Method and system for predicting baby brain function connection based on morphological characteristics and diffusion model, storage medium and electronic equipment

Abstract

The invention discloses a method and system for predicting infant brain function connection based on morphological characteristics and a diffusion model, a storage medium and electronic equipment. The method in the formula comprises the following steps: extracting cortex morphological characteristics in an infant structure magnetic resonance image (sMRI), constructing a morphological similarity network (MSN), and realizing infant brain function full-connection layer prediction by utilizing a diffusion model in combination with classifier irrelevant guidance and a cross-modal attention mechanism. According to the method, individual development features are stably captured by using a longitudinal information extraction module, accurate mapping from morphological features to function connection is realized in combination with a classifier irrelevant guidance diffusion model and a morphological guidance attention mechanism, the prediction precision and stability are remarkably improved, and the prediction efficiency is improved. The problem that infant functional magnetic resonance imaging (fMRI) data is scarce or low in quality can be effectively solved, functional networks in different development stages are accurately reconstructed, and the accuracy of infant early brain development monitoring is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis and relates to methods and systems for predicting functional connectivity of infant brain, specifically to a method and system, storage medium and electronic device for predicting functional connectivity of infant brain based on morphological features and diffusion models. Background Technology

[0002] Understanding the patterns of early human brain development is crucial for neuroscience research. Functional connectivity (FC), a key indicator measuring the functional collaboration between different brain regions, plays a vital role in studying cognitive and behavioral development, as well as monitoring early changes in neurodevelopmental disorders such as autism spectrum disorder and ADHD. In particular, longitudinal changes in brain functional connectivity during infancy directly reflect the formation and reorganization of brain functional networks in early development, serving as an important tool for revealing early brain plasticity and critical developmental periods.

[0003] However, acquiring resting-state functional MRI (rs-fMRI) data for infants faces numerous challenges, primarily including artifacts caused by vigorous subject movement, limited scan time, and low infant cooperation. These issues result in a severe scarcity of longitudinal functional connectivity data for infants, significantly restricting the ability to systematically study the early developmental patterns of infant brain function. Existing research attempts to generate missing structural connectivity (SC) data using deep learning models, but due to the dramatic reorganization of infant brain function, these methods perform poorly in predicting infant functional connectivity. Furthermore, while existing studies based on triplet contrastive learning frameworks can predict brain functional connectivity to some extent using longitudinal information, these methods lack stable and reliable age-specific references to pinpoint developmental stages, thus limiting the accuracy and biological plausibility of the predictions.

[0004] In recent years, some studies have proposed using Morphometric Similarity Networks (MSNs) for indirect prediction of functional connectivity. These networks extract various morphological features from structural MRI (sMRI) data, such as cortical thickness, surface area, gray matter volume, myelin content, and cortical curvature, and construct morphological feature similarity matrices between brain regions. Evidence suggests a significant correlation between these structural-morphological networks and functional networks, indicating that structural features can serve as an effective substitute for functional connectivity. However, current structure-to-function prediction methods are mostly limited to simple statistical association analysis, lacking robust deep learning and generative model frameworks, and thus failing to effectively capture the dynamically changing structure-function relationships during brain development.

[0005] Recently, diffusion models have demonstrated significant advantages in medical image generation tasks. Compared to traditional generative adversarial networks (GANs), diffusion models offer stable training, rich reconstruction details, and effectively address the common problems of low signal-to-noise ratio and missing data in medical images. Therefore, applying diffusion models to the prediction of longitudinal functional connectivity in infants holds promise for effectively solving the problems of data scarcity and low quality.

[0006] In summary, existing technologies have significant shortcomings in predicting longitudinal infant brain functional connectivity, especially in handling the complex relationship between structural features and functional connectivity, integrating age-dependent constraints, and processing low-quality data, which urgently require improvement. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, storage medium, and electronic device for predicting infant brain functional connectivity based on morphological features and diffusion models. It utilizes a longitudinal information extraction module to stably capture individual developmental features and combines a classifier-independent guided diffusion model with a morphology-guided attention mechanism to achieve accurate mapping from morphological features to functional connectivity, thereby enabling high-precision prediction of infant longitudinal functional connectivity.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A method for predicting infant brain functional connectivity based on morphological features and diffusion models includes the following steps: Step 1: Acquire and preprocess the infant's structural magnetic resonance imaging data, and reconstruct the cortical region from the preprocessed data to generate region segmentation labels; Step 2: Extract multidimensional morphological features based on the region segmentation labels obtained in Step 1, calculate the Pearson correlation coefficient between each cortical region, and construct a morphological similarity network; Step 3: Based on the construction of the whole-brain morphological similarity network in Step 2, a network structure is constructed including a longitudinal information extraction module, a classifier-independent guided diffusion model, and a morphology-guided attention mechanism. Step 4: Use the longitudinal information extraction module to extract individual-specific structural features at different time points from the morphological similarity network data obtained in Step 2; Step 5: Using the individual-specific structural features obtained in Step 4 as input, the fully connected layer map is denoised and predicted based on the classifier-independent diffusion model in the network structure of Step 3. In the process of model inference, a morphology-guided attention mechanism is integrated to generate a predicted fully connected layer map of infant brain function according to the specified age.

[0009] The present invention also has the following technical features: Preferably, the multidimensional morphological features mentioned in step two include: cortical thickness, surface area, gray matter volume, myelin content, and curvature.

[0010] Preferably, the longitudinal information extraction module described in step three constructs a shared encoder E to receive triplet inputs. The output corresponds to the latent representation. ; The following loss function is used during training to reconcile the triplet relationships: In the formula: , ; In the formula, The morphological similarity network features of the same subject i at different times. E represents the morphological similarity network features of other subjects j; E is the shared encoder. This represents the correlation coefficient, with a value ranging from -1 to 1.

[0011] Preferably, the classifier-independent guided diffusion model described in step three includes two stages: forward noise addition and reverse denoising; the forward noise addition process is defined as: , In the formula, This indicates the diffusion process at time step Noisy data, This represents the data obtained from the backdiffusion step; For the first The noise intensity of the step, This is the attenuation coefficient of the data from the previous step. The covariance of the injected noise; An identity matrix consistent with the data dimensions; The reverse denoising process is defined as: , In the formula, For neural network model parameters, The standard deviation of the back-diffusion noise. The mean is Covariance is The Gaussian distribution.

[0012] The reverse mean term is defined as: ; In the formula, The signal preservation coefficient, This represents the cumulative signal retention amount. For the model to input the current step data Time step and condition information The predicted noise is used to estimate the direction of data denoising and generate .

[0013] Furthermore, during the training process, the classifier-independent guidance diffusion model achieves classifier-independent guidance by randomly discarding the age condition. Combined with the individual-specific structural features and target age information, it optimizes the noise prediction network to ensure that the model can balance the accuracy of structural priors and functional connectivity generation. Introducing guidance coefficients during sampling The noise during final sampling is defined as: ; The sampling step is defined as follows: ; In the formula, For the diffusion process at time step Noisy data, Data generated for backdiffusion; Indicates the current step number and conditions Below, by parameters Noise prediction of neural network output, This represents unconditional noise prediction, where The condition is empty; These are conditional guiding coefficients, used to adjust the interpolation between conditional and unconditional predictions; The effective noise obtained by interpolation; and These are the reverse diffusion mean terms for conditional and unconditional conditions, respectively; Back diffusion in steps The corresponding noise intensity, It is Gaussian random noise; The model's loss function is: ; In the formula, The training objective is to address the noise prediction error. As expected, It is the square of the L2 norm.

[0014] Preferably, the morphology-guided attention mechanism described in step three is implemented in the U-Net structure of the classifier-agnostic guided diffusion model, targeting the fused feature map. The feature matrix, which has unique identifiers for each individual, obtained by the encoder through the longitudinal information extraction module, is used for cross-modal attention using the following formula: In the formula, The node feature matrix, The feature vector is obtained by the encoder from the MSN (Morphometric Similarity Network); A mapping for self-attention compression of functional connectivity features; A learnable projection matrix for query, key, and value. The dimension of the key vector. For attention normalization operation, This is the output result after fusion.

[0015] Preferably, the process of extracting individual-specific structural features at different time points from the morphological similarity network data using the longitudinal information extraction module in step four includes: 1) Input triples The latent vectors are obtained by passing each vector through the same encoder E: ; 2) For morphological similarity network data at any time t Stable individual feature vectors can be obtained through an encoder: ; 3) Select the individual feature vector at the most recent time point as the feature matrix with unique individual identification for the subject. And together with age conditions, it provides a noise prediction network.

