Multi-parameter MRI brain age prediction method and device based on uncertainty perception, equipment and medium

Abstract

The invention discloses a multi-parameter MRI brain age prediction method and device based on uncertainty perception, equipment and a medium, and the method comprises the steps: determining a first feature sequence of an sMRI image, a second feature sequence of a DTI image, and a third feature sequence of an fMRI image through a feature extraction module; and determining a GM sampling vector, a DTI sampling vector and an fMRI sampling vector based on the first feature sequence, the second feature sequence and the third feature sequence by using a probability distribution encoder, and performing brain age prediction by using the GM sampling vector, the DTI sampling vector and the fMRI sampling vector. According to the method and the device, the fusion feature vector determined by fusing the sMRI, DTI and fMRI information is utilized to perform brain age prediction, so that multi-dimensional information of the brain structure and function can be more comprehensively acquired, and the accuracy of brain age prediction is improved. Meanwhile, a probability distribution encoder is introduced in the process of fusing the multi-parameter MRI images, multi-parameter MRI feature mapping is converted into Gaussian distribution through the probability distribution encoder to model feature uncertainty, richer potential information in the multi-parameter MRI images is captured, and the accuracy of brain age prediction is further improved.

Description

Technical Field

[0001] This application relates to the field of biomedical engineering technology, and in particular to a method, device, equipment and medium for predicting brain age based on uncertainty perception using multi-parameter MRI. Background Technology

[0002] Currently, brain health is generally assessed through clinical manifestations and scale scores, supplemented by neuroimaging examinations. Compared to clinical manifestations and scale scores, MRI (Magnetic Resonance Imaging) can objectively reflect multi-dimensional information about brain structure and function, thereby predicting the state and health of an individual's brain relative to their actual age.

[0003] When using MRI for brain age prediction, a common approach is to construct a brain age prediction network using single-parameter MRI, and then use this network to predict brain age. While this method can predict brain age, the accuracy is low because single-parameter MRI provides limited information on brain structure and function.

[0004] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a method, device, equipment and medium for predicting brain age based on uncertainty perception, in order to address the shortcomings of the existing technology.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a multi-parameter MRI brain age prediction method based on uncertainty perception, which applies a trained brain age fusion prediction model, the brain age fusion prediction model including a feature extraction module, a probability distribution encoder, and a prediction module; the method includes: Acquire sMRI, DTI, and fMRI images of the brain; The feature extraction module determines a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images. Uncertainty extraction is performed on the first feature sequence by a probability distribution encoder to obtain a first mean vector and a first variance vector, and vector sampling is performed based on the first mean vector and the first variance vector to determine the GM sampling vector; Uncertainty extraction is performed on the second feature sequence by a probability distribution encoder to obtain a second mean vector and a second variance vector, and vector sampling is performed based on the second mean vector and the second variance vector to determine the DTI sampling vector; Uncertainty extraction is performed on the third feature sequence using a probability distribution encoder to obtain a third mean vector and a third difference vector, and vector sampling is performed based on the third mean vector and the third difference vector to determine the fMRI sampling vector; The prediction module determines a fusion feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector, and then performs brain age prediction based on the fusion feature vector to obtain the predicted brain age.

[0007] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein the uncertainty extraction process specifically includes: The input feature sequence is divided into a first input feature sequence and a second input feature sequence; An initial mean vector is determined based on a first input feature sequence using a multi-head attention mechanism, and a mean vector is determined based on the initial mean vector and the first input feature sequence. The variance vector is determined based on the second input feature sequence using a multi-head attention mechanism.

[0008] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein the step of determining the initial mean vector based on the first input feature sequence using a multi-head attention mechanism specifically includes: The first input feature sequence is divided equally according to the number of attention heads in the multi-head attention mechanism to obtain a number of sub-input feature sequences. Attention is calculated for each sub-input feature sequence to obtain the attention features corresponding to each sub-input feature sequence; The attention features corresponding to all sub-input feature sequences are concatenated to obtain the initial mean vector.

[0009] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein dividing the input feature sequence into a first input feature sequence and a second input feature sequence specifically involves: The input feature sequence is divided into a first input feature sequence and a second input feature sequence according to the dimensional direction.

[0010] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein the vector sampling process specifically includes: Sample random noise from a standard normal distribution; The mean vector and variance vector are weighted based on the random noise to obtain the sampled feature vector.

[0011] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein determining the fusion feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector specifically includes: The GM sampling vector is fused with the first mean vector to obtain the first fused feature vector; The DTI sampling vector is fused with the second mean vector to obtain the second fused feature vector; The fMRI sampling vector is fused with the third mean vector to obtain the third fused feature vector; The first fused feature vector, the second fused feature vector, and the third fused feature vector are fused to obtain a fused feature vector.

[0012] The aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, wherein, during the training process of the brain age fusion prediction model, a regularized loss function is introduced into the loss function terms corresponding to the GM sampling vector, DTI sampling vector, and fMRI sampling vector, respectively. The regularized loss function is expressed as: , , in, This represents the regularization loss function. The entropy of a multivariate Gaussian distribution is represented by its value. Represents the mean vector. Represents the variance vector. Representing feature dimension, This indicates a preset threshold.

