A Method and System for Predicting Fetal Brain Age Based on Cerebellar Vermis in Multimodal Feature Fusion

Abstract

This application belongs to the field of fetal brain age prediction technology, and relates to a method and system for predicting fetal brain age of the cerebellar vermis based on multimodal feature fusion. It employs the MST-Mamba segmentation network, and achieves synergy between local detail capture and global semantic modeling by embedding local-global aggregators at each level of the encoder. Simultaneously, a dynamic channel fusion unit is deployed at the jump connection between the encoder and decoder to avoid problems such as boundary ambiguity, missed classification, and misclassification. Through three parallel branches of a multi-granularity morphology-texture collaborative perception architecture, it simultaneously extracts two types of explicit features (macro-geometric and topological morphology) and two types of implicit features (micro-texture), achieving a comprehensive representation of the developmental features of the cerebellar vermis. After standardizing and calibrating the explicit features, they are spliced ​​and fused with the implicit features along the channel dimension to solve the problems of insufficient multimodal feature fusion and lack of calibration. Finally, prediction is performed using a multilayer perceptron regression head, ensuring the accuracy and stability of the brain age prediction results from the source.

Description

Technical Field

[0001] This application belongs to the field of fetal brain age prediction technology, and more specifically, relates to a method and system for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis. Background Technology

[0002] Fetal brain development is one of the core indicators for assessing fetal health in utero. Accurate prediction of brain age can provide important reference for clinicians to judge abnormalities such as delayed development and malformations of the fetal nervous system, and is of key significance for early intervention and prognosis assessment. At present, clinical fetal brain age assessment mainly relies on ultrasound and magnetic resonance imaging (MRI). Ultrasound examination is greatly affected by factors such as operator experience and fetal position, which limits the accuracy of assessment. MRI, on the other hand, has advantages such as high soft tissue resolution, no radiation, and clear visualization of brain anatomy, and has gradually become the preferred imaging method for fetal brain development assessment.

[0003] The core of MRI-based fetal brain age prediction methods lies in extracting brain features and establishing a prediction model to estimate brain age. Among these methods, the cerebellar vermis, as a region with significant morphological changes during fetal development, has a high correlation with overall brain age, making it an important research target for fetal brain age prediction. Existing MRI-based fetal brain age prediction methods using the cerebellar vermis can be broadly divided into two categories: traditional feature engineering methods and deep learning methods. Traditional feature engineering methods mainly rely on doctors manually segmenting the cerebellar vermis region and then extracting the macroscopic geometric features of this region (such as area and diameter) to construct a prediction model. This type of method is not only time-consuming and labor-intensive, but also highly subjective. Furthermore, by using a single explicit feature, it cannot capture the microscopic textural changes during the development of the cerebellar vermis, resulting in prediction accuracy that is difficult to meet the needs of precise clinical assessment.

[0004] To overcome the limitations of manual segmentation and single-feature analysis, deep learning methods have been widely applied in this field, using models such as convolutional neural networks to achieve automatic segmentation and feature extraction of the cerebellar vermis region. However, existing deep learning methods still face several technical bottlenecks, as follows:

[0005] Firstly, the segmentation accuracy is insufficient. The cerebellar vermis is small in size and has a delicate anatomical structure. In addition, the image heterogeneity caused by different scanning devices and scanning parameters makes it difficult for traditional segmentation networks to take into account both local detail capture and global semantic preservation. This can easily lead to problems such as blurred segmentation boundaries, missed or misclassified target regions, which in turn affects the reliability of subsequent feature extraction.

[0006] Secondly, the feature extraction dimension is singular. Existing methods mostly rely on implicit texture features automatically encoded by neural networks, or use explicit geometric and topological features alone. They have not achieved effective fusion of explicit and implicit features and cannot comprehensively characterize the developmental features of the cerebellar vermis.

[0007] Third, the rationality of feature fusion and transmission is lacking. Traditional network encoders and decoders often use simple splicing methods for feature transmission, which makes it difficult to solve the problem of deep feature semantic differences. Furthermore, there is a lack of standardized calibration of extracted multimodal features, resulting in poor feature fusion effects and ultimately affecting the accuracy and stability of brain age prediction.

[0008] In summary, current MRI-based fetal brain age prediction methods based on the cerebellar vermis generally suffer from technical problems such as insufficient targeted image preprocessing, limited feature capture capabilities of segmentation networks, inadequate multimodal feature fusion, and poor quality of input features to the prediction model. As a result, the accuracy and stability of brain age prediction are insufficient to meet the needs of precise clinical diagnosis and treatment. Summary of the Invention

[0009] This invention provides a method and system for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis, which aims to solve the technical problems mentioned in the background art.

[0010] On one hand, the present invention provides a method for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis, comprising the following steps:

[0011] Step 1: Perform standardized preprocessing on the input fetal brain MRI images to obtain standardized images;

[0012] Step 2: Input the standardized image into the MST-Mamba segmentation network, and output the cerebellar vermis binarized mask and the corresponding grayscale region of interest image through the MST-Mamba segmentation network; the MST-Mamba segmentation network uses the visual state space model as the backbone, embeds local-global aggregators at each level of the encoder, and deploys dynamic channel fusion at the jump connection between the encoder and the decoder.

[0013] Step 3: Based on the binarized mask and grayscale region of interest image, multimodal features are extracted through a multi-granularity morphology-texture collaborative perception architecture. The multi-granularity morphology-texture collaborative perception architecture includes three parallel branches, which respectively extract macroscopic geometric explicit features, topological morphological explicit features, and microscopic texture implicit features.

[0014] Step 4: Standardize the extracted macroscopic geometric explicit features and topological morphology explicit features, and then splice and fuse the processed explicit features with the microscopic texture implicit features in the channel dimension to obtain multimodal comprehensive features.

[0015] Step 5: Input the multimodal integrated features into the multilayer perceptron regression head and output the fetal brain age prediction results.

[0016] This invention employs the MST-Mamba segmentation network with a visual state-space model as its backbone. By embedding local-global aggregators at each level of the encoder, it achieves synergy between local detail capture and global semantic modeling. Simultaneously, a dynamic channel fusion processor is deployed at the jump connection between the encoder and decoder, replacing the traditional simple stitching method. This solves the problem of semantic differences between deep and shallow features, significantly improving the segmentation accuracy of the cerebellar vermis and effectively avoiding issues such as boundary blurring, missed segmentation, and missegmentation, ensuring the reliability of the binarized mask and grayscale region of interest images. Furthermore, through three parallel branches of a multi-granularity morphology-texture collaborative perception architecture, it simultaneously extracts two types of explicit features—macro-geometric and topological morphology—and two types of implicit features—micro-texture—breaking through the limitations of existing methods. This invention overcomes the limitations of single-dimensional extraction to achieve a comprehensive representation of cerebellar vermis development features. Furthermore, explicit features are standardized and calibrated before being spliced ​​and fused with implicit features along the channel dimension, resolving issues of insufficient multimodal feature fusion and lack of calibration, thus significantly improving the quality of multimodal integrated features. Finally, the multimodal integrated features are input into a multilayer perceptron regression head for prediction, ensuring the accuracy and stability of brain age prediction results from the source. Therefore, this invention not only achieves efficient collaboration between image preprocessing, segmentation, feature fusion, and prediction, but also significantly improves the accuracy and robustness of fetal brain age prediction through the full mining and reasonable fusion of multimodal features, providing more reliable technical support for clinical fetal nervous system development assessment.

[0017] Preferably, the standardization preprocessing includes:

[0018] A bicubic interpolation algorithm is used to resample MRI images to a uniform physical resolution, converting all images to a fixed isotropic resolution;

[0019] The Z-score normalization method is used to normalize the gray intensity. Based only on the gray mean and standard deviation calculated within the brain tissue mask region, the gray values ​​of the image are calibrated to map the gray values ​​of images acquired by different devices to the same standard normal distribution interval.

[0020] Preferably, the local-global aggregator processes the feature map output by the MST-Mamba encoder to output fused features, specifically including the following steps:

[0021] The local-global aggregator takes the feature map output by the MST-Mamba encoder as input features, performs block window processing on the input features through the local window attention subunit to extract local texture features; at the same time, it performs global average pooling, convolution, ReLU activation and Sigmoid activation on the input features in sequence through the global context gating subunit to generate pixel-level gating coefficients; the local texture features are element-wise weighted using the gating coefficients, and the global residual path subunit retains the global feature flow corresponding to the input features as a residual base, fuses the weighted local texture features with the global residual features, and outputs the fused features to the next level of the encoder.

[0022] Preferably, after receiving the deep features output by the encoder and the shallow features output by the decoder, the dynamic channel fusion unit first shuffles the inherent channel order of the two types of features through a channel shuffling strategy, so that each group of features contains mixed deep and shallow semantics; performs grouping processing on the channel-shuffled features, then generates a spatial weight map through the Sigmoid activation function, performs pixel-level filtering on the grouped features based on the spatial weight map, retains foreground features and filters noise features, and outputs the fused features to the next level of the decoder.

[0023] Preferably, the step of extracting the macroscopic geometric explicit features includes:

[0024] By analyzing the binarized mask, the absolute area is obtained by summing the foreground pixels in the mask, the boundary coordinates of the minimum bounding rectangle of the foreground region of the mask are determined, and the top and bottom diameters and front and back diameters of the minimum bounding rectangle are calculated to form a macroscopic geometric feature vector.

[0025] The steps for extracting explicit features of the topological morphology include:

[0026] Based on the binarized mask, the aspect ratio, compactness, and 7 Hu invariant moments are calculated to form a topological feature vector. The aspect ratio is the ratio of the top and bottom diameters to the front and back diameters. The compactness is the ratio of the square of the perimeter of the mask's foreground region to its absolute area. The Hu invariant moments are constructed through a nonlinear combination of the original moments, the central moments, and the normalized central moments.

[0027] The extraction steps for the implicit features of the micro-texture include:

[0028] A lightweight convolutional neural network with a spatial attention mechanism is introduced as input to a grayscale region of interest image. Multi-layer convolution is performed on the image to generate feature maps. Average pooling and max pooling operations are performed on the feature maps respectively. The results of the two pooling operations are concatenated and then generated into a two-dimensional spatial attention map through convolution and a sigmoid activation function. The feature maps are then element-wise weighted using the two-dimensional spatial attention map. Global average pooling is then performed on the weighted feature maps to output a high-dimensional implicit feature vector as micro-texture features.

[0029] Preferably, the standardization process in step 4 includes:

[0030] The explicit features of macroscopic geometry and explicit features of topology are concatenated to form an explicit feature set. Based on the mean and standard deviation of each dimension of the explicit feature set in the training set, the Z-score standardization method is used to standardize and calibrate each feature in the explicit feature set to obtain a standardized explicit feature vector.

[0031] On the other hand, the present invention provides a fetal brain age prediction system based on multimodal feature fusion of the cerebellar vermis, used to implement the fetal brain age prediction method based on multimodal feature fusion of the cerebellar vermis described in the present invention, comprising:

[0032] Preprocessing module: Used to perform standardized preprocessing on fetal brain MRI images and output standardized images;

[0033] Segmentation module: Built-in MST-Mamba segmentation network, which uses a visual state space model as its backbone and includes a local-global aggregator and a dynamic channel fusion unit. It is used to receive a normalized image and output a binarized mask of the cerebellar vermis and the corresponding grayscale region of interest image.

[0034] Multimodal feature extraction module: It adopts a multi-granularity morphology-texture collaborative perception architecture, which includes three parallel units: macro-geometry extraction unit, topological morphology extraction unit and micro-texture extraction unit.

[0035] Feature fusion module: It is used to standardize the explicit features output by the macro-geometry extraction unit and the topology extraction unit, and then splice and fuse the standardized explicit features with the implicit features output by the micro-texture extraction unit in the channel dimension to generate multimodal comprehensive features.

[0036] Regression prediction module: It has a built-in multilayer perceptron regression head, which is used to receive multimodal integrated features and output the fetal brain age prediction results after processing by the multilayer perceptron regression head.

[0037] Preferably, the local-global aggregator includes a local window attention subunit, a global context gating subunit, and a global residual path subunit;

[0038] The local window attention subunit is used to extract local texture features, the global context gating subunit is used to generate pixel-level gating coefficients and weight the local texture features, and the global residual path subunit is used to preserve the global feature flow and achieve feature fusion.

[0039] Preferably, the dynamic channel fusion unit includes a channel shuffling subunit, a grouping processing subunit, and a spatial weighting subunit;

[0040] The channel shuffling subunit is used to shuffle the order of feature channels, the grouping processing subunit is used to group the shuffled features, and the spatial weighting subunit is used to generate a spatial weight map and filter features.

[0041] Preferably, the macroscopic geometry extraction unit is used to parse the binarized mask, calculate the absolute area, top and bottom diameters and front and back diameters and generate macroscopic geometric feature vectors; the topology extraction unit is used to calculate the aspect ratio, compactness and 7 Hu invariant moments and generate topology feature vectors; the microscopic texture extraction unit includes a lightweight convolutional neural network and a spatial attention submodule, used to extract microscopic texture implicit feature vectors.

[0042] The beneficial effects of this invention include:

[0043] This invention employs the MST-Mamba segmentation network with a visual state-space model as its backbone. By embedding local-global aggregators at each level of the encoder, it achieves synergy between local detail capture and global semantic modeling. Simultaneously, a dynamic channel fusion processor is deployed at the jump connection between the encoder and decoder, replacing the traditional simple stitching method. This solves the problem of semantic differences between deep and shallow features, significantly improving the segmentation accuracy of the cerebellar vermis and effectively avoiding issues such as boundary blurring, missed segmentation, and missegmentation, ensuring the reliability of the binarized mask and grayscale region of interest images. Furthermore, through three parallel branches of a multi-granularity morphology-texture collaborative perception architecture, it simultaneously extracts two types of explicit features—macro-geometric and topological morphology—and two types of implicit features—micro-texture—breaking through the limitations of existing methods. This invention overcomes the limitations of single-dimensional extraction to achieve a comprehensive representation of cerebellar vermis development features. Furthermore, explicit features are standardized and calibrated before being spliced ​​and fused with implicit features along the channel dimension, resolving issues of insufficient multimodal feature fusion and lack of calibration, thus significantly improving the quality of multimodal integrated features. Finally, the multimodal integrated features are input into a multilayer perceptron regression head for prediction, ensuring the accuracy and stability of brain age prediction results from the source. Therefore, this invention not only achieves efficient collaboration between image preprocessing, segmentation, feature fusion, and prediction, but also significantly improves the accuracy and robustness of fetal brain age prediction through the full mining and reasonable fusion of multimodal features, providing more reliable technical support for clinical fetal nervous system development assessment. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0045] Figure 1 The overall process flowchart is provided for embodiments of the present invention.

[0046] Figure 2 This is a schematic diagram of the MST-Mamba segmentation network structure provided in an embodiment of the present invention.

[0047] Figure 3 This is a structural diagram of the LGA module (Local-Global Aggregator) provided in an embodiment of the present invention.

[0048] Figure 4 The diagram shows the structure of the DCF module (Dynamic Channel Fusion Unit) provided in this embodiment of the invention. Detailed Implementation

[0049] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0050] Example 1

[0051] See Figure 1 As shown, this embodiment provides a fetal brain age prediction method based on multimodal feature fusion of the cerebellar vermis. Guided by anatomical priors, it integrates the interpretability of explicit anatomical features with the texture representation ability of implicit deep features. The network structure addresses the problem of lost details in the segmentation of small targets, ultimately achieving accurate brain age prediction. The specific steps are as follows:

[0052] Step 1: Due to differences in parameters such as magnetic field strength, scanning sequence, and pixel spacing among different MRI devices during multi-center clinical data acquisition, images exhibit high heterogeneity in spatial resolution and grayscale distribution. Directly inputting these images into the model can affect the accuracy of feature calculation and generalization ability. Therefore, in this embodiment, a two-step standardization process is used to eliminate device interference and ensure data consistency, as detailed below:

[0053] Unified physical resolution resampling: The voxel spacing of MRI data typically fluctuates between 0.5mm and 1.2mm. Calculating explicit features based on the original pixels can lead to errors due to inconsistent brightness standards. In this embodiment, a bicubic interpolation algorithm is used to uniformly sample all images to a fixed isotropic resolution of 0.8mm × 0.8mm (which can be adjusted to 0.7mm or 0.9mm depending on the device parameters). Compared with bilinear interpolation and nearest neighbor interpolation, it has higher resolution conversion accuracy, can preserve image edge details, avoid artifacts, and ensure that the macroscopic geometric features between samples are physically comparable.

[0054] Gray-level intensity normalization: MRI is a non-quantitative imaging process, and different device gain settings result in no unified standard for the gray-level value range. Gray-level drift can affect network convergence and feature extraction stability. Therefore, in this embodiment, the Z-score normalization method is used, and the mean and standard deviation of gray levels are calculated only based on the brain tissue mask region to avoid background interference. The original gray-level value of each pixel is converted according to the following formula: ;

[0055] In the formula: This represents the original grayscale value in the image to be standardized. This represents the average grayscale value of all pixels within the brain tissue mask region; The standard deviation of all pixel grayscale values ​​within the brain tissue mask region; This represents the standardized grayscale value. The above formula maps the grayscale value to a standard normal distribution interval with a mean of 0 and a standard deviation of 1, so that subsequent network inputs are at the same grayscale scale.

[0056] It is important to note that MRI images are essentially 3D data composed of multiple cuts, each representing the structural information of a specific layer of the fetal brain. In this embodiment, the goal of processing this data is to accurately locate the midsagittal plane containing the clearest cerebellar vermis structure from hundreds of slices. Clinically, the midsagittal plane is characterized by the clearest sulcus and gyri texture. This method uses a gray level co-occurrence matrix (GLCM) to analyze the texture. Because the midsagittal plane has abundant sulcus and gyri structures, its texture information is the most complex. Therefore, by calculating the GLCM energy value of candidate slices, slices with lower energy values ​​(more complex textures) are selected as the final input data. Based on this, a unique and highest-quality 2D median sagittal image is obtained. Then, the 2D median sagittal image is resampled at a uniform physical resolution and its grayscale intensity is normalized to obtain a preprocessed 2D median sagittal image, which serves as the input to the segmentation network in step 2. Secondly, it should be noted that in this embodiment, the original MRI data input in step 1 can be 3D MRI data or a 2D median sagittal image processed by the above-mentioned grayscale co-occurrence matrix analysis. When the input is 3D MRI data, grayscale co-occurrence matrix analysis is performed first. Furthermore, when the input is a 2D median sagittal image, the acquisition of this 2D median sagittal image is not limited in this embodiment; it can be obtained through the above method or other existing techniques.

[0057] Step 2: The cerebellar vermis consists of tiny neural structures. In MRI images (2D midsagittal plane images), the boundaries are blurred and easily interfered with by surrounding tissues. The surface sulci and gyri are delicate, and traditional networks suffer from problems such as loss of detail and misjudgment of background. In this embodiment, the MST-Mamba (morpho-spatial-texture co-Mamba) segmentation network is used. With the visual state space model as the backbone, a local-global aggregator (LGA module) and a dynamic channel fusion (DCF module) are embedded to achieve fine segmentation and output a binary mask and grayscale region of interest images. The details are as follows:

[0058] See Figure 2 As shown, the MST-Mamba network adopts an encoder-decoder architecture. The encoder downsamples the standardized image to extract deep semantic features; the decoder upsamples the features to restore the spatial resolution, and finally outputs a binary mask with the same size as the input. Compared with CNN, the visual spatial model backbone in this embodiment uses linear complexity sequence modeling to reduce memory consumption. Compared with Transformer, it avoids the computational complexity from increasing twice with the image size. At the same time, it introduces convolutional inductive bias to enhance the ability to capture local textures and adapt to the segmentation of small targets.

[0059] See Figure 3 As shown, the LGA module is embedded in each level of the encoder. Through the coordinated action of local feature enhancement and global gating adjustment, it focuses on foreground details and suppresses background noise. The steps are as follows:

[0060] Input reception: The current level feature map (X) of the encoder is used as input. This feature map contains local texture and global semantic information, mixed with background noise.

[0061] Local texture extraction: By using local window attention subunits, X is divided into 7×7 pixel blocks (which can be adjusted to 5×5 or 9×9), and self-attention calculation is performed on the pixels within the window to extract local texture features (F) and capture the details of the grooves and cracks in the worm-like structure.

[0062] Gating coefficient generation: The global context gate subunit sequentially performs global average pooling, 1×1 convolution dimensionality reduction, ReLU activation, 1×1 convolution dimensionality increase, and Sigmoid activation on X to generate pixel-level gating coefficients (Gate, 0~1 range) with the same size as X. The coefficients in the foreground region are close to 1, and the coefficients in the background are close to 0.

[0063] Feature fusion: F is weighted at the element level using gating coefficients. The global residual path subunit retains the global feature flow (F) of X as the residual base. The weighted local features are fused with the global residual features at the element level and output to the next level of the encoder.

[0064] See Figure 2 and Figure 4As shown, in this embodiment, the DCF module is deployed at the jump connection between the encoder and decoder to solve the problem of isolation between deep and shallow feature channels. The steps are as follows:

[0065] Feature reception: Receiving deep semantic features (high semantics, low resolution) from the encoder and shallow spatial features (low semantics, high resolution) from the decoder;

[0066] Channel shuffling: The channel shuffling strategy is used to shuffle the order of the two feature channels, so that each group of features contains mixed semantics of deep and shallow layers, breaking the information isolation;

[0067] Grouping and weighting: The shuffled features are grouped (the number of groups is adaptive to the number of channels, with 8 groups being preferred), and a spatial weight map is generated by Sigmoid activation. Features are filtered at the pixel level to retain the foreground and filter out noise.

[0068] Feature output: The filtered features from each group are concatenated and output to the next level of the decoder to provide accurate fused features for resolution restoration.

[0069] Segmentation result output:

[0070] The decoder outputs a 1-channel feature map, which is activated by Sigmoid and mapped to the 0~1 range. A threshold of 0.5 is set. Pixel values ​​> the threshold are marked as foreground (vermulus, 1), and ≤ the threshold are marked as background (0), resulting in a binarized mask (M). Based on M, a grayscale region of interest image (I) is cropped from the normalized image for subsequent feature extraction.

[0071] Step 3: Extract macroscopic geometric explicit features, topological morphology explicit features, and microscopic texture implicit features through three parallel branches to quantify the developmental state of the vermicompost in all dimensions. The inputs are M and I, and parallel execution improves efficiency.

[0072] Macroscopic geometric explicit feature extraction:

[0073] Absolute area (A) calculation: Count the total number of foreground pixels with a pixel value of 1 in M, i.e., the projected area of ​​the vermiform part in the midsagittal plane. The formula is: ;

[0074] In the formula: Represents the absolute area of ​​the projection; Represents the horizontal and vertical coordinates of a pixel; Indicates the width and height of M; Indicates the mask pixel value (foreground 1, background 0);

[0075] Calculation of top and bottom diameters (CCD) and front and back diameters (APD): Traverse M pixels to find the minimum / maximum coordinates of the foreground. The calculation formulas for CCD and APD are as follows: ; ;

[0076] In the formula: CCD is the vertical diameter (reflecting height development); APD is the anteroposterior diameter (reflecting width development); Used to include boundary pixels to ensure accurate calculations;

[0077] Macro-feature vector generation: concatenation CCD and APD form a vector. Quantify the linear growth pattern.

[0078] Aspect Ratio (AR) Calculation: This reflects the longitudinal extension of the vermis, which increases with gestational age.

[0079] Compactness (C) calculation: Reflects the smoothness of the shape and is sensitive to grooves and cracks. The formula is: ,in: For compactness (the larger the value, the rougher the edges). M represents the foreground perimeter (the number of pixels counted after edge detection). This represents the absolute area of ​​the projection.

[0080] Hu invariant moments calculation: Calculate 7 Hu invariant moments It is constructed through a nonlinear combination of original moments, central moments, and normalized central moments, possessing invariance to translation, rotation, and scaling, and capturing subtle deformations in shape. The specific calculation formula is as follows:

[0081] First calculate Original Moments and central moments : ; ;

[0082] In the formula: for First-order primitive moments; correspond The power of the axis coordinates; correspond The power of the axis coordinates; We obtain the coordinates of the centroid; then we calculate the normalized central moments. Represents pixels The grayscale value at that location; express The first central moment;

[0083] in: Indicates when At that time, for all foreground pixels The coordinates are accumulated to reflect the horizontal distribution of the image, and... Combined with the x-coordinate used to calculate the centroid;

[0084] Indicates when When the formula becomes the sum of all foreground pixel values, in a binary image it represents the "area" (or mass) of the target and is the basic denominator for calculating the centroid and normalization.

[0085] Indicates when When the formula is applied to all foreground pixels, The coordinates are accumulated to reflect the horizontal distribution of the image, which is related to... Combined with the x-coordinate used to calculate the centroid.

[0086] ;

[0087] In the formula: express Normalized central moments of order; Represents the 0th order central moment;

[0088] Final utilization The nonlinear combination constructs 7 Hu moments For example, the first two moments are defined as: ; ;

[0089] In the formula: This represents the second-order normalized center distance, reflecting the image's position in the center. The degree of diffusion in the direction; This represents the second-order normalized central moment, reflecting the image's... The degree of diffusion in the direction; This represents the second-order normalized center distance, reflecting the image's position in the center. The degree of coupling diffusion in the direction; Indicates the first Hu moment, Indicates the second Hu moment, ..., This represents the seventh Hu matrix.

[0090] How technically is the above content disclosed? and This is a conventional technical method in this field, therefore it is omitted here. to The formula.

[0091] Morphological feature vector generation: concatenation of AR, C, , forming vectors The quantitative forms exhibit complex evolutionary patterns.

[0092] An 8-layer lightweight CNN with spatial attention mechanism is employed to avoid overfitting and reduce computational complexity. Input I is passed to the network and then processed through multiple convolutional layers to generate feature maps (C×H×W dimensions). Global average pooling and max pooling are performed on the feature maps, and the concatenated results are then processed by 7×7 convolution and sigmoid activation to generate a 2D attention map. The 2D attention map is used to element-wise weight the feature maps, focusing on groove and fissure regions. Finally, global average pooling is performed on the weighted feature maps to output a 1024-dimensional micro-texture feature vector. Encode deep texture patterns.