[0016] This invention also protects a system for predicting infant brain functional connectivity based on morphological features and diffusion models using the method described above, comprising a data acquisition and preprocessing module, a morphological feature extraction module, a longitudinal information extraction module, and a diffusion model module; The data acquisition and preprocessing module acquires structural magnetic resonance imaging data of the test infants and performs intensity non-uniformity correction, removal of cranial and non-brain tissues, white matter / gray matter / cerebrospinal fluid segmentation, and reconstruction of the left and right hemispheres and cortical regions on the acquired structural magnetic resonance imaging data. The morphological feature extraction module extracts morphological features of the cerebral cortex from the data preprocessed by the data acquisition and preprocessing module, including: cortical thickness, surface area, gray matter volume, myelin content and curvature, and performs Z-score normalization, calculates the Pearson correlation coefficient between each cortical region, and generates a whole-brain morphological similarity network. The longitudinal information extraction module utilizes a triplet network structure to encode the morphological similarity network features of the same subject at different time points with other subjects, thereby achieving stable extraction of individual-specific features. The diffusion model module achieves classifier-independent guidance by randomly discarding age conditions. It combines the individual-specific structural representation and target age information to optimize the noise prediction network, ensuring that the model can balance the accuracy of structural priors and functional connectivity generation. It also dynamically integrates structural features and age constraints, iteratively denoising to generate a brain functional connectivity map that conforms to biological laws.

[0017] The present invention also protects a storage medium storing a program that implements the method described above for predicting infant brain functional connectivity based on morphological features and diffusion models.

[0018] The present invention also protects an electronic device comprising at least one central processing unit, a graphics processing unit, an input / output device, a network communication module, and a storage unit electrically coupled to the central processing unit; the storage unit stores program code that can be read and executed by the central processing unit to implement the method described above for predicting infant brain functional connectivity based on morphological features and diffusion models.

[0019] Compared with the prior art, the present invention has the following technical effects: This invention extracts rich morphological features and constructs a morphological similarity network by processing and analyzing infant sMRI data. Combined with a longitudinal information extraction module, a classifier-independent guided diffusion model, and a morphology-guided attention mechanism, it achieves a precise mapping from morphological features to functional connections, significantly improving prediction accuracy and stability. It accurately reconstructs functional networks at different developmental stages, enhances the accuracy of early infant brain development monitoring, and provides an effective tool for infant brain development research and early diagnosis of neurological diseases. Attached Figure Description

[0020] Figure 1 A schematic diagram illustrating the construction process of the invention's method for predicting infant brain functional connectivity based on morphological features and diffusion models; Figure 2 This is a framework diagram of the noise prediction network used in the diffusion model; Figure 3 A schematic diagram illustrating the prediction results of different infant brain functional connectivity prediction methods; Figure 4 A schematic diagram of the structure of an electronic device according to Embodiment 4 of the present invention is shown. Detailed Implementation

[0021] The following detailed explanation of the specific content of the present invention is provided in conjunction with embodiments. These descriptions are intended to explain the present invention and not to limit it.

[0022] Example 1 This embodiment provides a method for predicting infant brain functional connectivity based on morphological features and diffusion models. Figure 1 This is a flowchart illustrating a method for predicting functional connectivity in the infant brain based on morphological features and diffusion models, as shown below. Figure 1 As shown, it includes the following steps: Step 1: Acquire and preprocess the infant's structural magnetic resonance imaging (sMRI) data. Specifically, the infants were placed on a 3T MRI scanner (equipped with a 32-channel head coil) under mild sedation (e.g., oral tranquilizers). T1-weighted images (TR / TE / TI = 2400 / 2.24 / 1600 ms, flip angle 8°, 0.8 mm³ isotropic) and T2-weighted images (TR / TE = 3200 / 564 ms, 0.8 mm³ isotropic) were acquired sequentially. The T2 images were then linearly registered to the T1 space to ensure anatomical consistency. After acquisition, the N3 method was used to eliminate intensity non-uniformity. Then, deep learning or threshold-based segmentation algorithms were used to remove cranial and non-brain tissues. Next, a deep learning model was used to segment gray matter, white matter, and cerebrospinal fluid, and the left and right hemispheres were automatically separated to obtain pure brain tissue images.

[0023] The preprocessed images are input into the FreeSurfer pipeline, and the inner and outer surfaces of the cortex are reconstructed based on the Desikan–Killiany template, automatically generating 68 anatomical region segmentation labels. Then, neuroimaging experts manually correct these labels in ITK-SNAP or 3D Slicer to obtain the true cortical region segmentation labels for the training set.