[0013] A second aspect of this application provides a multi-parameter MRI brain age prediction device based on uncertainty perception, which applies a trained brain age fusion prediction model, the brain age fusion prediction model including a feature extraction module, a probability distribution encoder, and a prediction module; the device includes: The acquisition module is used to acquire multi-parameter MRI images of the brain; The control module is configured to: determine a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images using the feature extraction module; extract uncertainty from the first feature sequence using a probability distribution encoder to obtain a first mean vector and a first variance vector, and perform vector sampling based on the first mean vector and the first variance vector to determine a GM sampling vector; extract uncertainty from the second feature sequence using a probability distribution encoder to obtain a second mean vector and a second variance vector, and perform vector sampling based on the second mean vector and the second variance vector to determine a DTI sampling vector; extract uncertainty from the third feature sequence using a probability distribution encoder to obtain a third mean vector and a third variance vector, and perform vector sampling based on the third mean vector and the third variance vector to determine an fMRI sampling vector; and determine a fused feature vector using a prediction module based on the GM sampling vector, the DTI sampling vector, and the fMRI sampling vector, and perform brain age prediction based on the fused feature vector to obtain a predicted brain age.

[0014] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the uncertainty-aware multi-parameter MRI brain age prediction method as described above.

[0015] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in any of the above-described uncertainty-aware multi-parameter MRI brain age prediction methods.

[0016] Beneficial Effects: Compared with existing technologies, this application determines a fusion feature vector by fusing sMRI, DTI, and fMRI information, and then uses this fusion feature vector for brain age prediction. This allows for a more comprehensive acquisition of multi-dimensional information on brain structure and function, overcoming the limitation of information provided by single-parameter MRI and improving the accuracy of brain age prediction. Furthermore, this application introduces a probability distribution encoder during the fusion of multi-parameter MRI images. This encoder transforms the multi-parameter MRI feature mapping into a Gaussian distribution to model feature uncertainty, capturing richer potential information from the multi-parameter MRI images. This improves the robustness and predictive performance of the brain age fusion prediction model, thereby further enhancing the accuracy of brain age prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This diagram illustrates an application scenario for the multi-parameter MRI brain age prediction method based on uncertainty perception provided in this application.

[0019] Figure 2 This is a network structure diagram of the brain age fusion prediction model.

[0020] Figure 3 A flowchart of a multi-parameter MRI brain age prediction method based on uncertainty perception provided in an embodiment of this application.

[0021] Figure 4 A schematic diagram of the principle of the multi-parameter MRI brain age prediction device based on uncertainty perception provided in the embodiments of this application.

[0022] Figure 5 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0023] This application provides a method, apparatus, device, and medium for predicting brain age using multi-parameter MRI based on uncertainty perception. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0025] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0026] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0027] Research has found that current methods for assessing brain health primarily rely on clinical manifestations and scale scores, supplemented by neuroimaging examinations. Compared to clinical manifestations and scale scores, MRI (Magnetic Resonance Imaging) can objectively reflect multidimensional information about brain structure and function, thereby predicting an individual's brain condition and health relative to their actual age.

[0028] When using MRI for brain age prediction, a common approach is to construct a brain age prediction network using single-parameter MRI, and then use this network to predict brain age. While this method can predict brain age, the accuracy is low because single-parameter MRI provides limited information on brain structure and function.

[0029] Therefore, in the embodiments of this application, sMRI (Structural Magnetic Resonance Imaging), DTI (Diffusion Tensor Imaging), and fMRI (functional Magnetic Resonance Imaging) images are acquired. In functional magnetic resonance imaging (fMRI), server 120 uses the feature extraction module to determine a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images. It then uses a probability distribution encoder to extract uncertainty from the first feature sequence to obtain a first mean vector and a first variance vector, and performs vector sampling based on these vectors to determine a GM sampling vector. Similarly, it uses the probability distribution encoder to extract uncertainty from the second feature sequence to obtain a second mean vector and a second variance vector, and performs vector sampling based on these vectors to determine a DTI sampling vector. Finally, it uses a prediction module to determine a fused feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector, and performs brain age prediction based on the fused feature vector to obtain a predicted brain age.

[0030] This application's embodiments determine a fused feature vector by fusing sMRI, DTI, and fMRI information, and then use this fused feature vector for brain age prediction. This allows for a more comprehensive acquisition of multi-dimensional information on brain structure and function, overcoming the limitation of information provided by single-parameter MRI and improving the accuracy of brain age prediction. Furthermore, this application introduces a probability distribution encoder during the fusion of multi-parameter MRI images. This encoder transforms the multi-parameter MRI feature mapping into a Gaussian distribution to model feature uncertainty, capturing richer potential information from the multi-parameter MRI images. This improves the robustness and predictive performance of the brain age fusion prediction model, thereby further enhancing the accuracy of brain age prediction.

[0031] An application environment diagram of the uncertainty-perception-based multi-parameter MRI brain age prediction method provided in this application embodiment can be shown as follows: Figure 1 As shown. (Refer to...) Figure 1This multi-parameter MRI brain age prediction system based on uncertainty perception includes a user terminal 110 and a server 120. The user terminal 110 and server 120 are connected via a network. The user terminal 110 can be a desktop user terminal or a mobile user terminal, and the mobile user terminal 110 can be at least one of a mobile phone, tablet computer, or laptop computer. The server 120 can be implemented as a standalone server or a cluster of multiple servers 120. The user terminal can display sMRI images, DTI images, and fMRI images. Server 120 acquires sMRI, DTI, and fMRI images from user terminal 110. Server 120 uses the feature extraction module to determine a first feature sequence based on the sMRI images, a second feature sequence based on the DTI images, and a third feature sequence based on the fMRI images. It then uses a probability distribution encoder to extract uncertainty from the first feature sequence to obtain a first mean vector and a first variance vector, and performs vector sampling based on the first mean vector and the first variance vector to determine a GM sampling vector. Similarly, it uses a probability distribution encoder to extract uncertainty from the second feature sequence to obtain a second mean vector and a second variance vector, and performs vector sampling based on the second mean vector and the second variance vector to determine a DTI sampling vector. Finally, it uses a prediction module to determine a fusion feature vector based on the GM sampling vector, the DTI sampling vector, and the fMRI sampling vector, and performs brain age prediction based on the fusion feature vector to obtain a predicted brain age.

[0032] The application content will be further explained below with reference to the accompanying drawings and through the description of the embodiments.

[0033] This embodiment provides a multi-parameter MRI brain age prediction method based on uncertainty perception, applying a trained brain age fusion prediction model. For example... Figure 2 As shown, the brain age fusion prediction model includes a feature extraction module, a probability distribution encoder, and a prediction module. The feature extraction module is used to extract features from multi-parameter MRI (i.e., sMRI, DTI, and fMRI images). The probability distribution encoder is used to convert the extracted feature maps into Gaussian distributions to obtain the corresponding feature vectors of sMRI, DTI, and fMRI images. The prediction module fuses the corresponding feature vectors of sMRI, DTI, and fMRI images to obtain a fused feature vector, and performs brain age prediction based on the fused feature vector to obtain the predicted brain age.

[0034] The feature extraction module includes a first feature extractor and a second feature extractor. The first feature extractor extracts features from sMRI images, while the second feature extractor extracts features from DTI and fMRI images. When extracting features from sMRI images, the images are processed into GM parameter maps, and then features are extracted from these maps. Similarly, when extracting features from DTI images, the images are processed into FA and MD parameter maps, and when extracting features from fMRI images, the images are processed into ALFF and ReHo parameter maps. Therefore, the first feature extractor is a single-parameter feature extractor, and the second feature extractor includes two parallel two-parameter feature extractors. Each two-parameter feature extractor includes two parallel feature extractors, a Transformer structure, and a residual module. Both parallel feature extractors are connected to the Transformer structure, which in turn is connected to the residual module. Both parallel feature extractors are used for feature extraction, the Transformer structure is used for feature fusion, and the residual module is used for dimensionality transformation. In addition, the GM feature map extracted by the feature extractor, the DTI feature map extracted by one two-parameter feature extractor, and the fMRI feature map extracted by another two-parameter feature extractor are fused through a Transformer structure to obtain the target feature map.

[0035] Furthermore, such as Figure 2 As shown, the feature extractor in the feature extraction module is the same as the two parallel feature extractors in the dual-parameter feature extractor. The feature extractor includes a cascaded 3D image patch embedding layer, a 3D absolute position encoding module, and several 3D attention modules. Each 3D attention module includes a first 3D normalization layer, a 3D multi-head attention unit, an adder, a second 3D normalization layer, and a feedforward network. The first 3D normalization layer, the 3D multi-head self-attention unit, the adder, the second 3D normalization layer, and the feedforward network are cascaded sequentially. The input to the adder includes the input to the first 3D normalization layer. The 3D multi-head attention unit includes a first matrix multiplier, a normalization layer, a 3D relative position encoding layer, and a second matrix multiplier, connected sequentially. The input to the first matrix multiplier is a query vector and a key vector. The input to the second matrix multiplier includes a value vector and the output of the 3D relative position encoding layer. The query vector, value vector, and key vector are constructed based on the output of the first 3D normalization layer.

[0036] The feature extractor in this embodiment effectively divides the 3D medical image (i.e., various parametric maps) into non-overlapping blocks. By introducing absolute position encoding, the spatial position information of each block is explicitly incorporated into the feature representation. Then, the multi-head self-attention mechanism in the multi-layer Transformer block is used to capture the global dependencies and long-distance relationships between the blocks (i.e., it can automatically capture the most relevant regions in two modal images, and through the cross-attention mechanism, the features of two different modalities are aligned across modalities, linking their deep semantics). Finally, the features are fused and dimensionality reduced through normalization operations and a feedforward network to generate high-quality 3D feature maps (i.e., the feature maps corresponding to each parametric map). This can enrich the feature information included in the corresponding feature maps, thereby enriching the feature information included in the target feature map, and thus improving the accuracy of brain age prediction.

[0037] like Figure 2 As shown, the residual module includes a first branch and a second branch. The first branch includes a cascaded 3D convolutional block, a normalization layer, a ReLU activation function, another 3D convolutional block, a normalization layer, an adder, another ReLU activation function, and a 3D max pooling layer. The second branch includes a 3D convolutional block, and the output of the 3D convolutional block in the second branch is the input of the adder. The inputs of the first branch and the second branch are the same. The residual module in the two-parameter feature extractor is used to perform dimensionality transformation on the fused features output by the Transformer structure in the two-parameter feature extractor, so that its output feature tensor maintains spatial dimensionality consistency with the GM features. This facilitates the subsequent fusion of GM feature maps, DTI feature maps, and fMRI feature maps.

[0038] This application utilizes the Transformer structure to leverage the global perception advantage of the self-attention mechanism, effectively capturing long-distance relationships between image features and achieving efficient fusion of multimodal medical image features. To this end, after extracting feature maps using a feature extractor, this application employs the Transformer structure to establish dependencies between multi-parameter MRI image features, thereby improving the performance of the brain age prediction model.

[0039] like Figure 2As shown, the probability distribution encoder includes a first feedforward network, a mean path, and a variance path. The first feedforward network is connected to both the mean path and the variance path, which run in parallel. The input feature sequence of the first feedforward network and the output data of the mean path are fused through an adder. Both the mean path and the variance path include a multi-head attention module, a concatenation module, and a second feedforward network cascaded in sequence. The mean path also includes an adder located between the concatenation module and the second feedforward network. The input to the adder includes the input feature sequence of the first feedforward network and the output data of the concatenation module. Furthermore, the probability distribution encoder includes a sampling unit for collecting the mean vector output by the mean path and the variance vector output by the variance path to determine the feature vector.

[0040] It should be noted that, since the probability distribution encoder needs to convert the feature maps of sMRI, DTI, and fMRI images into Gaussian distributions, three probability distribution encoders can be deployed in parallel to improve conversion speed, and each of the three encoders is connected to the prediction module. Specifically, the probability distribution encoder for sMRI images is connected to the first feature extractor; the probability distribution encoder for DTI images is connected to the two-parameter feature extractor for DTI images in the second feature extractor; and the probability distribution encoder for fMRI images is connected to the two-parameter feature extractor for fMRI images in the second feature extractor.

[0041] like Figure 3 As shown in the embodiments of this application, the multi-parameter MRI brain age prediction method based on uncertainty perception specifically includes: S10. Acquire sMRI, DTI, and fMRI images of the brain; S20. The feature extraction module determines a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images. S30. Uncertainty extraction is performed on the first feature sequence through a probability distribution encoder to obtain a first mean vector and a first variance vector, and vector sampling is performed based on the first mean vector and the first variance vector to determine the GM sampling vector; S40. Uncertainty extraction is performed on the second feature sequence through a probability distribution encoder to obtain a second mean vector and a second variance vector, and vector sampling is performed based on the second mean vector and the second variance vector to determine the DTI sampling vector. S50. Uncertainty extraction is performed on the third feature sequence through a probability distribution encoder to obtain a third mean vector and a third difference vector, and vector sampling is performed based on the third mean vector and the third difference vector to determine the fMRI sampling vector; S60. The prediction module determines the fusion feature vector based on the GM sampling vector, DTI sampling vector and fMRI sampling vector, and performs brain age prediction based on the fusion feature vector to obtain the predicted brain age.

[0042] This application's embodiments determine a fused feature vector by fusing sMRI, DTI, and fMRI information, and then use this fused feature vector for brain age prediction. This allows for a more comprehensive acquisition of multi-dimensional information on brain structure and function, overcoming the limitation of information provided by single-parameter MRI and improving the accuracy of brain age prediction. Furthermore, this application introduces a probability distribution encoder during the fusion of multi-parameter MRI images. This encoder transforms the multi-parameter MRI feature mapping into a Gaussian distribution to model feature uncertainty, capturing richer potential information from the multi-parameter MRI images. This improves the robustness and predictive performance of the brain age fusion prediction model, thereby further enhancing the accuracy of brain age prediction.

[0043] Specifically, in step S10, the sMRI, DTI, and fMRI images are obtained from cranial MRI images of the same patient, differing only in the parameters used. sMRI images provide clear anatomical information with high spatial resolution, DTI images can be used to analyze topological changes at the microstructural level in the brain's white matter, and fMRI images provide information on brain activity with high temporal resolution. Therefore, multi-parameter MRI images, including sMRI, DTI, and fMRI, can not only display brain anatomical structures but also present multi-dimensional information such as tissue microstructure, functional characteristics, and physiological states, enabling the tracking of brain structural and functional development and aging.

[0044] Furthermore, in step S20, the feature extraction module is used to extract features from sMRI images, DTI images, and fMRI images. Before extracting features from sMRI images, DTI images, and fMRI images, they need to be converted into parametric maps. Then, feature extraction is performed on their respective parametric maps to obtain a first feature sequence, a second feature sequence, and a third feature sequence.

[0045] Specifically, sMRI images can be processed into GM parameter maps, which are parametric maps of gray matter volume or gray matter density. DTI images can be processed into FA parameter maps and MD parameter maps. FA parameter maps are parametric maps of anisotropy fraction, used to quantify the degree of directional diffusion of water molecules in brain white matter fibers, reflecting the integrity of brain white matter fibers. MD parameter maps are parametric maps of average diffusion coefficient, used to measure the overall magnitude of water molecule diffusion in various directions, reflecting the integrity of brain tissue. fMRI images can be processed into ALFF parameter maps and ReHo parameter maps. ALFF parameter maps are parametric maps of low-frequency amplitude, used to measure the intensity of signal fluctuations in spontaneous brain activity in the low-frequency range, reflecting the intensity of neuronal activity. ReHo parameter maps are parametric maps of regional consistency, used to measure the temporal synchronicity of adjacent voxel activity in local brain regions.

[0046] After obtaining the GM parameter map, FA parameter map, MD parameter map, ALFF parameter map and ReHo parameter map, the feature extraction module extracts features from the GM parameter map to determine the first feature sequence, the feature extraction model extracts features from the FA parameter map and MD parameter map to determine the second feature sequence, and the feature extraction module extracts features from the ALFF parameter map and ReHo parameter map to determine the third feature sequence.

[0047] Furthermore, in practical applications, to reduce computational load and better adapt to network input, the GM, FA, MD, ALFF, and ReHo parameter maps can be cropped to remove surrounding non-brain regions. These cropped maps are then resampled to a fixed size. The fixed sizes for the GM, FA, MD, ALFF, and ReHo parameter maps can be the same or different. For example, the fixed size for the GM, FA, and MD parameter maps is 96×128×96, and the fixed size for the ALFF and ReHo parameter maps is 48×64×48. Then, the converted fixed-size GM, FA, MD, ALFF, and ReHo parameter maps undergo maximum-minimum normalization, mapping their grayscale values ​​to between 0 and 1 to reduce their impact on training. Maximum-minimum normalization can be expressed as:

[0048] in, This represents the minimum value. This represents the maximum value.

[0049] In this embodiment, the FA and MD parameter maps both belong to diffusion tensor imaging features, while the ALFF and ReHo parameter maps both belong to functional MRI features. FA and MD reflect the diffusion characteristics of water molecules in white matter fiber bundles from different perspectives. By first extracting and fusing features from the FA and MD parameter maps using a dual-parameter feature extractor, a more robust high-level feature representation of "white matter integrity" can be formed. Similarly, the ALFF and ReHo parameter maps characterize neuronal activity from different dimensions. By first extracting and fusing features from the ALFF and ReHo parameter maps using a dual-parameter feature extractor, a more accurate high-level feature representation of "neuronal activity characteristics" can be formed. Furthermore, fusing parameter maps from the same MRI image first allows the brain age prediction model to more effectively capture the intrinsic, strongly correlated microstructural information of both, reducing computational complexity and improving model interpretability.

[0050] Furthermore, in steps S30-S50, although multi-parameter MRI can increase MRI information, it also introduces more noise and introduces varying degrees of uncertainty in MRI. Simultaneously, noise can make the estimation of points in the point representation space unclear. These uncertainties affect the accuracy of brain age prediction. Therefore, after obtaining the first, second, and third feature sequences, this application does not directly fuse them to determine the fused feature map. Instead, it converts the first, second, and third feature sequences into a Gaussian distribution in the probability space to capture more potential information, thereby overcoming the impact of uncertainty on brain age prediction and further improving the accuracy of brain age prediction.

[0051] Based on this, in steps S30-S50 of this embodiment, a Gaussian distribution is introduced into the multimodal uncertainty modeling of medical images through a probability distribution encoder in the brain age fusion prediction model. Through interaction at the feature level and sequence level, the point representation space is transformed into a Gaussian distribution, estimating the mean vector and variance vector of different MRI sequence features, so that the variance vector represents the measure of feature uncertainty. Specifically, the probability distribution encoder performs uncertainty extraction on the first feature sequence, the second feature sequence, and the third feature sequence respectively, so as to transform the first feature sequence, the second feature sequence, and the third feature sequence into a Gaussian distribution in the probability space, thereby estimating the mean vector and variance vector of the MRI sequence features corresponding to sMRI images, DTI images, and fMRI images respectively. The uncertainty extraction process of the probability distribution encoder for the first feature sequence, the second feature sequence, and the third feature sequence is the same, only the input feature sequence and output data are different. Therefore, this embodiment uses the specific extraction process of uncertainty extraction as an example for explanation.

[0052] For example, the extraction process for uncertainty extraction specifically includes: H10. Divide the input feature sequence into a first input feature sequence and a second input feature sequence; H20. Utilize a multi-head attention mechanism to determine an initial mean vector based on the first input feature sequence, and determine the mean vector based on the initial mean vector and the first input feature sequence; H30. The variance vector is determined based on the second input feature sequence using a multi-head attention mechanism.

[0053] In step H10, the input feature sequence is the input item of the probability distribution encoder, and the first and second input feature sequences are obtained by splitting the input feature sequence. When dividing the input feature sequence into the first and second input feature sequences, an equal-proportional division method can be used, for example, dividing the input feature sequence equally along its dimensional direction. This ensures that the information contained in the first and second input feature sequences is relatively balanced in subsequent processing, laying the foundation for accurately determining the mean vector and variance vector.

[0054] Furthermore, when dividing the input feature sequence into a first input feature sequence and a second input feature sequence, the input feature sequence can first be fed into a feedforward neural layer to complete feature-level interaction to obtain an interactive feature sequence. Then, the interactive feature sequence can be equally divided into the first input feature sequence and the second input feature sequence according to the sequence dimension. For example, if the interactive feature sequence has a sequence length of 288 and a sequence dimension of 32, then the first input feature sequence has a column length of 288 and a sequence dimension of 16, and the second input feature sequence also has a column length of 288 and a sequence dimension of 16.

