcirculation, D is the diffusion coefficient of the tissue water molecules, D * Is pseudo-dispersion coefficient, f, D of water molecules in microcirculation blood * Are all model parameters to be fitted.
4. The deep learning based diffusion microcirculation model driving parameter estimation method according to claim 1, wherein in the S3, each sample in the training data of default b values contains a sequence data combination of voxel x and a model parameter truth value label corresponding to the voxel x, the sequence data combination of voxel x includes sequence data corresponding to different b values in the b value combination, and each sequence data corresponding to b value is represented by a normalized diffusion weighting signal S of each voxel in a graphic block centered on the voxel x in the magnetic resonance diffusion weighted image corresponding to b value b /S 0 Sequentially forming; the number of b values in the b value combination is preferably not less than 3.
5. The deep learning based dispersion microcirculation model driving parameter estimation method according to claim 1, wherein in the step S3, the deep learning network performs network parameter optimization by minimizing a total loss function for each training batch, the total loss function is a sum of loss function values of all input samples in each training batch, and an absolute error of an output model parameter estimation result and a true value label of the loss function of each input sample is as follows:
Wherein [ f, D * ]Is the gold standard value of three parameters in the model parameters,and estimating three parameters in the model parameters output by the deep learning network.
6. The method for estimating parameters according to claim 1, wherein in S4, the registration and the region of interest division are performed according to the same method as the sample in the training data on the magnetic resonance diffusion weighted image of the object to be estimated.
7. The dispersion microcirculation model driving parameter estimation device based on deep learning is characterized by comprising a memory and a processor;
the memory is used for storing a computer program;
the processor is configured to implement the diffusion microcirculation model driving parameter estimation method based on deep learning according to any one of claims 1 to 6 when executing the computer program.
8. A computer-readable storage medium, wherein a computer program is stored on the storage medium, which when executed by a processor, implements the deep learning-based dispersion microcirculation model driving parameter estimation method according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111107046.XA CN113838105B (en) | 2021-09-22 | 2021-09-22 | Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111107046.XA CN113838105B (en) | 2021-09-22 | 2021-09-22 | Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113838105A CN113838105A (en) | 2021-12-24 |
| CN113838105B true CN113838105B (en) | 2024-02-13 |
Family
ID=78960378
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111107046.XA Active CN113838105B (en) | 2021-09-22 | 2021-09-22 | Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113838105B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117173173A (en) * | 2023-11-01 | 2023-12-05 | 华中科技大学同济医学院附属同济医院 | Non-essential disease classification screening method for diffusion weighted imaging |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111445553A (en) * | 2020-03-31 | 2020-07-24 | 浙江大学 | Depth learning-based intra-voxel incoherent motion imaging acceleration method and device |
| CN112184845A (en) * | 2020-09-02 | 2021-01-05 | 东软医疗系统股份有限公司 | Generation method and device of diffusion weighted image reconstruction model |
| CN113421247A (en) * | 2021-06-30 | 2021-09-21 | 首都医科大学附属北京天坛医院 | Method and device for determining brain tissue microstructure estimation parameters and electronic equipment |
-
2021
- 2021-09-22 CN CN202111107046.XA patent/CN113838105B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111445553A (en) * | 2020-03-31 | 2020-07-24 | 浙江大学 | Depth learning-based intra-voxel incoherent motion imaging acceleration method and device |
| CN112184845A (en) * | 2020-09-02 | 2021-01-05 | 东软医疗系统股份有限公司 | Generation method and device of diffusion weighted image reconstruction model |
| CN113421247A (en) * | 2021-06-30 | 2021-09-21 | 首都医科大学附属北京天坛医院 | Method and device for determining brain tissue microstructure estimation parameters and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113838105A (en) | 2021-12-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | Deep learning for image enhancement and correction in magnetic resonance imaging—state-of-the-art and challenges | |
| CN110378980B (en) | Multichannel magnetic resonance image reconstruction method based on deep learning | |
| Xiang et al. | Deep-learning-based multi-modal fusion for fast MR reconstruction | |
| US10088544B2 (en) | Tractography framework with magnetic resonance imaging for brain connectivity analysis | |
| CN109523584B (en) | Image processing method and device, multi-modality imaging system, storage medium and equipment | |
| CN113496495B (en) | Medical image segmentation model building method capable of realizing missing input and segmentation method | |
| CN110619635B (en) | Hepatocellular carcinoma magnetic resonance image segmentation system and method based on deep learning | |
| CN111784788A (en) | PET (positron emission tomography) rapid imaging method and system based on deep learning | |
| US11835612B2 (en) | System and method for motion correction of magnetic resonance image | |
| Wang et al. | High fidelity direct-contrast synthesis from magnetic resonance fingerprinting in diagnostic imaging | |
| CN114299185A (en) | Magnetic resonance image generation method, magnetic resonance image generation device, computer equipment and storage medium | |
| CN113838105B (en) | Diffusion microcirculation model driving parameter estimation method, device and medium based on deep learning | |
| US20150016701A1 (en) | Pulse sequence-based intensity normalization and contrast synthesis for magnetic resonance imaging | |
| Tamilselvan et al. | Brain tumor detection from clinical CT and MRI images using WT-FCM algorithm | |
| Giraldo et al. | Super-Resolution Reconstruction of Multi-Slice T2-W FLAIR MRI Improves Multiple Sclerosis Lesion Segmentation | |
| Xu | A Robust and Efficient Framework for Slice-to-Volume Reconstruction: Application to Fetal MRI | |
| CN113066145B (en) | Deep learning-based rapid whole-body diffusion weighted imaging method and related equipment | |
| CN117333571B (en) | Reconstruction method, system, equipment and medium of magnetic resonance image | |
| CN116740217B (en) | Arterial spin marking method, device and storage medium based on artificial intelligence technology | |
| CN112505598B (en) | Quantitative magnetic susceptibility imaging reconstruction method and system, storage medium and terminal | |
| CN114972562B (en) | Fast magnetic resonance imaging method combining coil sensitivity estimation and image reconstruction | |
| CN113808176B (en) | MR image registration method, analysis method and device based on device imaging coordinates | |
| US12019133B2 (en) | Systems, methods, and media for estimating a mechanical property based on a transformation of magnetic resonance elastography data using a trained artificial neural network | |
| Xu et al. | A motion assessment method for reference stack selection in fetal brain MRI reconstruction based on tensor rank approximation | |
| WO2024227134A2 (en) | Systems and methods for deep learning-based mri reconstruction with artificial fourier transform (aft) |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |