Brain age prediction method and system based on magnetic resonance imaging based on quantitative index text

Abstract

The present invention discloses a method and system for predicting brain age from a nuclear magnetic resonance image based on quantitative indicator text, the method comprising the following steps: obtaining a brain nuclear magnetic resonance image data set, slicing the brain nuclear magnetic resonance image to obtain a two-dimensional image, segmenting and extracting the two-dimensional image to obtain a brain tissue region map, and calculating the quantitative indicator of the brain tissue corresponding to the brain tissue region map; inputting the quantitative indicator into a large language model to construct a quantitative indicator text; inputting the quantitative indicator text into a text feature extraction network, an attention cross enhancement module, and a residual fusion module in sequence to obtain a quantitative indicator regulation factor; splicing the nuclear magnetic resonance image and the tissue attention map to obtain an input matrix, extracting the brain structure features of the nuclear magnetic resonance image through a two-dimensional convolution residual network feature; splicing the quantitative indicator regulation factor and the brain structure features of the nuclear magnetic resonance image through a linear layer regression to obtain a brain age prediction value. The present invention improves the accuracy and interpretability of brain age prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of brain age prediction, and in particular to a method and system for predicting brain age from a nuclear magnetic resonance image based on quantitative indicator text. Background Art

[0002] Brain age prediction has been one of the research hotspots in the field of neuroscience and medical imaging in recent years. Existing brain age prediction methods have the following defects: (1) Most of them rely on single feature input and cannot capture the complex changes of the brain; (2) When processing complex brain structures and multimodal data, there is a lack of effective inter-regional feature fusion, which affects the reliability of the prediction results; (3) Relying on traditional image processing methods, they ignore deep semantic information, resulting in the failure to fully explore the potential of magnetic resonance imaging; (4) They are only applicable to specific age groups and lack generalization capabilities. Summary of the invention

[0003] In order to overcome the defects and shortcomings of the prior art, the present invention provides a method and system for predicting brain age from magnetic resonance images based on quantitative indicator text. The present invention constructs morphological indicators and signal strength indicators into quantitative indicator text through a large language model, thereby enhancing the expression ability of text features, and improving the effectiveness of text features by using a text feature extraction network. The text features are enhanced based on an attention cross-enhancement module, and the text features are fused by a residual fusion module, thereby achieving deep synergistic enhancement of morphological features and signal strength features. Attention is guided according to the attention weighting of the brain tissue area, and the fine-grained feature extraction of brain age prediction is improved to the brain tissue level. Dynamic correction is performed based on the quantitative indicator regulation factor, thereby solving the weight imbalance problem caused by simple splicing of multimodal features, and achieving directional correction of image features by text features through linear layer regression regulation, thereby improving the accuracy and interpretability of the prediction results.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] The present invention provides a method for predicting brain age from magnetic resonance imaging based on quantitative indicator text, comprising the following steps:

[0006] Acquire a brain magnetic resonance image data set with known brain age results, slice the brain magnetic resonance image to obtain a two-dimensional image, segment and extract the two-dimensional image to obtain a brain tissue regional map, and calculate quantitative indicators of the brain tissue corresponding to the brain tissue regional map, wherein the quantitative indicators include morphological indicators and signal intensity indicators;

[0007] Input the quantitative indicators into the large language model to construct the quantitative indicator text;

[0008] The quantitative indicator text is sequentially passed through the text feature extraction network to extract text features, the attention cross enhancement module to enhance text features, and the residual fusion module to fuse text features to obtain the quantitative indicator regulation factor;

[0009] Brain tissue area Figure 2 The tissue attention map is obtained by quantization, and the MRI image and the tissue attention map are spliced ​​to obtain the input matrix. The two-dimensional convolutional residual network extracts features from the input matrix to obtain the brain structure features of the MRI image.

[0010] The quantitative index regulatory factors and brain structure characteristics of magnetic resonance imaging were spliced ​​and then subjected to linear layer regression to obtain the predicted value of brain age.

[0011] As a preferred technical solution, the two-dimensional image is segmented and extracted to obtain a brain tissue area map, and specifically, the image segmentation is performed using a nnUnet segmentation tool, a freesurfer segmentation tool or a freesurfer segmentation network.