[0055] In step H20, the mean vector reflects the average feature information of the input feature sequence. This mean vector is obtained by performing attention learning on the first input feature sequence using a multi-head attention mechanism. The multi-head attention mechanism can extract information from different representation subspaces, enhancing the model's ability to capture complex dependencies in the input feature sequence. Specifically, when processing the first input feature sequence, the multi-head attention mechanism maps it to multiple low-dimensional subspaces, each with an independent attention head. These attention heads assign different weights to different elements based on the correlation between elements in the input feature sequence, thereby highlighting important feature information. Then, the outputs of each attention head are concatenated and linearly transformed to obtain the initial mean vector.

[0056] Since the mean vector is correlated with the input feature sequence, after obtaining the initial mean vector, it can be fused with the input feature sequence. This allows the fused mean vector to carry the original information from the input feature sequence, supplementing the initial mean vector. Thus, the mean vector not only contains the important feature information extracted from the first input feature sequence by the multi-head attention mechanism but also retains the original information of the input feature sequence, more accurately reflecting the average features of the input feature sequence. Specifically, when fusing the initial mean vector with the input feature sequence, an adder can be used to add the initial mean vector to the input feature sequence.

[0057] In one embodiment, determining the initial mean vector based on the first input feature sequence using a multi-head attention mechanism specifically includes: The first input feature sequence is divided equally according to the number of attention heads in the multi-head attention mechanism to obtain a number of sub-input feature sequences. Attention is calculated for each sub-input feature sequence to obtain the attention features corresponding to each sub-input feature sequence; The attention features corresponding to all sub-input feature sequences are concatenated to obtain the initial mean vector.

[0058] Specifically, the multi-head attention mechanism includes multiple attention heads, each independently performing attention calculations on the sub-input feature sequences. Therefore, when processing the first input feature sequence using the multi-head attention mechanism, the first input feature sequence can be divided into sub-input feature sequences according to the sequence dimension, corresponding to the number of attention heads. Each equally divided sub-input feature sequence is then used as input data for each attention head, and attention calculations are performed by each attention head. During attention calculation, the sub-input feature sequences are first transformed linearly to obtain a query vector, a key vector, and a value vector. The query vector is used to calculate similarity with the key vector to determine the relevance weights between different elements. Then, by calculating the dot product of the query vector and the key vector, and after scaling and normalization, attention scores are obtained. These attention scores reflect the importance of each element in the sub-input feature sequence. Finally, the attention scores are multiplied by the value vectors to obtain a weighted value vector, which is the attention feature.

[0059] By performing attention calculations on each sub-input feature sequence, attention features corresponding to each sub-input feature sequence can be obtained. These attention features reflect the information of the sub-input feature sequences from different perspectives. Finally, the attention features corresponding to all sub-input feature sequences are concatenated to obtain an initial mean vector. This integrates the information extracted by multiple attention heads, making the initial mean vector contain richer feature information, thus providing a foundation for subsequently determining a more accurate mean vector.

[0060] In step H30, the variance vector is a measure of the uncertainty of the input feature sequence, used to measure the dispersion of the input feature sequence. A larger variance vector indicates a higher degree of dispersion and greater uncertainty in the input feature sequence; conversely, a smaller variance vector indicates a more concentrated input feature sequence and lower uncertainty. The variance vector accurately characterizes the uncertainty of the input feature sequence, providing important quantitative basis for handling the uncertainty of multimodal medical images in subsequent brain age prediction models. Specifically, the variance vector is obtained by processing the second input feature sequence using a multi-head attention mechanism, and the processing procedure is the same as that for determining the initial mean vector, so it will not be repeated here.

[0061] This embodiment of the application, through the aforementioned uncertainty extraction process, transforms the first feature sequence, the second feature sequence, and the third feature sequence into a Gaussian distribution in the probability space, obtaining the mean vector and variance vector of the MRI sequence features corresponding to the sMRI image, DTI image, and fMRI image, respectively. These mean vectors and variance vectors contain more potential information about the features, effectively compensating for the impact of uncertainty introduced by multi-parameter MRI on the accuracy of brain age prediction. This allows the brain age fusion prediction model to analyze and process medical image data more accurately, thereby improving the accuracy and reliability of brain age prediction.

[0062] Furthermore, the mean vector represents the center position of the Gaussian distribution in the probability space, while the variance vector represents the distribution range of each dimension. Using the mean and variance vectors, a feature sequence can be transformed into a Gaussian distribution with the mean and variance vectors as parameters. For example, the Gaussian distribution of the first feature sequence can be represented as: , in, Indicates the first characteristic sequence. i Sampling features of the characteristic distribution of a sequence, The first feature sequence represents the first feature sequence. i A sequence, The first feature sequence represents the first feature sequence. i The mean vector of the sequences The first feature sequence represents the first feature sequence. i The variance vector of each sequence. Represents the identity matrix.

[0063] If we consider the covariance matrix, the mean vector of the first characteristic sequence It can be used as a predictive feature of the GM parameter map, the variance vector of the first feature sequence. A measure of characteristic uncertainty can be derived from... Random sampling yields a sampled feature vector. However, this sampling operation may prevent the brain age fusion prediction model from backpropagating during the training phase.

[0064] Therefore, in this embodiment of the application, the sampling process of the quantity sampling specifically includes: Sample random noise from a standard normal distribution; The mean vector and variance vector are weighted based on the random noise to obtain the sampled feature vector.

[0065] Specifically, random noise follows a standard normal distribution. Sampling a random noise ,Right now Then, the mean vector and variance vector are weighted using random noise, where the sampled feature vector can be represented as: , in, Represents the sampled feature vector. Represents the mean vector. Represents the variance vector. This represents random noise.

