Fetal brain maturity assessment method based on three-dimensional point cloud geometry multi-scale feature fusion

Abstract

The application provides a fetal brain maturity evaluation method based on three-dimensional point cloud geometry multi-scale feature fusion. The method comprises the following steps: acquiring a fetal brain ultrasound image, and extracting a brain three-dimensional ultrasound volume data of the ultrasound image; preprocessing the brain three-dimensional ultrasound volume data to generate a point cloud feature; constructing a point cloud recognition network; the point cloud recognition network comprises a front feature mapping layer, a four-stage serial SA module, a cross-scale geometry feature fusion module GMSFA and a prediction head; constructing an output layer consistency loss function to associate a continuous gestational age estimation value and a discrete maturity category probability distribution; constructing an overall optimization target loss function; based on the overall optimization target loss function and the point cloud feature, the point cloud recognition network is trained for subsequent fetal brain maturity evaluation based on the trained point cloud recognition network. The application has the advantages of preventing information leakage, high robustness and self-adaptive multi-scale geometry feature extraction for evaluating fetal brain maturity.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent assessment of fetal brain maturity, and more specifically, to a method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features. Background Technology

[0002] Fetal brain development is a continuous and highly structured biological process. Throughout pregnancy, the fetal brain expands in volume while the shape of its hemispheres changes. As the curvature of the brain surface gradually changes, structures such as sulci and gyri appear. Quantitatively characterizing these changes and evaluating fetal brain maturity is of significant clinical importance for assessing fetal congenital central nervous system abnormalities and long-term susceptibility to neurological diseases.

[0003] Currently, prenatal ultrasound is the primary method for assessing fetal brain maturity. However, existing technologies face the following technical bottlenecks when evaluating fetal brain maturity:

[0004] (1) Limited ultrasound image quality: Ultrasound images have speckle noise, skull acoustic shadow and matrix attenuation, resulting in non-uniform density and local missingness in the extracted 3D surface point cloud.

[0005] (2) Poor adaptability of general point cloud networks: Existing general 3D point cloud networks (RF, PointNet, PointNet++, DGCNN, PointMamba, PointNet++ + GMSFA) use a fixed radius grouping method, which cannot adapt to the feature differences between complex sulci and smooth regions of the brain.

[0006] (3) Information leakage leads to model degradation: Existing assessment models often use the actual gestational age as an auxiliary feature directly or indirectly, which leads to "circular reasoning" (inferring brain maturity directly from the input gestational age, rather than learning the actual brain morphological features), which seriously affects the model's generalization ability and actual clinical application value.

[0007] Therefore, we urgently need a new method for assessing fetal brain maturity to address the technical shortcomings of existing technologies in assessing fetal brain maturity. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art or related art.

[0009] Therefore, the purpose of this invention is to propose a method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features.

[0010] To achieve the above objectives, the present invention provides a method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features. This method includes: Step S1: acquiring a fetal cranial ultrasound image and extracting three-dimensional ultrasound volume data of the cranium from the ultrasound image; Step S2: preprocessing the three-dimensional ultrasound volume data of the cranium to generate point cloud features, which serve as input to a point cloud recognition network; Step S3: constructing a point cloud recognition network for adaptive extraction of multi-scale geometric features; wherein the point cloud recognition network includes: a pre-feature mapping layer, a four-level cascaded SA module, a cross-scale geometric feature fusion module GMSFA, and a prediction head; wherein the pre-feature mapping layer is a Shared Multilayer Perceptron. The MLP performs dimensionality mapping on the point cloud features, and the output features are sequentially fed into the first-level SA module, the second-level SA module, the third-level SA module, and the fourth-level SA module. The execution flow of each level SA module is as follows: input point set xyz features → perform FPS farthest point sampling to filter out the center point → adopt curvature adaptive radius grouping mechanism to construct local neighborhoods → complete coarse aggregation of neighborhood features using local max pooling → use curvature attention module to modulate feature weights → use Shared MLP to extract neighborhood principal point features to complete feature dimensionality upgrade → output the geometric features of this level; through the GMSFA module, the multi-scale features output by the four-level SA modules are uniformly mapped to the M-dimensional feature dimension to complete feature resolution alignment, and learnable gating weights are used to achieve adaptive fusion of shallow fine-grained geometric features and deep global semantic features. An independent MLP is configured for each level scale feature to complete differentiated feature transformation and generate the fused global geometric representation; the global geometric representation is input into the prediction head Shared. The MLP completes feature mapping; the output of the prediction head is divided into a three-level maturity classification head and a gestational age regression head; the three-level maturity classification head and the gestational age regression head output the discrete maturity category probability distribution and continuous gestational age estimate of the fetal brain, respectively; Step S4: Construct an output layer consistency loss function to correlate the continuous gestational age estimate with the discrete maturity category probability distribution; Step S5: Based on the output layer consistency loss function, construct an overall optimization target loss function; Step S6: Based on the overall optimization target loss function and the point cloud features, train the point cloud recognition network for subsequent fetal brain maturity assessment based on the trained point cloud recognition network; wherein, during the training process of the point cloud recognition network, only the real gestational age is used as the supervision target, and the real gestational age label does not participate in the network feature input to avoid data leakage.

[0011] Preferably, step S2 specifically includes: step S2.1: scaling the three-dimensional ultrasound volume data of the brain; step S2.2: manually segmenting the scale-normalized three-dimensional ultrasound volume data of the brain to extract the region of interest mesh on the surface of the fetal brain, and performing surface reconstruction, mesh cleaning and hole filling; step S2.3: uniformly sampling the processed mesh to obtain a surface point cloud containing N sampling points, and calculating the local surface normal and average curvature of each sampling point; step S2.4: constructing the input features corresponding to the sampling points based on the three-dimensional coordinates, local surface normal and average curvature of each sampling point to generate point cloud features, which are used as input to the subsequent point cloud recognition network.

[0012] Preferably, in step S2.4, the expression corresponding to the input feature is:

[0013] ;

[0014] in, In three-dimensional coordinates, For local surface normals, The mean curvature.

[0015] Preferably, N is 4096.

[0016] Preferably, in step S3, the pre-feature mapping layer performs a 7→64→64 operation on the original 4096×7-dimensional point cloud features. Dimensional mapping; the FPS farthest point sampling iteratively selects the point farthest from the selected samples in space to ultimately select a specified number of points to ensure that the sampling points are uniformly distributed on the 3D surface; the number of key points after downsampling by the four-level SA module is 1024, 512, 256 and 128 respectively; the curvature adaptive radius grouping mechanism dynamically adjusts the neighborhood search radius based on the average curvature of the center point itself, and the base radius rbase configured by the four-level SA module is 0.12, 0.18, 0.28 and 0.4 respectively, while limiting the maximum number of neighborhood points k of a single center point to 24, 32, 48 and 64 respectively; the curvature attention module uses the curvature contrast between the center point and the neighborhood sampling points to modulate the attention weight coefficient to eliminate the prior bias of fixed assignment of high curvature feature weights; the four-level SA module outputs multi-scale geometric features of 128, 256, 384 and 512 dimensions respectively; M is 128; the prediction head Shared The network structure of the MLP is 128→18→64. The three-level maturity classification head outputs the discrete maturity category probability distribution of the fetal brain based on 64-dimensional features, and the gestational age regression head outputs the continuous gestational age estimate based on 64-dimensional features.

[0017] Preferably, for the j-th center point of the l-th layer, the curvature adaptive radius grouping mechanism is used to calculate the local grouping radius. The formula is:

[0018] ;

[0019] in, The preset base radius, It is the Sigmoid activation function. For learnable adjustment of weight coefficients, Let be the normalized curvature at the center point j.

[0020] Preferably, a curvature attention module is used. Within the adaptive neighborhood, the formula for calculating the attention weight coefficient of neighborhood point i is:

[0021] ;

[0022] in, , For a learnable projection matrix, For feature dimension, This is a hyperparameter for controlling curvature contrast. Let i be a feature of point i in layer l. The features of center point j in layer l, To traverse the features at layer l, In order to be in The points traversed within the inner loop; For the l-th layer, the set of local neighborhood points of the center point j is obtained by adaptive radius grouping. Let i be the average curvature of point i. The average curvature at point j; The average curvature of the point traversal;

[0023] From attention weight coefficient The updated local features are as follows:

[0024] ;

[0025] in, It is a learnable projection matrix.

[0026] Preferably, in step S3, the formula for the learnable gating weights is:

[0027] ;

[0028] in, It is a learnable gating weight. It is the global multi-scale feature vector obtained by fusion. It is the layer point feature of the improved output of layer l. It is an MLP specific to layer l.

[0029] Preferably, in step S4, the formula for the output layer consistency loss function is:

[0030] ;

[0031] in, Preset maturity category range The midpoint value, To predict successive gestational ages, To predict the probability of discrete maturity categories, C is the number of auxiliary categories, and C is 3.

[0032] Preferably, in step S5, the formula for the overall optimization objective loss function is:

[0033] ;

[0034] in, and These are the true gestational age and the true discrete maturity labels, respectively. For the prediction class, y represents the actual maturity level analogy. For Huber's return loss, For weighted cross-entropy loss, It is 0.35. It is 0.08.

[0035] Compared with existing technologies, the fetal brain maturity assessment method based on the fusion of three-dimensional point cloud geometric multi-scale features provided by this invention has the following beneficial effects:

[0036] (1) Prevents data leakage and has strong generalization: This invention eliminates the risk of “target leakage” from the root by cutting off the connection between the real gestational age and the network input / hidden layer, relying only on pure geometric deformation features for prediction, and using the output layer consistency loss to implement constraints.

[0037] (2) Adaptive multi-scale extraction, robust to noise: The introduced adaptive radius grouping and curvature-guided attention mechanism can automatically expand the receptive field in complex areas such as brain sulci and gyri, and tighten the neighborhood in smooth areas, effectively overcoming the interference caused by ultrasound image artifacts and non-uniform density of point clouds.

[0038] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0039] Figure 1A schematic diagram of the overall process of a fetal brain maturity assessment method based on three-dimensional point cloud geometric multi-scale feature fusion is shown in an embodiment of the present invention.

[0040] Figure 2 This diagram illustrates a specific process for preprocessing three-dimensional ultrasound volume data of the cranium according to an embodiment of the present invention.

[0041] Figure 3 A schematic diagram of the specific architecture of a point cloud recognition network according to an embodiment of the present invention is shown;

[0042] Figure 4 This diagram illustrates the execution flow of each SA module in a point cloud recognition network according to an embodiment of the present invention. Detailed Implementation

[0043] To better understand the above-mentioned objects, features, and advantages of the present invention, such as Figures 1 to 4 As shown in the accompanying drawings and specific embodiments, the present invention will be further described in detail below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0045] Figure 1 This diagram illustrates the overall workflow of a fetal brain maturity assessment method based on three-dimensional point cloud geometric multi-scale feature fusion, according to an embodiment of the present invention. Figure 1 As shown, this method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features includes:

[0046] Step S1: Obtain fetal cranial ultrasound images and extract three-dimensional ultrasound volume data of the cranium from the ultrasound images;

[0047] Step S2: Preprocess the three-dimensional ultrasound volume data of the cranium to generate point cloud features, which serve as input to the point cloud recognition network;

[0048] Step S3: Construct a point cloud recognition network for adaptive extraction of multi-scale geometric features; wherein, the point cloud recognition network includes: a pre-feature mapping layer, a four-level cascaded SA module, a cross-scale geometric feature fusion module GMSFA, and a prediction head;

[0049] The pre-feature mapping layer is a shared multilayer perceptron (Shared MLP), which performs dimension mapping on the original point cloud features and outputs features that are sequentially fed into the first-level SA module, the second-level SA module, the third-level SA module, and the fourth-level SA module.

[0050] The execution flow of each SA module is as follows: input point set xyz features → perform FPS farthest point sampling to filter out the center point → adopt curvature adaptive radius grouping mechanism to construct local neighborhood → local max pooling to complete the coarse aggregation of neighborhood features → adopt curvature attention module to modulate feature weights → adopt Shared MLP to extract neighborhood principal point features to complete feature dimensionality upgrade → output the center point features of this level.

[0051] The GMSFA module maps the multi-scale features output by the four-level SA module to the M-dimensional feature dimension to achieve feature resolution alignment. Learnable gated weights are used to achieve adaptive fusion of shallow fine-grained geometric features and deep global semantic features. An independent MLP is configured for each scale feature to complete differentiated feature transformation and generate the fused global geometric representation.

[0052] The global geometric representation is input into the prediction head Shared MLP to complete feature mapping; the output of the prediction head is divided into a three-level maturity classification head and a gestational age regression head; the three-level maturity classification head and the gestational age regression head output the discrete maturity category probability distribution and continuous gestational age estimate of the fetal brain respectively.

[0053] Step S4: Construct an output layer consistency loss function to correlate continuous gestational age estimates with discrete maturity category probability distributions;

[0054] Step S5: Construct the overall optimization objective loss function based on the output layer consistency loss function;

[0055] Step S6: Based on the overall optimized target loss function and point cloud features, train the point cloud recognition network for subsequent fetal brain maturity assessment based on the trained point cloud recognition network; during the training process of the point cloud recognition network, only the real gestational age is used as the supervision target, and the real gestational age label is not included in the network feature input to avoid data leakage.

[0056] In this embodiment, the present invention provides a fetal brain maturity assessment method that is preventive of information leakage, highly robust, and capable of adaptively extracting multi-scale geometric features.

[0057] The technical solution of the present invention will be illustrated below with a specific embodiment. This specific embodiment of the fetal brain maturity assessment method based on three-dimensional point cloud geometric multi-scale feature fusion is implemented through the following steps:

[0058] (1) Step S1: Data acquisition and standardization preprocessing.

[0059] Ultrasound images of the fetal brain are acquired, and three-dimensional ultrasound volume data (DICOM) of the fetal brain is extracted from the ultrasound images. The three-dimensional ultrasound data is scaled and then manually segmented to extract the region of interest (ROI) mesh of the fetal brain surface. Surface reconstruction, mesh cleaning, and hole filling are then performed. A surface point cloud containing N sampling points is obtained by uniformly sampling from the mesh, and the local surface normal and average curvature corresponding to each sampling point are calculated to construct point cloud features, thereby capturing different geometric aspects of fetal brain maturation and stimulating data-level fusion design.

[0060] The expression corresponding to the feature vector of the input point is:

[0061] ;

[0062] in, In three-dimensional coordinates, For local surface normals, The mean curvature is N, which is 4096.

[0063] The specific process is as follows: Figure 2 As shown: De-identification and format conversion mainly convert the original DICOM ultrasound images into NIfTI format for further processing; Resampling and grayscale normalization: Resampling refers to unifying the voxel spacing of each case to the same size, in this example, it is unified to 0.6*0.6*0.6mm; Grayscale normalization uses Min-Max normalization to map the grayscale values ​​of the image to the [0,1] interval; Fetal brain segmentation: The annotation of fetal brain segmentation was completed collaboratively by two clinicians and one senior ultrasound physician. After initial annotation, double review and correction, and consensus on disagreements, a unified and reliable manual segmentation standard was formed; Marching cubes surface reconstruction: Based on Marching Cubes The algorithm extracts isosurfaces from brain tissue and constructs closed, connected triangular patch surfaces to eliminate geometric distortions caused by discrete voxels. Mesh cleaning and hole filling are performed: the reconstructed mesh undergoes connected component filtering, hole filling, and mild smoothing to remove redundancy and defects, constructing a topologically sound brain surface to support subsequent curvature calculation and point cloud sampling. Normal and average curvature calculations are performed: after mesh cleaning, the vertex normals are calculated, and the average curvature of the brain surface is solved to quantify the degree of curvature of the brain tissue surface and extract geometric features. Sampling: 4096 points are uniformly sampled from each acceptable mesh to generate a point cloud.

[0064] (2) Step S2: Construct a point cloud recognition network to perform multi-scale geometric feature adaptive extraction.

[0065] The input point feature vector is input into the hierarchical point cloud backbone network, which performs farthest point sampling and local grouping through four abstract levels. In the local grouping, an adaptive radius grouping mechanism is adopted to dynamically adjust the receptive field radius of each centroid based on the average curvature of the sampling points. This is used to increase the receptive field of geometrically complex regions, thereby addressing the problems introduced by uneven surface sampling and alleviating insufficient sampling and segmentation uncertainty in high curvature regions.

[0066] The specific architecture of the point cloud recognition network is as follows: Figure 3 As shown: The point cloud recognition network includes: a pre-feature mapping layer, a four-level cascaded SA module, a cross-scale geometric feature fusion module GMSFA, and a prediction head; wherein, the pre-feature mapping layer is a shared multilayer perceptron (Shared MLP), which performs dimensional mapping on the point cloud features, and the output features are sequentially fed into the first-level SA module, the second-level SA module, the third-level SA module, and the fourth-level SA module; as shown... Figure 4 As shown, the execution flow of each SA module is as follows: input point set xyz features → perform FPS farthest point sampling to filter out the center point → adopt curvature adaptive radius grouping mechanism to construct local neighborhood → local max pooling to complete the coarse aggregation of neighborhood features → adopt curvature attention module to modulate feature weights → adopt Shared MLP to extract the principal point features of the neighborhood to complete feature dimensionality upgrade → output the center point features of this level; through the GMSFA module, the multi-scale features output by the four-level SA modules are uniformly mapped to the M-dimensional feature dimension to complete feature resolution alignment. Learnable gating weights are adopted to achieve adaptive fusion of shallow fine-grained geometric features and deep global semantic features. An independent MLP is configured for each level of scale features to complete differentiated feature transformation and generate the fused global geometric representation; the global geometric representation is input to the prediction head Shared MLP to complete feature mapping; the output of the prediction head is divided into a three-level maturity classification head and a gestational age regression head; the three-level maturity classification head and the gestational age regression head output the discrete maturity category probability distribution and continuous gestational age estimate of the fetal brain.

[0067] In this specific embodiment, the pre-feature mapping layer performs a 7→64→64 operation on the original 4096×7-dimensional point cloud features. Dimensional mapping; the FPS farthest point sampling iteratively selects the point farthest from the selected samples in space to ultimately select a specified number of points to ensure that the sampling points are uniformly distributed on the 3D surface; the number of key points after downsampling by the four-level SA module is 1024, 512, 256 and 128 respectively; the curvature adaptive radius grouping mechanism dynamically adjusts the neighborhood search radius based on the average curvature of the center point itself, and the base radius rbase configured by the four-level SA module is 0.12, 0.18, 0.28 and 0.4 respectively, while limiting the maximum number of neighborhood points k of a single center point to 24, 32, 48 and 64 respectively; the curvature attention module uses the curvature contrast between the center point and the neighborhood sampling points to modulate the attention weight coefficient to eliminate the prior bias of fixed assignment of high curvature feature weights; the four-level SA module outputs multi-scale geometric features of 128, 256, 384 and 512 dimensions respectively; M is 128.

[0068] like Figure 4 As shown, FPS stands for Farthest Point Sampling, iteratively selecting the point farthest from the selected samples in space, and finally selecting a specified number of points to ensure that the sampled points are evenly distributed on the 3D curved surface and do not cluster locally; Adaptive radius grouping: Curvature adaptively constructs local neighborhoods by dynamically radii for each center point; Neighborhood pooling: uses local maxpooling for pooling operations; Curvature attention: learns attention based on curvature modulation to remove the prior influence of "high curvature must be important"; Shared MLP: extracts the features of the principal points in the neighborhood, performs dimensionality upscaling, and then outputs the point features.

[0069] For the j-th centroid of the l-th layer, the local grouping radius is calculated using the aforementioned curvature adaptive radius grouping mechanism. The formula is:

[0070] ;

[0071] in, The preset base radius, It is the Sigmoid activation function. For learnable adjustment of weight coefficients, Let be the normalized curvature at the center point j.