[0012] As a preferred technical solution, the quantitative indicators include morphological indicators and signal strength indicators, specifically including:

[0013] When calculating the morphological features, each brain tissue region map is binarized and the area ratio m is calculated. 1 , roundness m 2 , firmness m 3 , and use Shannon entropy to calculate the fractal dimension m 4 , the boundary curvature is calculated by the gradient after Gaussian filtering to obtain the average curvature m 5 , divide the image into left and right hemispheres, and calculate the area of ​​the left hemisphere m 6 , right hemisphere area m 7 and the asymmetry ratio m 8 ;

[0014] When calculating the signal intensity feature, the two-dimensional image and the binarized brain tissue area map are multiplied to calculate the average signal intensity s 1 , standard deviation 2 , using standard deviation to approximate contrast s 3 And calculate the homogeneity s 4 , calculate the entropy s through the histogram 5 , and calculate the peak value s of the signal strength histogram 6 , skewness s 7 , Kurtosis s 8 .

[0015] As a preferred technical solution, the quantitative indicators are input into the large language model to construct the quantitative indicator text, which specifically includes:

[0016] General description of structural morphological features M 0, the template is "This is a nuclear magnetic resonance image taken from [viewing angle], and the morphological characteristics of [tissue] are calculated", [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and the morphological index description text M is constructed. 1 ~M 8 , the construction template is "[index] is m x ”, [Indicator] and m x Replace with the corresponding indicator type and indicator value to obtain the morphological indicator text M 0 ~M 8 ;

[0017] Construct the overall description text S of the signal strength feature 0 , the template is constructed as "This is a nuclear magnetic resonance image captured from [viewing angle], and the signal intensity characteristics of [tissue] are calculated", [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and the signal intensity index description text S is constructed. 1 ~S 8 , the construction template is "[Indicator] is s x ”, [Indicator] and s x Replace with the corresponding indicator type and indicator value to obtain the signal strength text S 0 ~S 8 .

[0018] As a preferred technical solution, the quantitative indicator text is sequentially subjected to the text feature extraction network to extract text features, the attention cross enhancement module to enhance text features, and the residual fusion module to fuse text features to obtain the quantitative indicator regulation factor, specifically including:

[0019] The quantitative indicator text is input into the text feature extraction network to obtain the quantitative indicator text features, which are specifically expressed as:

[0020] ;

[0021] ;

[0022] in, Represents morphological text features, Indicates the signal strength text feature, and They represent the sentences corresponding to the morphological index text and the signal strength index respectively. The value range of x is 0~8, and n represents the length of the feature;

[0023] The attention cross-enhancement module enhances the morphological text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get:

[0024] ;

[0025] ;

[0026] ;

[0027] Weighted summation to get the first multi-head attention output :

[0028] ;

[0029] ;

[0030] ;

[0031] in, express function;

[0032] Perform linear transformation to obtain morphologically enhanced text features :

[0033] ;

[0034] in, represents a linear transformation;

[0035] The attention cross-enhancement module enhances the signal strength text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get:

[0036] ;

[0037] ;

[0038] ;

[0039] Weighted summation to get the second multi-head attention output :

[0040] ;

[0041] ;

[0042] ;

[0043] Perform linear transformation to obtain signal strength enhanced text features :

[0044] ;

[0045] To concatenate the morphologically enhanced text features and the signal strength enhanced text features and input them into the residual fusion module, the projection features are obtained by projecting through the linear projection layer. :

[0046] ;

[0047] in, Indicates spelling operation. Represents a projection operation;

[0048] The projection features are transformed nonlinearly through the GELU activation function, and then linear transformation, dropout operation and residual connection operation are performed to obtain the residual features. :

[0049] ;

[0050] The normalization process yields the quantitative index control factor:

[0051] ;

[0052] in, represents the quantitative index control factor, It is the data length of the quantitative indicator control factor.

[0053] As a preferred technical solution, the magnetic resonance image and tissue attention map are concatenated to obtain a 2-channel matrix, which is converted into a 3-channel matrix through a layer of convolution and input into a 2D convolutional residual network.

[0054] As a preferred technical solution, the two-dimensional convolutional residual network is provided with multiple residual block groups, a global pooling layer, and a linear layer, each residual block group contains a different number of residual blocks, and each residual block contains two convolutional layers and a jump connection;

[0055] The global pooling layer aggregates the output feature map of the last residual block group in the spatial dimension, converts the feature map of each feature channel into a numerical value, reduces the three-dimensional feature map into a two-dimensional feature vector, and inputs the two-dimensional feature vector into the linear layer to obtain the brain structure features of the magnetic resonance imaging.

[0056] As a preferred technical solution, the quantitative index control factor and the brain structure characteristics of the magnetic resonance image are spliced ​​and then subjected to linear layer regression to obtain the brain age prediction value, which specifically includes:

[0057] Quantitative index control factor and MRI brain structural features Spliced ​​into a multimodal regulation feature vector, expressed as:

[0058] ;

[0059] in, represents the multimodal control feature vector, Indicates the data length of the quantitative indicator control factor, The length of data representing brain structural features of MRI images;

[0060] Output brain age prediction value after linear layer regression :

[0061] ;

[0062] in, represents the connection parameters of the linear layer, Represents the activation function.