[0093] Step 4: The three types of features have large differences in magnitude and attributes, and direct fusion can easily lead to training instability. This step achieves effective fusion through standardization and channel concatenation:

[0094] Explicit feature set construction: splicing and This forms an 11-dimensional explicit feature set. ;

[0095] Explicit feature standardization: based on training set pairs Various dimensions Z-score standardization is adopted: ;

[0096] In the formula: Represents the original eigenvalues; , The mean and standard deviation of the corresponding dimensions of the training set are respectively used; Represents the standardized eigenvalues; based on explicit feature standardization, the standardized explicit feature set is obtained. ;

[0097] Multimodal fusion: Channel-dimensional stitching and This forms a 1036-dimensional comprehensive feature vector. To achieve complementary advantages.

[0098] Step 5: Input the multimodal integrated features into the multilayer perceptron regression head and output the fetal brain age prediction result. The MLP (multilayer perceptron) contains two hidden layers and one output layer, all of which are fully connected layers with ReLU activation function. The input dimension is 1036, the first hidden layer outputs 512 dimensions, the second hidden layer outputs 64 dimensions, and the output layer outputs 1 dimension (predicted brain age).

[0099] Will Inputting into an MLP, the output is a predicted brain age after nonlinear transformation. In this embodiment, the mean squared error (MSE) is used as the loss function, and the optimization objective is... ,in To reflect the actual gestational age, the Adam optimizer was used to iteratively update the parameters until the model converged.

[0100] In summary, this embodiment effectively identifies and suppresses background noise by introducing an LGA (Local-Global Aggregator) module into the encoder, utilizing its unique context gating mechanism, thus solving the problem of traditional networks easily misidentifying surrounding tissues as targets. Secondly, by deploying a dynamic channel fusion unit at skip connections, the interaction between deep and shallow features is enhanced through a channel shuffling strategy, significantly improving the model's ability to capture minute sulcus boundaries and subtle fissures. Furthermore, a macro-morphological-texture multi-granularity collaborative perception architecture is constructed, simulating and surpassing the comprehensive diagnostic thinking of experienced radiologists who "look at size, shape, and texture." Compared to traditional techniques that rely solely on diameter measurements, this model can utilize a significantly increased number of information dimensions. In particular, the addition of morphological and textural flows enables the model to keenly capture the nonlinear developmental feature of the cerebellar vermis in late pregnancy, where the volume growth slows down but the sulcus texture continues to become more complex. This effectively solves the problem of significantly reduced prediction accuracy in late pregnancy using traditional methods, significantly improving the prediction accuracy throughout the entire pregnancy (especially the critical late pregnancy).

[0101] This invention abandons the pure "black box" deep learning model. The introduction of explicit anatomical features (size, shape) gives the prediction results a clear physical meaning. Clinicians can analyze the weights of explicit and implicit texture features in the regression model to determine whether the prediction results are mainly based on volume development or cortical texture development. This characteristic has extremely high application value for pathological diagnosis, helping doctors to distinguish between different types of neurological abnormalities such as "cerebellar hypoplasia" (small volume but normal shape and texture) and "cortical developmental disorders" (normal volume but no sulci / gyri / smooth brain), providing stronger evidence for prenatal eugenic counseling.

[0102] In addition, to address the problem of scarce medical image annotation data, the offloading architecture adopted in this invention reduces the model's dependence on massive amounts of data. Among them, explicit features (such as area and Hu moment) calculated based on anatomical prior knowledge play a strong regularization role in model training, effectively constraining the search space of the neural network and preventing the overfitting problem that pure deep learning models are prone to on small sample data. This significantly improves the model's generalization ability and robustness on data collected from different centers and different devices.

[0103] Example 2

[0104] This embodiment provides a fetal brain age prediction system based on multimodal feature fusion for the cerebellar vermis, used to implement the method described in the above embodiment. The system includes five core modules: a preprocessing module, a segmentation module, a multimodal feature extraction module, a feature fusion module, and a regression prediction module. These modules are sequentially connected and data is transmitted in an orderly manner. The preprocessing module provides consistent input for subsequent modules, the segmentation module outputs accurate target regions, the feature extraction module captures developmental information in all dimensions, the fusion module eliminates feature differences, and the regression prediction module outputs the final result. Details are as follows:

[0105] Preprocessing module: Used to perform standardized preprocessing on fetal brain MRI images, eliminate device heterogeneity, output standardized images, and provide qualified input for the segmentation module; the preprocessing module has built-in bicubic interpolation unit and Z-score normalization unit, which work together to complete the preprocessing;

[0106] The bicubic interpolation unit receives MRI images and performs resampling at a preset isotropic resolution (0.8mm × 0.8mm) to unify the physical resolution, ensure consistent measurement standards for different samples, preserve edge details, and avoid artifacts.

[0107] After receiving the resampled image and the brain tissue mask, the mean gray value (μ) and standard deviation (σ) are calculated based only on the mask region. The image gray value is calibrated using the Z-score formula mentioned above and mapped to the standard normal distribution range to eliminate gray-level drift interference.

[0108] See Figure 2 As shown, the segmentation module receives a standardized image, performs fine segmentation of the cerebellar vermis using the MST-Mamba segmentation network, and outputs a binarized mask and grayscale region of interest image, providing a precise target region for the feature extraction module. The segmentation module incorporates the MST-Mamba segmentation network, which includes a visual state space model backbone, an embedded local-global aggregator (LGA module), and a dynamic channel fusion (DCF module). These modules work together to achieve the segmentation function, as detailed below:

[0109] The visual state space model backbone serves as the core architecture of the encoder-decoder, undertaking the tasks of feature downsampling and upsampling, modeled with linear complexity, and balancing segmentation accuracy and computational efficiency.