[0066] This application's embodiments transform random sampling operations into a deterministic calculation process, which not only allows for the calculation of the mean vector... Sum of variance vectors from The separation of brain age into the model also enables backpropagation during the training phase of the brain age fusion prediction model.

[0067] Furthermore, in step H60, after obtaining the GM sampling vector, DTI sampling vector, and fMRI sampling vector, the GM sampling vector, DTI sampling vector, and fMRI sampling vector are fused to obtain a fused feature vector. Then, the fused feature vector is used to predict brain age. For example, the prediction module includes a fully connected layer. The fused feature vector is input into the fully connected layer, and the fully connected layer performs a nonlinear transformation on the fused feature vector to uncover the complex relationships between features, thereby outputting the predicted brain age.

[0068] In one embodiment, determining the fusion feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector specifically includes: The GM sampling vector is fused with the first mean vector to obtain the first fused feature vector; The DTI sampling vector is fused with the second mean vector to obtain the second fused feature vector; The fMRI sampling vector is fused with the third mean vector to obtain the third fused feature vector; The first fused feature vector, the second fused feature vector, and the third fused feature vector are fused to obtain a fused feature vector.

[0069] Specifically, the GM sampling vector and the first mean vector, the DTI sampling vector and the second mean vector, and the fMRI sampling vector and the third mean vector can all be fused using feature concatenation. This allows the fused feature vector to simultaneously contain information from both the sampling vectors and the average feature information represented by the mean vector. Feature concatenation can be performed along the feature dimension; for example, the GM sampling vector and the first mean vector can be concatenated in the column direction to form a new feature vector, namely the first fused feature vector. Similarly, the DTI sampling vector and the second mean vector, and the fMRI sampling vector and the third mean vector, are also concatenated using feature concatenation to obtain the second and third fused feature vectors, respectively.

[0070] Subsequently, when fusing the first, second, and third fusion feature vectors, a concatenation method can also be used. This layer-by-layer fusion approach can fully integrate information from different modalities of image features, resulting in a final fusion feature vector that contains more and more comprehensive feature information, providing a richer and more accurate basis for subsequent brain age prediction.

[0071] In practical applications, sampling operations increase the training difficulty of the brain age fusion prediction model. The sampling vector may converge to its optimal position under L1 loss function optimization, leading to variance "collapse" and the distribution representation ultimately degenerating into point representations. This causes the brain age fusion prediction model to lose its ability to learn multi-parameter MRI uncertainties. Therefore, during the training process of the brain age fusion prediction model, a regularized loss function is introduced to prevent the uncertainty level of different MRI image feature distributions from falling below a certain threshold. In other words, during the training process of the brain age fusion prediction model, a regularized loss function is introduced into the loss function terms corresponding to the GM feature vector, DTI feature vector, and fMRI feature vector, respectively. The regularized loss function is expressed as: , , in, This represents the regularization loss function. The entropy of a multivariate Gaussian distribution is represented by its value. Represents the mean vector. Represents the variance vector. Representing feature dimension, This indicates a preset threshold.

[0072] Furthermore, the brain age fusion prediction model can be obtained by weighting the loss function terms corresponding to the GM feature vector, DTI feature vector, and fMRI feature vector. In addition to the regularization loss function, the loss function terms corresponding to the GM feature vector, DTI feature vector, and fMRI feature vector can also include the L1 loss function and the InfoNCE contrast loss function.

[0073] In summary, this embodiment provides a multi-parameter MRI brain age prediction method based on uncertainty perception. By fusing sMRI, DTI, and fMRI information to determine a fused feature vector, and then using this fused feature vector for brain age prediction, a more comprehensive multi-dimensional information on brain structure and function can be obtained, overcoming the limitation of information provided by single-parameter MRI and improving the accuracy of brain age prediction. Furthermore, this application introduces a probability distribution encoder during the fusion of multi-parameter MRI images. This encoder transforms the multi-parameter MRI feature mapping into a Gaussian distribution to model feature uncertainty, capturing richer potential information from the multi-parameter MRI images, improving the robustness and predictive performance of the brain age fusion prediction model, thereby further enhancing the accuracy of brain age prediction.

[0074] Based on the aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, this embodiment provides a multi-parameter MRI brain age prediction device based on uncertainty perception, which applies a trained brain age fusion prediction model. The brain age fusion prediction model includes a feature extraction module, a probability distribution encoder, and a prediction module; for example... Figure 4 As shown, the device includes: Acquisition module 100 is used to acquire multi-parameter MRI images of the brain; The control module 200 is configured to: determine a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images using the feature extraction module; extract uncertainty from the first feature sequence using a probability distribution encoder to obtain a first mean vector and a first variance vector, and perform vector sampling based on the first mean vector and the first variance vector to determine a GM sampling vector; extract uncertainty from the second feature sequence using a probability distribution encoder to obtain a second mean vector and a second variance vector, and perform vector sampling based on the second mean vector and the second variance vector to determine a DTI sampling vector; extract uncertainty from the third feature sequence using a probability distribution encoder to obtain a third mean vector and a third variance vector, and perform vector sampling based on the third mean vector and the third variance vector to determine an fMRI sampling vector; and determine a fusion feature vector using a prediction module based on the GM sampling vector, the DTI sampling vector, and the fMRI sampling vector, and perform brain age prediction based on the fusion feature vector to obtain a predicted brain age.