[0072] Using a curvature attention module, the formula for calculating the attention weight coefficient of point i within the adaptive neighborhood is as follows:

[0073] ;

[0074] in, , For a learnable projection matrix, For feature dimension, This is a hyperparameter for controlling curvature contrast. Let i be a feature of point i in layer l. The features of center point j in layer l, To traverse the features at layer l, point, for In the l-th layer, the set of local neighborhood points of the center point j (obtained by adaptive radius grouping). Let i be the average curvature of point i. Let j be the average curvature. The average curvature of the point traversal;

[0075] From attention weight coefficient The updated local features are as follows:

[0076] ;

[0077] in, It is a learnable projection matrix.

[0078] (3) Step S3: Curvature-guided local attention association.

[0079] Within each adaptive neighborhood, a curvature-guided attention mechanism is used to adjust the attention weight allocation based on the curvature contrast between each point in the neighborhood and the centroid, and to fuse the local geometric features of each level. Specifically: 1) Linearly project the features of the center point and the neighborhood points; 2) Calculate the scaled dot product to obtain the feature similarity; 3) Modulate by superimposing curvature difference terms; 4) Perform Softmax normalization in the local neighborhood to obtain the attention weights; 5) Aggregate the neighborhood features using the weights.

[0080] (4) Step S4: Cross-scale geometric feature fusion.

[0081] By aligning all hierarchical features to the coarsest scale through the above sampling, the problem of inconsistent point counts and spatial resolution across encoder layers is resolved, enabling the fusion of shallow fine-grained geometric features and deep global semantic features; and the learned gating weights are then used. The system performs fusion to generate a fused global geometric representation. Compared with fixed rule fusion methods, it can dynamically adjust the feature contributions of each level and fully combine shallow geometric details with deep global semantics. Each layer is configured with a dedicated MLP, which allows the network to perform differentiated transformations on features at different scales and adapt to the different representation characteristics of coarse and fine level features.

[0082] The formula for the learnable gating weights is:

[0083] ;

[0084] in, It is a learnable gating weight. It is the global multi-scale feature vector obtained by fusion. It is the layer point feature of the improved output of layer l. It is an MLP specific to layer l.

[0085] (5) Step S5: Leakage prevention multi-task prediction and joint consistency constraint.

[0086] The global geometric representation is input into the main prediction head and the auxiliary prediction head respectively, and the continuous gestational age estimate and the discrete maturity category probability distribution are output respectively. The output layer consistency loss function is introduced to associate the continuous gestational age estimate and the discrete maturity category probability distribution, and during the training process, it is ensured that the real gestational age is only used as a supervision target and does not participate in the feature input.

[0087] To ensure the prediction of continuous gestational age With predicting discrete maturity category probabilities Regarding physical and logical consistency, the formula for the output layer consistency loss function is:

[0088] ;

[0089] in, Preset maturity category range The midpoint value, To predict successive gestational ages, To predict the probability of discrete maturity categories, C is the number of auxiliary categories, and C is 3.

[0090] The formula for the overall optimization objective loss function is:

[0091] ;

[0092] in, and These are the true gestational age and the true discrete maturity labels, respectively. For the prediction class, y represents the actual maturity level analogy. For Huber's return loss, For weighted cross-entropy loss, It is 0.35. It is 0.08.

[0093] (6) Verify the experimental results.