[0110] See Figure 3 As shown, the LGA module is embedded in each level of the encoder; the LGA module includes a local window attention subunit, a global context gating subunit, and a global residual path subunit; the local window attention subunit extracts local texture, the global context gating subunit generates gating coefficients, and the global residual path subunit preserves global features, and the three work together to enhance foreground details and suppress background noise.

[0111] See Figure 2 and Figure 4 As shown, the DCF module is deployed at the jump connection between the encoder and decoder; the DCF module includes a channel shuffling subunit, a grouping processing subunit, and a spatial weighting subunit; the channel shuffling subunit is used to shuffle the order of feature channels, the grouping processing subunit is used to group features, and the spatial weighting subunit is used to generate a weight map to filter features, thereby achieving effective fusion of deep and shallow features and improving the accuracy of boundary segmentation.

[0112] The multimodal feature extraction module receives the binarized mask and grayscale region of interest image output by the segmentation module. Through a multi-granularity morphology-texture collaborative perception architecture, it extracts three types of features in parallel, providing full-dimensional feature data for the fusion module. The multimodal feature module contains three parallel feature extraction units, corresponding to macroscopic, morphological, and texture feature extraction respectively, and are operated independently, as detailed below:

[0113] The macro-geometry extraction unit is used to parse the binarized mask and generate macro-features through the area statistics subunit and the diameter calculation subunit. The area statistics subunit calculates the absolute area by counting the total number of foreground pixels according to the formula provided in the method embodiment. The diameter calculation subunit locates the coordinates of the minimum bounding rectangle of the foreground, calculates the upper and lower diameters (CCD) and the front and back diameters (APD), and finally stitches them together to form macro-geometry features.

[0114] The topology extraction unit calculates morphological features based on a binary mask, including aspect ratio calculation subunit, compactness calculation subunit, and Hu invariant moment calculation subunit. The specific calculation formulas are given in the method embodiment. The results obtained from the aspect ratio calculation subunit, compactness calculation subunit, and Hu invariant moment calculation subunit are spliced ​​together to form the topology morphological features.

[0115] The micro-texture extraction unit includes a lightweight CNN sub-network and a spatial attention sub-module. The lightweight CNN sub-network performs multi-layer convolution on grayscale regions of interest to generate feature maps. The spatial attention sub-module performs pooling, convolution, and activation on the feature maps to generate attention maps. After weighting, it is then subjected to global average pooling to output micro-texture features.

[0116] The feature fusion module receives three types of features output by the multimodal feature extraction module, and generates multimodal comprehensive features through standardization processing and channel concatenation to eliminate feature differences and provide input for the regression prediction module. Specifically, the feature fusion module has a built-in feature concatenation unit and a Z-score standardization unit.

[0117] The feature splicing unit first splices macroscopic geometric features and topological morphological features to obtain explicit features, which are then standardized and spliced ​​with microscopic texture feature channels to generate multimodal comprehensive features.

[0118] The Z-score normalization unit calculates the mean and standard deviation of each dimension of the explicit features based on the training set data, performs normalization, and outputs the normalized explicit features. These are then used to concatenate with the micro-texture feature channels to generate multimodal integrated features.

[0119] The regression prediction module receives multimodal integrated features, constructs a nonlinear mapping between features and fetal brain age through an MLP regression head, and outputs accurate brain age prediction results; the regression prediction module has an MLP regression head and a model training unit built in it.

[0120] The MLP regression head includes two hidden layers and one output layer, all of which are fully connected layers. It is then combined with the ReLU activation function to achieve non-linear feature transformation. The input dimension is 1036 and the output is 1-dimensional, which is the prediction of brain age.

[0121] The model training unit uses MSE as the loss function and Adam as the optimizer to iteratively update the MLP parameters, minimize the error between the predicted brain age and the actual gestational age, and ensure model convergence and prediction accuracy.