[0063] As a preferred technical solution, the activation function adopts ReLU function, sigmoid function or tanh function.

[0064] The present invention also provides a nuclear magnetic resonance image brain age prediction system based on quantitative indicator text, comprising: a brain nuclear magnetic resonance image data set acquisition module, a quantitative indicator calculation module, a quantitative indicator text construction module, a quantitative indicator regulation factor construction module, a nuclear magnetic resonance image brain structure feature extraction module, and a brain age prediction module;

[0065] The brain magnetic resonance image data set acquisition module is used to acquire a brain magnetic resonance image data set with known brain age results;

[0066] The quantitative index calculation module calculates quantitative indexes, slices the brain magnetic resonance image to obtain a two-dimensional image, segments and extracts the two-dimensional image to obtain a brain tissue region map, and calculates quantitative indexes of brain tissue corresponding to the brain tissue region map, wherein the quantitative indexes include morphological indexes and signal intensity indexes;

[0067] The quantitative indicator text construction module is used to input the quantitative indicator into the large language model to construct the quantitative indicator text;

[0068] The quantitative indicator control factor construction module is used to extract text features from the quantitative indicator text in sequence through a text feature extraction network, enhance text features through an attention cross enhancement module, and fuse text features through a residual fusion module to obtain a quantitative indicator control factor;

[0069] The magnetic resonance imaging brain structure feature extraction module is used to extract the brain tissue area Figure 2 The tissue attention map is obtained by quantization, the MRI image and the tissue attention map are concatenated to obtain the input matrix, and the input matrix is ​​subjected to feature extraction based on a two-dimensional convolutional residual network to obtain the brain structure features of the MRI image;

[0070] The brain age prediction module is used to obtain a brain age prediction value by splicing the quantitative index control factor and the brain structure features of the magnetic resonance imaging image and performing linear layer regression.

[0071] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0072] (1) The present invention applies morphological indicators and signal strength indicators to brain age prediction, and constructs the two indicators into quantitative indicator text through a large language model, thereby solving the technical problem of a single structured quantitative indicator text mode and enhancing the text feature expression capability through semantic diversity;

[0073] By leveraging the powerful capabilities of the text feature extraction network, the detailed feature information of each brain tissue is fully extracted, solving the problem of poor adaptability of general text models to medical terminology, and achieving accurate encoding of biomedical semantics in the feature extraction stage, thus improving the effectiveness of text features.

[0074] The text features are enhanced by the attention cross-enhancement module and the text features are fused by the residual fusion module, which solves the technical problem of insufficient interaction of multimodal features, realizes the deep synergistic enhancement of morphological features and signal intensity features, and effectively improves the feature fusion and information transmission capabilities in the brain age prediction process;

[0075] The MRI images and tissue attention maps are spliced ​​together to guide attention according to the attention weighting of brain tissue regions, solving the technical problem that traditional convolutional networks ignore the importance of brain tissue regions and improving the granularity of feature extraction for brain age prediction to the brain tissue level.

[0076] The quantitative indicator regulation factors and the brain structure features of magnetic resonance imaging are spliced ​​together and then regressed through a linear layer to obtain the predicted value of brain age. Dynamic correction is performed based on the quantitative indicator regulation factors to solve the weight imbalance problem caused by simple splicing of multimodal features. The linear layer regression regulation is used to achieve the directional correction of image features by text features, thereby improving the accuracy and interpretability of the prediction results.

[0077] (2) The present invention adopts an area ratio m 1, roundness m 2 , firmness m 3 , fractal dimension m 4 , mean curvature m 5 , left hemisphere area m 6 , right hemisphere area m 7 and the asymmetry ratio m 8 As a morphological indicator, it solves the problem that traditional morphological characteristics are insufficient in representing complex structures. The newly added indicators increase the overall sensitivity to abnormal brain development.

[0078] (3) The present invention adopts the average signal strength s 1 , standard deviation 2 , Contrast 3 Homogeneity 4 , Entropy 5 , the peak value s of the signal strength histogram 6 , skewness s 7 , Kurtosis s 8 As a signal strength indicator, it solves the defect of traditional strength analysis that ignores distribution characteristics. Its skewness s 7 , Kurtosis s 8 Equivalent statistics expand the signal feature dimension to 8 dimensions, increasing the amount of feature information.