[0024] Step 2: Extract multi-dimensional morphological features based on region segmentation labels and construct a morphological similarity network; Specifically, for each cortical anatomical region obtained in Step 1, five morphological features are calculated: cortical thickness, surface area, gray matter volume, myelin content, and surface curvature. These five features are then Z-score standardized to eliminate the influence of dimensions and individual differences. The standardized five-dimensional features of each region are then combined into a feature vector. For any two anatomical regions a and b, the Pearson correlation coefficient S_a,b of their feature vectors is calculated, and all S_a,b are organized into an N×N original morphological similarity matrix. To highlight significant structural associations, thresholding or sparsification can be applied to this matrix to remove low-correlation edges, ultimately obtaining a morphological similarity network (MSN) for subsequent longitudinal information extraction and classifier-independent guidance of the diffusion model prediction.

[0025] Step 3: Construct the network structure for the vertical information extraction module, the classifier-independent guided diffusion model, and the morphology-guided attention mechanism, as shown in the attached diagram. Figure 2 As shown; Specifically, this step divides the entire prediction network into three main sub-modules: the Longitudinal Information Extraction (LIEB) module, the Classifier-Independent Guided Diffusion (CFG–Diffusion) model, and the Morphology-Guided Attention (MGA) mechanism, designed as follows: The longitudinal information extraction module (LIEB) constructs a shared encoder E to receive triplet inputs. In the formula, the first two are the MSNs of the same subject i at different times, and the latter is the MSN of other subjects j, outputting the corresponding latent representation. .

[0026] The following loss function is used during training to reconcile the triplet relationships: In the formula: , ; In the formula, The morphological similarity network features of the same subject i at different times. E represents the morphological similarity network features of other subjects j; E is the shared encoder. This represents the correlation coefficient, with a value ranging from -1 to 1.

[0027] This design can simultaneously reduce the representational distance of the same subject at different time points and widen the differences with other individuals, ensuring that the encoder extracts stable individual-specific structural features.

[0028] For classifier-independent guided diffusion (CFG–Diffusion) models, the forward noise addition process is defined as: , In the formula, This indicates the diffusion process at time step Noisy data, This represents the data obtained from the backdiffusion step; For the first The noise intensity of the step, This is the attenuation coefficient of the data from the previous step. The covariance of the injected noise; An identity matrix consistent with the data dimensions; The reverse denoising process is defined as: , In the formula, For neural network model parameters, The standard deviation of the back-diffusion noise. The mean is Covariance is The Gaussian distribution.

[0029] The reverse mean term is defined as: ; In the formula, The signal preservation coefficient, This represents the cumulative signal retention amount. For the model to input the current step data Time step and condition information The predicted noise is used to estimate the direction of data denoising and generate ; To achieve classifier-independent guidance, the condition c (MSN + age) needs to be randomly dropped during training, and the conditional model needs to be trained jointly. With unconditional models A guidance coefficient is introduced during sampling. A guidance coefficient is introduced during the final sampling. The noise during final sampling is defined as: ; The sampling step is defined as follows: ; In the formula, For the diffusion process at time step Noisy data, Data generated for backdiffusion; Indicates the current step number and conditions Below, by parameters Noise prediction of neural network output, This represents unconditional noise prediction, where The condition is empty; These are conditional guiding coefficients, used to adjust the interpolation between conditional and unconditional predictions; The effective noise obtained by interpolation; and These are the reverse diffusion mean terms for conditional and unconditional conditions, respectively; Back diffusion in steps The corresponding noise intensity, It is Gaussian random noise; The model's loss function is: ; In the formula, The training objective is to address the noise prediction error. As expected, It is the square of the L2 norm.

[0030] This module uses the U-Net structure to predict noise. Furthermore, by employing the aforementioned guidance strategy, MSN and target age are precisely integrated into the reverse denoising process to achieve high-fidelity FC map generation.