[0075] Based on the above-described multi-parameter MRI brain age prediction method based on uncertainty perception, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the multi-parameter MRI brain age prediction method based on uncertainty perception as described in the above embodiment.

[0076] Based on the aforementioned multi-parameter MRI brain age prediction method based on uncertainty perception, this application also provides a terminal device, such as... Figure 5 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0077] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0078] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0079] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0080] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-parameter MRI brain age prediction method based on uncertainty perception, characterized in that, The method involves applying a trained brain age fusion prediction model, which includes a feature extraction module, a probability distribution encoder, and a prediction module; the method includes: Acquire sMRI, DTI, and fMRI images of the brain; The feature extraction module determines a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images. Uncertainty extraction is performed on the first feature sequence by a probability distribution encoder to obtain a first mean vector and a first variance vector, and vector sampling is performed based on the first mean vector and the first variance vector to determine the GM sampling vector; Uncertainty extraction is performed on the second feature sequence by a probability distribution encoder to obtain a second mean vector and a second variance vector, and vector sampling is performed based on the second mean vector and the second variance vector to determine the DTI sampling vector; Uncertainty extraction is performed on the third feature sequence using a probability distribution encoder to obtain a third mean vector and a third difference vector, and vector sampling is performed based on the third mean vector and the third difference vector to determine the fMRI sampling vector; The prediction module determines a fusion feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector, and then performs brain age prediction based on the fusion feature vector to obtain the predicted brain age.

2. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 1, characterized in that, The uncertainty extraction process specifically includes: The input feature sequence is divided into a first input feature sequence and a second input feature sequence; An initial mean vector is determined based on a first input feature sequence using a multi-head attention mechanism, and a mean vector is determined based on the initial mean vector and the first input feature sequence. The variance vector is determined based on the second input feature sequence using a multi-head attention mechanism.

3. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 2, characterized in that, The step of determining the initial mean vector based on the first input feature sequence using a multi-head attention mechanism specifically includes: The first input feature sequence is divided equally according to the number of attention heads in the multi-head attention mechanism to obtain a number of sub-input feature sequences. Attention is calculated for each sub-input feature sequence to obtain the attention features corresponding to each sub-input feature sequence; The attention features corresponding to all sub-input feature sequences are concatenated to obtain the initial mean vector.

4. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 2, characterized in that, The specific steps of dividing the input feature sequence into a first input feature sequence and a second input feature sequence are as follows: The input feature sequence is divided into a first input feature sequence and a second input feature sequence according to the dimensional direction.

5. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 1, characterized in that, The sampling process for the vector sampling specifically includes: Sample random noise from a standard normal distribution; The mean vector and variance vector are weighted based on the random noise to obtain the sampled feature vector.

6. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 1, characterized in that, The determination of the fusion feature vector based on the GM sampling vector, DTI sampling vector, and fMRI sampling vector specifically includes: The GM sampling vector is fused with the first mean vector to obtain the first fused feature vector; The DTI sampling vector is fused with the second mean vector to obtain the second fused feature vector; The fMRI sampling vector is fused with the third mean vector to obtain the third fused feature vector; The first fused feature vector, the second fused feature vector, and the third fused feature vector are fused to obtain a fused feature vector.

7. The multi-parameter MRI brain age prediction method based on uncertainty perception according to claim 1, characterized in that, During the training process of the brain age fusion prediction model, a regularized loss function is introduced into the loss function terms corresponding to the GM sampling vector, DTI sampling vector, and fMRI sampling vector, where the regularized loss function is expressed as: ， ， in, This represents the regularization loss function. The entropy of a multivariate Gaussian distribution is represented by its value. Represents the mean vector. Represents the variance vector. Representing feature dimension, This indicates a preset threshold.

8. A multi-parameter MRI brain age prediction device based on uncertainty perception, characterized in that, The device employs a trained brain age fusion prediction model, which includes a feature extraction module, a probability distribution encoder, and a prediction module; the device includes: The acquisition module is used to acquire multi-parameter MRI images of the brain; The control module is configured to: determine a first feature sequence based on sMRI images, a second feature sequence based on DTI images, and a third feature sequence based on fMRI images using the feature extraction module; extract uncertainty from the first feature sequence using a probability distribution encoder to obtain a first mean vector and a first variance vector, and perform vector sampling based on the first mean vector and the first variance vector to determine a GM sampling vector; extract uncertainty from the second feature sequence using a probability distribution encoder to obtain a second mean vector and a second variance vector, and perform vector sampling based on the second mean vector and the second variance vector to determine a DTI sampling vector; extract uncertainty from the third feature sequence using a probability distribution encoder to obtain a third mean vector and a third variance vector, and perform vector sampling based on the third mean vector and the third variance vector to determine an fMRI sampling vector; and determine a fused feature vector using a prediction module based on the GM sampling vector, the DTI sampling vector, and the fMRI sampling vector, and perform brain age prediction based on the fused feature vector to obtain a predicted brain age.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the multi-parameter MRI brain age prediction method based on uncertainty perception as described in any one of claims 1-7.

10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the multi-parameter MRI brain age prediction method based on uncertainty perception as described in any one of claims 1-7.

Images