[0079] (4) The present invention fully mines the potential information in brain magnetic resonance images based on multi-view slices, brain tissue segmentation and multi-dimensional index calculation, and solves the technical problems of insufficient single-view information and single feature dimension in traditional brain age prediction methods. Compared with the existing methods that can only predict the brain age of a single age group, the present invention is suitable for brain age prediction tasks at all age groups in the human life cycle. The tests were conducted on the fetal dataset FetalBrain-GD, the infant dataset DHCP and the elderly dataset OASIS3. The average brain age errors were 0.739 weeks, 0.771 weeks and 3.882 years, respectively, with a high accuracy rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a flow chart of the method for predicting brain age from magnetic resonance images based on quantitative indicator texts of the present invention;

[0081] Figure 2 Schematic diagram of the overall framework of the two-dimensional convolutional residual network module of the present invention;

[0082] Figure 3 Schematic diagram of the overall network framework of the linear layer of the present invention;

[0083] Figure 4 This is a data comparison chart of the predicted gestational age and the actual gestational age in this embodiment 2;

[0084] Figure 5This is a data comparison chart of the predicted gestational age and the actual gestational age in this embodiment 3;

[0085] Figure 6 This is a data comparison chart of the predicted gestational age and the actual gestational age in this embodiment 4;

[0086] Figure 7 This is a schematic diagram of the overall implementation framework of the magnetic resonance imaging brain age prediction system based on quantitative indicator text in Example 5. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0088] Example 1

[0089] like Figure 1 As shown, this embodiment provides a method for predicting brain age from magnetic resonance imaging based on quantitative indicator text, comprising the following steps:

[0090] S1: Obtain a brain MRI image dataset with known brain age results, slice the dataset, segment brain tissue, and calculate indicators;

[0091] Slice refers to the three middle two-dimensional images cut out from the three perspectives of sagittal, coronal and transverse planes in the brain MRI image. Finally, one brain MRI image can obtain 9 two-dimensional images.

[0092] Brain tissue segmentation refers to segmenting the two-dimensional image separately through segmentation tools, extracting the regional maps of various brain tissues from the MRI image. The extracted brain tissue categories are gray matter, white matter, cerebellum, brain stem, etc. Finally, the two-dimensional image and the binarized brain tissue regional map constitute a data pair. The brain tissue regional map is binarized to obtain the tissue attention map;

[0093] Index calculation refers to calculating the morphological index and signal intensity index of the corresponding brain tissue in the data pair. First, the morphological features are calculated, and the area, perimeter, etc. of each brain tissue are binarized and calculated, and then the area ratio m is obtained. 1 , roundness m 2 , firmness m 3 , and use Shannon entropy to calculate the fractal dimension m 4 , the boundary curvature is calculated by the gradient after Gaussian filtering to obtain the average curvature m 5 The image is also divided into left and right hemispheres to calculate the left hemisphere area m 6 , right hemisphere area m 7 and the asymmetry ratio m 8; Then calculate the signal intensity feature, multiply the two-dimensional image and the binarized brain tissue area map, and calculate its average signal intensity s 1 , standard deviation 2 , approximate the contrast s using the standard deviation 3 And calculate the homogeneity s 4 , calculate the entropy s through the histogram 5 , and calculate the peak value s of the signal strength histogram 6 , skewness s 7 , Kurtosis s 8 ;

[0094] S2: The quantitative indicators are used to construct a preliminary text according to a fixed template, and then sent to a large language model for polishing and rewriting to ensure the diversity of the text. The large language model in this embodiment is preferably the large language model ChatGPT4.0;

[0095] First, construct the overall description text M of the morphological features 0 , the template is constructed as "This is a nuclear magnetic resonance image taken from [viewing angle], and the morphological characteristics of [tissue] are calculated", where [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and then the morphological index description text M is constructed. 1 ~M 8 , the construction template is "[index] is m x ”, where [index] and m x Replace with the corresponding indicator type and indicator value;

[0096] Similarly, construct the overall description text S of the signal strength feature 0 , the template is constructed as "This is a nuclear magnetic resonance image taken from [viewing angle], and the signal intensity characteristics of [tissue] are calculated", where [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and then the signal intensity index description text S is constructed. 1 ~S 8 , the construction template is "[Indicator] is s x ”, where [index] and s x Replace with the corresponding indicator type and indicator value;

[0097] S3: Extract quantitative indicator text features based on the text feature extraction network, send the extracted quantitative indicator text features to the attention cross enhancement module to obtain morphological enhanced text features and signal strength enhanced text features, and then input the morphological enhanced text features and signal strength enhanced text features into the residual fusion module to obtain the quantitative indicator regulation factor , is the length;

[0098] In this embodiment, the specific implementation process of step S3 is as follows:

[0099] S31: The morphological index text M constructed in step S2 0 ~M 8 and signal strength indicator text S 0 ~S 8 Input the text feature extraction network in sequence to obtain the quantitative index text features. The features of each sentence can be expressed as a one-dimensional vector. The morphological text features are expressed as , the signal strength text feature is expressed as ,in, and The subscript x represents different sentences, the value range of x is 0~8, and n represents the length of the feature;

[0100] S32: The attention cross enhancement module enhances the morphological text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get: , and , and then calculate the attention score: ,pass Function to calculate attention weights , use the attention weights to weight the values ​​and get the first multi-head attention output , adjust its dimension and perform linear transformation to obtain morphologically enhanced text features ,in, represents a linear transformation;

[0101] Similarly, the attention cross-enhancement module enhances the signal strength text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get: , and Then calculate the attention score: ,pass Function to calculate attention weights , use the attention weights to weight the values ​​and get the second multi-head attention output , adjust its dimension and perform linear transformation to obtain signal strength enhanced text features ;

[0102] S33: Enhance text features with morphology and signal strength enhancement text features After splicing, the input is input into the residual fusion module, and the input is projected through the linear projection layer to obtain the projection feature , It represents the concatenation operation, which transforms the projection features nonlinearly through the GELU activation function, and then performs linear transformation, dropout and residual connection operations to obtain the residual features. , where + represents the residual connection operation, and the quantitative index control factor is obtained by normalization, which is expressed as: , is the data length;

[0103] S4: The MRI image and the tissue attention map are concatenated to form a 2-channel matrix as input, which is first converted into a 3-channel matrix by a convolution layer and then sent to a 2D convolutional residual network to extract the brain structure features of the MRI image with tissue attention, expressed as: , is the data length;

[0104] like Figure 2 As shown in Figure 1, the 2D convolutional residual network has 4 residual block groups, each of which contains a different number of residual blocks, namely 3, 4, 6, and 3 residual blocks, for a total of 16 residual blocks. Each residual block contains two convolutional layers and a skip connection.

[0105] The first residual block group contains 3 residual blocks, and the input of the first residual block is the image data to be processed. For each residual block in this group, its output is jump-joined with the input and used as the input of the next residual block. The output processed by the first residual block group is then used as the input of the second residual block group. The second residual block group also contains 4 residual blocks, which are processed in sequence in the same jump-joining manner. Similarly, after being processed by the third residual block group (containing 6 residual blocks) and the fourth residual block group (containing 3 residual blocks), the obtained feature map is output to the global pooling layer.

[0106] The global pooling layer aggregates the output feature map of the last residual block group in the spatial dimension and converts the feature map of each feature channel into a numerical value, thereby reducing the three-dimensional feature map to a two-dimensional feature vector. Finally, the two-dimensional feature vector is input into the linear layer to obtain the brain structure feature of the MRI image, which is expressed as , is the length.

[0107] S5: Quantitative indicators control factors and MRI brain structural features Spliced ​​into a multimodal regulation feature vector, expressed as:

[0108] ;

[0109] in, Preferably 768, Preferably, it is 2048, and the superscripts q and i are used to distinguish different vectors;

[0110] like Figure 3 As shown in the figure, the quantitative indicator regulation prediction module outputs the predicted value of brain age through linear layer regression, which is expressed as:

[0111] ;

[0112] in, represents the connection parameters of the linear layer, represents an activation function. The activation function of this embodiment can be any one of the ReLU function, sigmoid function, and tanh function.

[0113] In this embodiment, based on the generated quantitative indicator text, the magnetic resonance image and the tissue attention map are spliced ​​to form a matrix with 2 channels as input, and trained through the text feature extraction network, the attention cross enhancement module, the residual fusion module, the two-dimensional convolution residual network, and the quantitative indicator regulation and prediction module, a model capable of predicting fetal brain age is obtained, wherein the text feature extraction network uses "biobert-v1.1", the training process uses the gradient descent method, and the network uses the loss function ,in, To predict age, is the real age, and the learning rate is set at 1e -3 To 2e -3 The experimental method used five-fold three-fold cross validation, and the experimental evaluation method used three evaluation indicators, namely, the average brain age error, R 2 Correlation coefficient, accuracy;

[0114] Among them, the average brain age error is: , R 2 The correlation coefficient is the Pearson correlation coefficient between the predicted brain age and the actual brain age. The accuracy is: , It indicates the number of brain age errors less than 1 in the dataset, and N is the total number of data in the dataset.