[0031] Morphological Guided Attention (MGA) in the upsampling branch of U-Net targets the fused feature maps. Cross-modal attention is performed using the following formula, based on the individual-specific structural features F obtained by the encoder via LIEB: , In the formula, The node feature matrix, The feature vector is obtained by the encoder from the MSN (Morphometric Similarity Network); A mapping for self-attention compression of functional connectivity features; A learnable projection matrix for query, key, and value. The dimension of the key vector. For attention normalization operation, This is the output result after fusion. First, the functional connectivity features are self-attentionally compressed, and then combined with structural priors and age conditions, so that the network dynamically adjusts the structure-function fusion weights of each brain region in each denoising iteration, thereby enhancing the biological rationality of the prediction results.

[0032] Step 4: Extract individual-specific structural features at different time points from the morphological similarity network feature data using the longitudinal information extraction module (LIEB). Specifically, in this step, the multi-time-point morphological similarity network features (MSN) are input into the shared encoder E through the longitudinal information extraction module (LIEB), and stable individual-specific vectors are extracted during training using triplet loss constraints. The process is as follows: Input triples The latent vectors are obtained by passing each vector through the same encoder E: , After training, the MSN at any time t is... Stable individual feature vectors can be obtained through encoders. .

[0033] In the subsequent diffusion model, the individual feature vector at the most recent time point will be selected as the ID feature of the subject. And together with age conditions, it provides a noise prediction network.

[0034] Step 5: Implement denoising prediction of functional connectivity graphs based on a diffusion model guided by classifier independence; Specifically, the individual feature vectors obtained in step four and the target developmental age combination are first used as the conditional inputs to the diffusion model. Then, through iterative forward noise addition and backward denoising, a high-quality functional connectivity matrix is ​​finally recovered. The implementation process is as follows: Step 1, Condition Construction: The individual feature vectors output by the longitudinal information extraction module... With the target developmental age Each feature vector is mapped to the same dimension through a fully connected embedding layer and then summed to obtain the joint conditional vector. .

[0035] Step 2, forward noise addition, to the actual function connection matrix According to the preset noise plan Perform T-step forward diffusion: Iterate to As pure noise input.

[0036] Step 3, inverse denoising independent of the classifier, uses the U-Net noise prediction network. And combined with unconditional networks Implementation Guidelines: , In the formula This serves as a guideline for strength, used to balance sample quality and conditional consistency.

[0037] During the reverse sampling process, for each time step First, calculate the conditional mean; ; Then according to Sampling was performed to gradually reconstruct the image. ; Step 4: During the training phase, the following formula is used as the optimization objective for the noise prediction network: , In the formula ; Step 5, Output and Post-processing; when iterating backwards to... At that time, the model output This is the predictive functional connectivity matrix.

[0038] Through the above steps, the system accurately recovers the infant brain functional connectivity map during diffusion denoising, relying on structural morphology priors and age guidance, without requiring a large amount of high-quality fMRI data.

[0039] On a pre-split test set, the following metrics are used to measure prediction accuracy: Mean Absolute Error (MAE) Pearson correlation coefficient (r) To illustrate the accuracy of the functional connectivity prediction method in this embodiment, this embodiment compares the method of this invention with specific implementation data of existing deep learning-based functional connectivity prediction methods, including MLP, MGCN-GAN, and CITN. Quantitative results are shown in Table 1, and qualitative results are shown in... Figure 3 As shown.

[0040] Table 1. Prediction results of MLP, MGCN-GAN, CITN, and MAD-Net Table 1 summarizes the performance metrics of four different methods in the functional connectivity prediction task. It can be seen that the method proposed in this paper significantly outperforms those proposed in other literature. Specifically, our method outperforms other methods in both the MAE and Pearson correlation coefficient metrics. This is attributed to our method's reliance on structural morphology priors and the introduction of age-guided information, thereby improving prediction quality.

[0041] Figure 3This diagram illustrates the results of two individuals using four different methods in a functional connectivity prediction task in Embodiment 1 of the present invention; from Figure 3 The observations show that the method of this invention achieves state-of-the-art performance in both MAE and Pearson correlation coefficients, and it also generates connectivity patterns that are very similar to the fundamental facts. Qualitative results further validate the effectiveness of the proposed functional connectivity prediction method.

[0042] The above results verify that the method of the present invention can accurately restore the functional connectivity of the infant brain by relying on structural morphology priors and age guidance through diffusion models, without requiring a large amount of high-quality fMRI data. This approach provides strong technical support for early monitoring of infant brain development and early diagnosis of neurodevelopmental disorders.