[0122] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting fetal brain age based on cerebellar vermis using multimodal feature fusion, characterized in that, Includes the following steps: Step 1: Perform standardized preprocessing on the input fetal brain MRI images to obtain standardized images; Step 2: Input the standardized image into the MST-Mamba segmentation network. The MST-Mamba segmentation network outputs a binarized mask of the cerebellar vermis and the corresponding grayscale region of interest image. The MST-Mamba segmentation network uses a visual state space model as its backbone, embedding local-global aggregators at each level of the encoder and deploying dynamic channel fusion units at the jump connections between the encoder and decoder. After receiving the deep features output by the encoder and the shallow features output by the decoder, the dynamic channel fusion unit first shuffles the inherent channel order of the two types of features through a channel shuffling strategy, so that each group of features contains mixed deep and shallow semantics. The shuffled features are then grouped, and a spatial weight map is generated through the Sigmoid activation function. Based on the spatial weight map, the grouped features are pixel-level filtered to retain foreground features and filter out noise features. The fused features are then output to the next level of the decoder. Step 3: Based on the binarized mask and grayscale region of interest image, multimodal features are extracted through a multi-granularity morphology-texture collaborative perception architecture. The multi-granularity morphology-texture collaborative perception architecture includes three parallel branches, which respectively extract macroscopic geometric explicit features, topological morphological explicit features, and microscopic texture implicit features. The steps for extracting the macroscopic geometric explicit features include: By analyzing the binarized mask, the absolute area is obtained by summing the foreground pixels in the mask, the boundary coordinates of the minimum bounding rectangle of the foreground region of the mask are determined, and the top and bottom diameters and front and back diameters of the minimum bounding rectangle are calculated to form a macroscopic geometric feature vector. The steps for extracting explicit features of the topological morphology include: Based on the binarized mask, the aspect ratio, compactness, and 7 Hu invariant moments are calculated to form a topological feature vector. The aspect ratio is the ratio of the top and bottom diameters to the front and back diameters. The compactness is the ratio of the square of the perimeter of the mask's foreground region to its absolute area. The Hu invariant moments are constructed through a nonlinear combination of the original moments, the central moments, and the normalized central moments. The extraction steps for the implicit features of the micro-texture include: A lightweight convolutional neural network with a spatial attention mechanism is introduced as input to a grayscale region of interest image. Multi-layer convolution is performed on the image to generate feature maps. Average pooling and max pooling operations are performed on the feature maps respectively. The two pooling results are concatenated and then generated into a two-dimensional spatial attention map through convolution and a sigmoid activation function. The feature maps are then element-wise weighted using the two-dimensional spatial attention map. Global average pooling is then performed on the weighted feature maps to output a high-dimensional implicit feature vector as micro-texture features. Step 4: Standardize the extracted macroscopic geometric explicit features and topological morphology explicit features, and then splice and fuse the processed explicit features with the microscopic texture implicit features in the channel dimension to obtain multimodal comprehensive features. Step 5: Input the multimodal integrated features into the multilayer perceptron regression head and output the fetal brain age prediction results.

2. The method for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis according to claim 1, characterized in that, The standardized preprocessing includes: A bicubic interpolation algorithm is used to resample MRI images to a uniform physical resolution, converting all images to a fixed isotropic resolution; The Z-score normalization method is used to normalize the gray intensity. Based only on the gray mean and standard deviation calculated within the brain tissue mask region, the gray values ​​of the image are calibrated to map the gray values ​​of images acquired by different devices to the same standard normal distribution interval.

3. The method for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis according to claim 1, characterized in that, The local-global aggregator processes the feature map output by the MST-Mamba encoder to output fused features, specifically including the following steps: The local-global aggregator takes the feature map output by the MST-Mamba encoder as input features, performs block window processing on the input features through the local window attention subunit to extract local texture features; at the same time, it performs global average pooling, convolution, ReLU activation and Sigmoid activation on the input features in sequence through the global context gating subunit to generate pixel-level gating coefficients; the local texture features are element-wise weighted using the gating coefficients, and the global residual path subunit retains the global feature flow corresponding to the input features as a residual base, fuses the weighted local texture features with the global residual features, and outputs the fused features to the next level of the encoder.

4. The method for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis according to claim 1, characterized in that, The standardization process in step 4 includes: The explicit features of macroscopic geometry and explicit features of topology are concatenated to form an explicit feature set. Based on the mean and standard deviation of each dimension of the explicit feature set in the training set, the Z-score standardization method is used to standardize and calibrate each feature in the explicit feature set to obtain a standardized explicit feature vector.

5. A fetal brain age prediction system based on multimodal feature fusion of the cerebellar vermis, characterized in that, The method for predicting fetal brain age based on multimodal feature fusion of the cerebellar vermis, as described in any one of claims 1 to 4, comprises: Preprocessing module: Used to perform standardized preprocessing on fetal brain MRI images and output standardized images; Segmentation Module: Built-in MST-Mamba segmentation network, which uses a visual state space model as its backbone and includes a local-global aggregator and a dynamic channel fusion unit. It is used to receive a normalized image and output a binarized mask of the cerebellar vermis and the corresponding grayscale region of interest image. The dynamic channel fusion unit includes a channel shuffling subunit, a grouping processing subunit, and a spatial weighting subunit. The channel shuffling subunit is used to shuffle the order of feature channels, the grouping processing subunit is used to group the shuffled features, and the spatial weighting subunit is used to generate a spatial weight map and filter features. The multimodal feature extraction module adopts a multi-granularity morphology-texture collaborative perception architecture, comprising three parallel units: a macro-geometry extraction unit, a topological morphology extraction unit, and a micro-texture extraction unit. The macro-geometry extraction unit is used to parse the binarized mask, calculate the absolute area, top and bottom diameters, and front and back diameters, and generate macro-geometry feature vectors. The topological morphology extraction unit is used to calculate the aspect ratio, compactness, and 7 Hu invariant moments, and generate topological morphology feature vectors. The micro-texture extraction unit includes a lightweight convolutional neural network and a spatial attention submodule, used to extract implicit feature vectors of micro-texture. Feature fusion module: It is used to standardize the explicit features output by the macro-geometry extraction unit and the topology extraction unit, and then splice and fuse the standardized explicit features with the implicit features output by the micro-texture extraction unit in the channel dimension to generate multimodal comprehensive features. Regression prediction module: It has a built-in multilayer perceptron regression head, which is used to receive multimodal integrated features and output the fetal brain age prediction results after processing by the multilayer perceptron regression head.

6. The cerebellar vermis fetal brain age prediction system based on multimodal feature fusion according to claim 5, characterized in that, The local-global aggregator includes a local window attention subunit, a global context gating subunit, and a global residual path subunit; The local window attention subunit is used to extract local texture features, the global context gating subunit is used to generate pixel-level gating coefficients and weight the local texture features, and the global residual path subunit is used to preserve the global feature flow and achieve feature fusion.

Images