[0115] Example 2

[0116] This embodiment is based on the implementation process of the method for predicting brain age from magnetic resonance images based on quantitative indicator text in Embodiment 1, and specifically selects the FetalBrain-GD dataset, which contains 202 and 50 samples in the training set and the test set, respectively, with the brain age ranging from gestational weeks 20 to 38;

[0117] The nnUnet segmentation tool was used for brain tissue segmentation. The extracted brain tissue categories were 7 types of brain tissues: extracerebral spinal fluid, gray matter, white matter, ventricle, cerebellum, deep gray matter, and brainstem. The morphological index and signal intensity index of each brain tissue were calculated through the index calculation program.

[0118] In this embodiment, the final average brain age prediction error, R 2 The correlation coefficient and accuracy can reach 0.739, 0.970, and 72.8% respectively. Table 1 is the comparison results of this embodiment with other methods based on convolutional neural networks and some of their derivatives on the FetalBrain-GD dataset. It can be seen that the method proposed in the present invention has achieved the best results. Figure 4 As shown, the comparison results of the predicted brain age and the actual brain age of 250 data on FetalBrain-GD are obtained. The horizontal axis in the figure is the actual brain age, and the vertical axis is the predicted brain age. The dotted line in the figure is the fitting line of the scatter plot, and each point corresponds to a different sample. It can be seen that the brain age predicted by the method proposed in the present invention is very small compared with the actual brain age.

[0119] Table 1 Comparison of experimental results of the brain age prediction method of this embodiment and other methods

[0120]

[0121] Example 3

[0122] This embodiment is based on the implementation process of the method for predicting brain age from magnetic resonance images based on quantitative indicator text in embodiment 1, and specifically selects a DHCP dataset, which includes 703 and 175 samples in the training set and the test set, respectively, with a brain age range of 26 to 45 weeks for newborns;

[0123] The brain tissue segmentation was performed using the FreeSurfer segmentation tool. The extracted brain tissue categories were 8 types of brain tissues, including extracerebral spinal fluid, gray matter, white matter, ventricle, cerebellum, deep gray matter, brainstem, and hippocampus. The morphological index and signal intensity index of each brain tissue were then calculated using an index calculation program.

[0124] The final average brain age prediction error, R 2The correlation coefficient and accuracy can reach 0.771, 0.960, and 70.5% respectively. Table 2 is the comparison results of this embodiment on DHCP with other methods based on convolutional neural networks and some of their derivatives. It can be seen that the method proposed in the present invention has achieved the best results. Figure 5 As shown, the comparison results of the predicted brain age and the actual brain age of 250 data on DHCP by the brain age prediction method of this embodiment are obtained. In the figure, the horizontal axis is the actual brain age, and the vertical axis is the predicted brain age. It can be seen that the difference between the brain age predicted by the method proposed in the present invention and the actual brain age is very small.

[0125] Table 2 Comparison of experimental results of the brain age prediction method of this embodiment and other methods

[0126]

[0127] Example 4

[0128] This embodiment is based on the implementation process of the method for predicting brain age from magnetic resonance images based on quantitative indicator text in Embodiment 1, and specifically selects the OASIS3 dataset, which contains 922 and 230 samples in the training set and the test set, respectively, with brain ages ranging from 42 to 97 years old;

[0129] The freesurfer segmentation network is used for brain tissue segmentation. The extracted brain tissue categories are gray matter, white matter, ventricle, cerebellum, deep gray matter and brain stem. Then the morphological indicators and signal intensity indicators of each brain tissue are calculated through the indicator calculation program.

[0130] The final average brain age prediction error, R 2 The correlation coefficient and accuracy can reach 3.882, 0.854, and 70.6% respectively. Table 3 is the comparison results of the classification method provided in this embodiment with other methods based on convolutional neural networks and some of their derivatives on OASIS3. It can be seen that the method proposed in the present invention has achieved the best results. Figure 6 As shown, the comparison results of the predicted brain age and the actual brain age of 250 data on OASIS3 by the brain age prediction method of this embodiment are obtained. The horizontal axis is the actual brain age and the vertical axis is the predicted brain age. It can be seen that the difference between the brain age predicted by the method proposed in the present invention and the actual brain age is very small.