[0043] Example 2 To address the aforementioned technical problems in the prior art, this embodiment provides a system for predicting infant brain functional connectivity based on morphological features and diffusion models, including a data acquisition and preprocessing module, a morphological feature extraction module, a longitudinal information extraction module, and a diffusion model module; The data acquisition and preprocessing module acquires structural magnetic resonance imaging data of the infants and performs intensity non-uniformity correction, removal of cranial and non-brain tissues, white matter / gray matter / cerebrospinal fluid segmentation, and reconstruction of the left and right hemispheres and cortical regions on the acquired structural magnetic resonance imaging data. The morphological feature extraction module extracts morphological features of the cerebral cortex from the data preprocessed by the data acquisition and preprocessing module, including: cortical thickness, surface area, gray matter volume, myelin content and curvature, and performs Z-score normalization, calculates the Pearson correlation coefficient between each cortical region, and generates a whole-brain morphological similarity network. The longitudinal information extraction module utilizes a triplet network structure to encode the morphological similarity network features of the same subject at different time points with other subjects, thereby achieving stable extraction of individual-specific features. The diffusion model module achieves classifier-independent guidance by randomly discarding age conditions. It combines the individual-specific structural representation and target age information to optimize the noise prediction network, ensuring that the model can balance the accuracy of structural priors and functional connectivity generation. It also dynamically integrates structural features and age constraints, iteratively denoising to generate brain functional connectivity maps that conform to biological laws.

[0044] The system for predicting infant brain functional connectivity based on morphological features and diffusion models in this embodiment acquires infant structural magnetic resonance imaging (sMRI) data through a data acquisition and preprocessing module, extracts cortical morphological features from infant sMRI images through a morphological feature extraction module, constructs a morphological similarity network (MSN), and then uses a diffusion model combined with classifier-independent guidance and a longitudinal information extraction module to predict the fully connected layers of infant brain function. This achieves a precise mapping from morphological features to functional connectivity, significantly improving prediction accuracy and stability. It can effectively address the problem of scarce or low-quality infant functional magnetic resonance imaging (fMRI) data, accurately reconstruct functional networks at different developmental stages, and improve the accuracy of early infant brain development monitoring.

[0045] Example 3 This embodiment provides a computer-readable storage medium on which a program for implementing a method for predicting infant brain functional connectivity based on morphological features and diffusion models is stored. When the program is executed by a computer, it performs the following steps: Image preprocessing: Denoising, enhancing, and normalizing the images to improve the input quality of the prediction model.

[0046] Model loading and initialization: Load the infant brain functional connectivity model based on morphological features and diffusion model, and initialize the model parameters.

[0047] Inverse denoising sampling: Input the preprocessed image data, sample from Gaussian space, use a diffusion model to denoise and complete the sampling, and generate prediction results.

[0048] Results storage and display: The prediction results are saved as image files or data files for subsequent diagnosis or analysis, and then visualized on the user interface.

[0049] The storage medium can be a hard disk, solid-state drive, optical disk, flash drive, or cloud storage device, capable of storing the above program in electronic form for execution.

[0050] Example 4 This embodiment provides an electronic device, comprising: at least one central processing unit (CPU) and a graphics processing unit (GPU); a storage unit electrically coupled to the CPU; and input / output devices and a network communication module. The storage unit stores program code that can be read and executed by the CPU for performing the processing methods related to this invention.

[0051] Figure 4 A schematic diagram of the electronic device according to Embodiment 4 of the present invention is shown. (Refer to...) Figure 4 The electronic device of the present invention includes: Computational Units: The computational units include a Central Processing Unit (CPU) and a Graphics Processing Unit (GPU). The CPU possesses multi-core parallel processing capabilities, used for data preprocessing, task coordination and scheduling, and supports efficient thread management to meet real-time requirements. The GPU significantly improves the efficiency of deep learning-related computations through hardware acceleration, supporting large-scale matrix operations and complex model network inference capabilities. Its architecture should be optimized for parallel computing performance and adapted to the needs of deep learning frameworks and algorithms. Working together, these two components can efficiently handle the large amounts of data computation and real-time result output involved in medical image analysis, thereby meeting the processing requirements of high performance and low latency.