[0131] Table 3 Comparison of experimental results of the brain age prediction method of this embodiment and other methods

[0132]

[0133] Example 5

[0134] like Figure 7As shown, this embodiment provides a nuclear magnetic resonance image brain age prediction system based on quantitative indicator text, which is used to implement the nuclear magnetic resonance image brain age prediction method based on quantitative indicator text in the above embodiment 1, and the system includes: a brain nuclear magnetic resonance image data set acquisition module, a quantitative indicator calculation module, a quantitative indicator text construction module, a quantitative indicator regulation factor construction module, a nuclear magnetic resonance image brain structure feature extraction module, and a brain age prediction module;

[0135] In this embodiment, the brain magnetic resonance image data set acquisition module is used to acquire a brain magnetic resonance image data set with known brain age results;

[0136] In this embodiment, the quantitative index calculation module calculates quantitative indexes, slices the brain magnetic resonance image to obtain a two-dimensional image, segments and extracts the two-dimensional image to obtain a brain tissue region map, and calculates quantitative indexes of the brain tissue corresponding to the brain tissue region map, wherein the quantitative indexes include morphological indexes and signal intensity indexes;

[0137] In this embodiment, the quantitative indicator text construction module is used to input the quantitative indicator into the large language model to construct the quantitative indicator text;

[0138] In this embodiment, the quantitative indicator control factor construction module is used to extract text features from the quantitative indicator text through the text feature extraction network, enhance text features through the attention cross enhancement module, and fuse text features through the residual fusion module to obtain the quantitative indicator control factor;

[0139] In this embodiment, the brain structure feature extraction module of the magnetic resonance image is used to extract the brain tissue area Figure 2 The tissue attention map is obtained by quantization, the MRI image and the tissue attention map are concatenated to obtain the input matrix, and the input matrix is ​​subjected to feature extraction based on a two-dimensional convolutional residual network to obtain the brain structure features of the MRI image;

[0140] In this embodiment, the brain age prediction module is used to obtain a brain age prediction value by concatenating the quantitative index control factor and the brain structure features of the magnetic resonance imaging image and performing linear layer regression.

[0141] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for predicting brain age from magnetic resonance imaging based on quantitative indicator text, characterized in that: The steps include: Acquire a brain magnetic resonance image data set with known brain age results, slice the brain magnetic resonance image to obtain a two-dimensional image, segment and extract the two-dimensional image to obtain a brain tissue regional map, and calculate quantitative indicators of the brain tissue corresponding to the brain tissue regional map, wherein the quantitative indicators include morphological indicators and signal intensity indicators; Input the quantitative indicators into the large language model to construct the quantitative indicator text, including: Construct the overall description text M0 of the morphological characteristics, and the construction template is "This is a nuclear magnetic resonance image taken from [viewing angle], and the morphological characteristics of [tissue] are calculated", [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and construct the morphological index description text M1~M8, and the construction template is "[index] is m x ”, [Indicator] and m x Replace with the corresponding indicator type and indicator value to obtain the morphological indicator text M0~M8; Construct the overall description text S0 of the signal intensity feature, and the construction template is "This is a nuclear magnetic resonance image intercepted from [viewing angle], and the signal intensity characteristics of [tissue] are calculated", [viewing angle] and [tissue] are replaced by the corresponding viewing angle and tissue type for each data pair, and construct the signal intensity index description text S1~S8, and the construction template is "[index] is s x ”, [Indicator] and s x Replace with the corresponding indicator type and indicator value to obtain the signal strength text S0~S8; The quantitative indicator text is sequentially passed through the text feature extraction network to extract text features, the attention cross enhancement module to enhance text features, and the residual fusion module to fuse text features to obtain the quantitative indicator regulation factor; The brain tissue area map is binarized to obtain a tissue attention map, and the MRI image and the tissue attention map are concatenated to obtain an input matrix. A two-dimensional convolutional residual network is used to extract features from the input matrix to obtain brain structure features of the MRI image. The quantitative index regulatory factors and brain structure characteristics of magnetic resonance imaging were spliced ​​and then subjected to linear layer regression to obtain the predicted value of brain age.

2. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The two-dimensional image is segmented and extracted to obtain a brain tissue region map, and the image segmentation is performed using a nnUnet segmentation tool, a freesurfer segmentation tool or a freesurfer segmentation network.

3. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The quantitative indicators include morphological indicators and signal strength indicators, specifically including: When calculating morphological features, each brain tissue region map was binarized and the area ratio m1, roundness m2, and compactness m3 were calculated. At the same time, the fractal dimension m4 was calculated using Shannon entropy. The boundary curvature was calculated by the gradient after Gaussian filtering to obtain the average curvature m5. The image was divided into left and right hemispheres, and the left hemisphere area m6, right hemisphere area m7, and asymmetry ratio m8 were calculated. When calculating the signal intensity feature, the two-dimensional image and the binarized brain tissue area map are multiplied to calculate the average signal intensity s1 and standard deviation s2. The standard deviation is used to approximate the contrast s3 and the homogeneity s4. The entropy s5 is calculated through the histogram. At the same time, the peak s6, skewness s7, and kurtosis s8 of the signal intensity histogram are calculated.

4. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The quantitative indicator text is sequentially passed through the text feature extraction network to extract text features, the attention cross enhancement module to enhance text features, and the residual fusion module to fuse text features to obtain the quantitative indicator regulation factor, which specifically includes: The quantitative indicator text is input into the text feature extraction network to obtain the quantitative indicator text features, which are specifically expressed as: ； ； in, Represents morphological text features, Indicates the signal strength text feature, and They represent the sentences corresponding to the morphological index text and the signal strength index respectively. The value range of x is 0~8, and n represents the length of the feature; The attention cross-enhancement module enhances the morphological text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get: ； ； ； Weighted summation to get the first multi-head attention output : ； ； ； in, express function; Perform linear transformation to obtain morphologically enhanced text features : ； in, represents a linear transformation; The attention cross-enhancement module enhances the signal strength text features. As a query vector , As a key vector Sum value vector , for the query vector , key vector Sum value vector Perform linear transformation respectively to get: ； ； ； Weighted summation to get the second multi-head attention output : ； ； ； Perform linear transformation to obtain signal strength enhanced text features : ； To concatenate the morphologically enhanced text features and the signal strength enhanced text features and input them into the residual fusion module, the projection features are obtained by projecting through the linear projection layer. : ； in, Indicates spelling operation. Represents a projection operation; The projection features are transformed nonlinearly through the GELU activation function, and then linear transformation, dropout operation and residual connection operation are performed to obtain the residual features. : ； The normalization process yields the quantitative index control factor: ； in, represents the quantitative index control factor, It is the data length of the quantitative indicator control factor.

5. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The MRI image and tissue attention map are concatenated to obtain a 2-channel matrix, which is converted into a 3-channel matrix through a layer of convolution and input into the 2D convolutional residual network.

6. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The two-dimensional convolutional residual network is provided with a plurality of residual block groups, a global pooling layer, and a linear layer, each residual block group contains a different number of residual blocks, and each residual block contains two convolutional layers and a skip connection; The global pooling layer aggregates the output feature map of the last residual block group in the spatial dimension, converts the feature map of each feature channel into a numerical value, reduces the three-dimensional feature map into a two-dimensional feature vector, and inputs the two-dimensional feature vector into the linear layer to obtain the brain structure features of the magnetic resonance imaging.

7. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 1, characterized in that: The quantitative index regulatory factors and the brain structure characteristics of the MRI images are spliced ​​and then subjected to linear layer regression to obtain the predicted value of brain age, which includes: Quantitative index control factor and MRI brain structural features Spliced ​​into a multimodal regulation feature vector, expressed as: ； in, represents the multimodal control feature vector, Indicates the data length of the quantitative indicator control factor, The length of data representing brain structural features in MRI images; Output brain age prediction value after linear layer regression : ； in, represents the connection parameters of the linear layer, Represents the activation function.

8. The method for predicting brain age from magnetic resonance images based on quantitative indicator text according to claim 7, characterized in that: The activation function uses a ReLU function, a sigmoid function or a tanh function.

9. A magnetic resonance imaging brain age prediction system based on quantitative indicator text, characterized in that: The method for predicting brain age based on nuclear magnetic resonance images based on quantitative indicator text according to any one of claims 1 to 8 comprises: a brain nuclear magnetic resonance image data set acquisition module, a quantitative indicator calculation module, a quantitative indicator text construction module, a quantitative indicator regulation factor construction module, a nuclear magnetic resonance image brain structure feature extraction module, and a brain age prediction module; The brain magnetic resonance image data set acquisition module is used to acquire a brain magnetic resonance image data set with known brain age results; The quantitative index calculation module is used to slice the brain magnetic resonance image to obtain a two-dimensional image, segment and extract the two-dimensional image to obtain a brain tissue area map, and calculate the quantitative index of the brain tissue corresponding to the brain tissue area map, wherein the quantitative index includes a morphological index and a signal intensity index; The quantitative indicator text construction module is used to input the quantitative indicator into the large language model to construct the quantitative indicator text; The quantitative indicator control factor construction module is used to extract text features from the quantitative indicator text in sequence through the text feature extraction network, enhance text features through the attention cross enhancement module, and fuse text features through the residual fusion module to obtain the quantitative indicator control factor; The brain structure feature extraction module of the nuclear magnetic resonance image is used to binarize the brain tissue area map to obtain a tissue attention map, splice the nuclear magnetic resonance image and the tissue attention map to obtain an input matrix, and extract features from the input matrix based on a two-dimensional convolutional residual network to obtain brain structure features of the nuclear magnetic resonance image; The brain age prediction module is used to obtain a brain age prediction value by splicing the quantitative index control factor and the brain structure features of the magnetic resonance imaging image and performing linear layer regression.

Images