[0052] Storage Units: Storage units include Random Access Memory (RAM) and non-volatile memory (such as solid-state drives). RAM is used to store temporary data generated during operation and intermediate program state data. Its capacity and access speed are crucial to system performance and should support high-concurrency access to meet the needs of multi-threaded or multi-tasking processing, thereby ensuring the smoothness and real-time performance of the computation process. Non-volatile memory is used to store program files, model parameters, and long-term retained data. Solid-state drives (SSDs) or other non-volatile storage media can be used to provide fast data write and read capabilities. Its capacity design must support the storage and loading of massive amounts of high-resolution format files of medical images. In addition, storage units can adopt a tiered storage architecture, combining cache and main memory to optimize data access; equipped with a data backup module, using mirrored storage or RAID (Redundant Array of Independent Disks) to improve data reliability; and support expansion of storage capacity through external interfaces (such as USB, SATA, or PCIe) to meet future data growth needs. Through the above design, storage units can efficiently support the data processing needs of electronic devices, ensuring the stability and efficiency of the system in high-performance computing and medical image processing.

[0053] Input Devices: Input devices support the acquisition of medical images, including but not limited to CT, MRI, and other medical imaging equipment, enabling the acquisition of high-resolution image data. Input devices must be compatible with standard medical image formats (such as DICOM) to ensure consistency in data exchange and processing between devices, and support multiple data input interfaces to accommodate the connection requirements of different acquisition devices. Through input devices, medical image data can be efficiently acquired and transmitted, providing a foundation for subsequent processing and analysis.

[0054] Output devices include high-resolution displays for intuitively displaying functional connectivity prediction results, ensuring clear visibility of details. Additionally, output devices may include interactive visualization devices (such as those supporting 3D displays) to enable in-depth analysis and interactive operation of functional connectivity prediction results. These devices support multiple display modes to meet the needs of clinical diagnosis or research analysis, providing users with more comprehensive and intuitive image visualizations.

[0055] Network communication module: including a high-speed Ethernet adapter and an optional 5G communication module, for remote data transmission and edge computing scenarios.

[0056] The configuration of the aforementioned electronic device enables efficient data processing and real-time interactive analysis of functional connectivity prediction results, as described in this invention. The device structure described in this embodiment is merely an exemplary configuration; specific implementation schemes can be adjusted according to requirements.

[0057] While the embodiments disclosed in this invention are as described above, the content is merely for the purpose of facilitating understanding of the invention and is not intended to limit the invention. Any person skilled in the art to which this invention pertains may make any modifications and changes in form and detail of the implementation without departing from the spirit and scope disclosed herein; however, the scope of protection of this invention shall still be determined by the scope defined in the appended claims.

Claims

1. A method for predicting infant brain functional connectivity based on morphological features and diffusion models, characterized in that, The method comprises the following steps: Step one, obtaining and preprocessing structural magnetic resonance imaging data of infants, reconstructing the preprocessed data in the cortical region, and generating a region segmentation label; Step two, extracting multi-dimensional morphological features according to the region segmentation label obtained in step one, calculating the Pearson correlation coefficient between each cortical region, and constructing a morphological similarity network; Step three, constructing a network structure comprising a longitudinal information extraction module, a classifier-independent guided diffusion model, and a morphologically guided attention mechanism based on the whole brain morphological similarity network constructed in step two; Step four, using the longitudinal information extraction module to obtain individual-specific structural features at different time points from the morphological similarity network data obtained in step two; Step five, taking the individual-specific structural features obtained in step four as input, implementing denoising prediction of the whole connection layer atlas based on the classifier-independent guided diffusion model in the network structure of step three, and fusing the morphologically guided attention mechanism in the model inference process to generate a predicted infant brain function whole connection layer atlas according to a specified age.

2. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 1, characterized in that, The multi-dimensional morphological features in step two include cortical thickness, surface area, gray matter volume, myelin content, and curvature.

3. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 1, characterized in that, The longitudinal information extraction module described in step three builds a shared encoder E, which receives a triple input and outputs a corresponding latent representation ; The following loss function is used to coordinate the triple relationship during training: In the formula: ， ； wherein is a morphological similarity network feature for the same subject i at different time instances, is a morphological similarity network feature for other subjects j; E is a shared encoder; represents a correlation coefficient, and takes a value of -1-1.

4. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 1, characterized in that, The classifier-independent guided diffusion model in step three includes two stages of forward noise addition and reverse denoising; the forward noise addition process is defined as: ， In the formula, This indicates the diffusion process at time step Noisy data, This represents the data obtained from the backdiffusion step; For the first The noise intensity of the step, This is the attenuation coefficient of the data from the previous step. The covariance of the injected noise; An identity matrix consistent with the data dimensions; The reverse denoising process is defined as: ， wherein are neural network model parameters, is the standard deviation of the noise of the back diffusion, denotes a Gaussian distribution with mean and covariance . The reverse mean term is defined as: ； wherein is the signal preservation coefficient, is the accumulated signal preservation amount; is the model prediction of the noise at the current step data , time step and conditional information , used to estimate the data denoising direction and generate .

5. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 4, characterized in that, In the training process of the classifier-independent guided diffusion model, the classifier-independent guidance is realized by randomly discarding the age condition, the individual-specific structural features and the target age information are combined to optimize the noise prediction network, and the accuracy of the model in balancing structural prior and functional connection generation is ensured; Introducing a guide coefficient at sampling time The noise at final sampling time is defined as ； The sampling step is defined as: ； wherein is the data at time step with noise, is the data generated by the reverse diffusion; denotes the noise prediction output by the neural network with parameters at the current step number and condition ; denotes the unconditional noise prediction, where the condition is empty; is the condition guided coefficient used to adjust the interpolation between the conditional and unconditional predictions; is the effective noise obtained by interpolation; and are the conditional and unconditional reverse diffusion mean terms, respectively; the reverse diffusion at step number corresponds to the noise intensity is a Gaussian random noise; The loss function of the model is: ； wherein is the training target for the noise prediction error, is the expectation, is the L2 norm square.

6. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 1, characterized in that, The modality-guided attention mechanism described in step three is implemented in a U-Net structure of the classifier-agnostic guided diffusion model, and the fusion feature map is used as the input of the model The feature matrix with individual unique identification obtained by the longitudinal information extraction module of the encoder is used for cross-modal attention in the following formula: In the formula, is a node feature matrix, is a feature vector obtained by the encoder of the MSN; is a mapping for self-attention compression of the functional connection features; is a learnable projection matrix of the query, key, and value, is a key vector dimension, is an attention normalization operation, is a fused output result.

7. The method of predicting infant brain functional connectivity based on morphological features and diffusion model according to claim 1, characterized in that, The process of extracting individual-specific structural features at different time points from the morphological similarity network data by the longitudinal information extraction module in step four comprises: 1) input triplets are obtained by passing the same encoder E, respectively, to obtain latent vectors: ； 2) Morphological similarity network data for any time instant t , both of which can get stable individual feature vectors through the encoder: ； 3) select the individual feature vector at the latest time point as the feature matrix with individual unique identification for the subject and supply the noise prediction network together with the age condition.

8. A system for predicting infant brain functional connectivity based on morphological features and diffusion models based on the method of any one of claims 1 to 7, characterized in that, It comprises a data acquisition and preprocessing module, a morphological feature extraction module, a longitudinal information extraction module, and a diffusion model module; The data acquisition and preprocessing module acquires structural magnetic resonance imaging data of the test infant, and performs intensity non-uniformity correction, head and non-brain tissue removal, white matter / gray matter / cerebrospinal fluid segmentation, and left and right hemisphere and cortical region reconstruction on the acquired structural magnetic resonance imaging data; The morphological feature extraction module extracts brain cortex morphological features from the data preprocessed by the data acquisition and preprocessing module, including cortical thickness, surface area, gray matter volume, myelin content, and curvature, and performs Z-score normalization, calculates the Pearson correlation coefficient between each cortical region, and generates a whole brain morphological similarity network; The longitudinal information extraction module uses a three-tuple network structure to encode the morphological similarity network features from the same subject at different time points and other subjects, and realizes stable extraction of individual-specific features; The diffusion model module realizes classifier-independent guidance through random age condition dropping, combines the individual-specific structural representation and target age information, optimizes a noise prediction network, ensures that the model can balance the accuracy of structural prior and functional connection generation, and dynamically fuses structural features and age constraints to iteratively denoise and generate a brain functional connection atlas conforming to biological laws.

9. A storage medium, characterized by A program for implementing the method of predicting infant brain functional connectivity based on morphological features and diffusion models according to any one of claims 1 to 7 is stored.

10. An electronic device, comprising: The computer device comprises at least one central processing unit, a graphics processing unit, an input / output device, a network communication module, and a storage unit electrically coupled to the central processing unit; the storage unit stores program codes readable and executable by the central processing unit, which implement the method of predicting infant brain functional connectivity based on morphological features and diffusion models according to any one of claims 1 to 7.

Images