[0094] The proposed model achieves the lowest error, with a MAE of 0.98 ± 0.05 weeks, an RMSE of 1.25 ± 0.06 weeks, and a pooled cross-validation R² of 0.73. Compared to the PointNet++ baseline, the full model reduces the MAE by 0.28 weeks, equivalent to a relative reduction of 22.2%, and improves the pooled R² from 0.55 to 0.73. Compared to PointMamba, the full model reduces the MAE by 0.07 weeks, equivalent to a relative reduction of 6.7%. The performance improvements of this model compared to existing and recent point cloud baseline models suggest that the fusion of normals, curvature, and multi-scale geometry provides supplementary evidence for fetal brain maturity.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features, characterized in that, include: Step S1: Acquire fetal cranial ultrasound images and extract the three-dimensional ultrasound volume data of the cranium from the ultrasound images; Step S2: Preprocess the three-dimensional ultrasound volume data of the cranium to generate point cloud features, which serve as input to the point cloud recognition network; Step S3: Construct a point cloud recognition network for adaptive extraction of multi-scale geometric features; wherein, the point cloud recognition network includes: a pre-feature mapping layer, a four-level cascaded SA module, a cross-scale geometric feature fusion module GMSFA, and a prediction head; The preceding feature mapping layer is a shared multilayer perceptron (Shared MLP), which performs dimension mapping on the point cloud features and outputs features that are sequentially fed into the first-level SA module, the second-level SA module, the third-level SA module, and the fourth-level SA module. The execution flow of each SA module is as follows: input point set xyz features → perform FPS farthest point sampling to filter out the center point → adopt curvature adaptive radius grouping mechanism to construct local neighborhood → local max pooling to complete the coarse aggregation of neighborhood features → adopt curvature attention module to modulate feature weights → adopt Shared MLP to extract neighborhood principal point features to complete feature dimensionality upgrade → output the geometric features of this level. The GMSFA module maps the multi-scale features output by the four-level SA module to the M-dimensional feature dimension to achieve feature resolution alignment. Learnable gated weights are used to achieve adaptive fusion of shallow fine-grained geometric features and deep global semantic features. An independent MLP is configured for each scale feature to complete differentiated feature transformation and generate the fused global geometric representation. The global geometric representation is input into the prediction head Shared MLP to complete feature mapping; the output of the prediction head is divided into a three-level maturity classification head and a gestational age regression head; the three-level maturity classification head and the gestational age regression head output the discrete maturity category probability distribution and continuous gestational age estimate of the fetal brain respectively. Step S4: Construct an output layer consistency loss function to correlate continuous gestational age estimates with discrete maturity category probability distributions; Step S5: Based on the output layer consistency loss function, construct the overall optimization target loss function; Step S6: Based on the overall optimization target loss function and the point cloud features, train the point cloud recognition network for subsequent fetal brain maturity assessment based on the trained point cloud recognition network; wherein, during the training process of the point cloud recognition network, only the real gestational age is used as the supervision target, and the real gestational age label is not included in the network feature input to avoid data leakage.

2. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 1, characterized in that, Step S2 specifically includes: Step S2.1: Scale normalize the three-dimensional ultrasound volume data of the cranium; Step S2.2: Manually segment the scale-normalized three-dimensional ultrasound volume data of the cranium to extract the region of interest mesh on the surface of the fetal brain, and perform surface reconstruction, mesh cleaning and hole filling; Step S2.3: Uniformly sample the processed mesh to obtain a surface point cloud containing N sampling points, and calculate the local surface normal and average curvature of each sampling point; Step S2.4: Based on the three-dimensional coordinates, local surface normals, and average curvature of each sampling point, construct the input features corresponding to the sampling point to generate point cloud features, which will then serve as the input for the subsequent point cloud recognition network.

3. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 2, characterized in that, In step S2.4, the expression corresponding to the input feature is: ； in, In three-dimensional coordinates, For local surface normals, The mean curvature.

4. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 3, characterized in that, N is 4096.

5. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 1, characterized in that, In step S3, the pre-feature mapping layer performs a 7→64→64 dimension mapping on the original 4096×7-dimensional point cloud features; The FPS farthest point sampling iteratively selects the point that is farthest from the selected sample in space, so as to finally select a specified number of points to ensure that the sampling points are evenly distributed on the three-dimensional curved surface. The number of key points after downsampling by the Level 4 SA module is 1024, 512, 256 and 128 respectively; The curvature adaptive radius grouping mechanism dynamically adjusts the neighborhood search radius based on the average curvature of the center point itself. The base radius rbase configured for the four-level SA module is 0.12, 0.18, 0.28 and 0.4 respectively, while limiting the maximum number of neighborhood points k of a single center point to 24, 32, 48 and 64 respectively. The curvature attention module uses the curvature contrast between the center point and the neighboring sampling points to modulate the attention weight coefficients, so as to eliminate the prior bias of fixed weight assignment for high curvature features. The four-level SA module outputs multi-scale geometric features of 128, 256, 384 and 512 dimensions in sequence; M is 128; The prediction head Shared MLP has a network structure of 128→18→64. The three-level maturity classification head outputs the discrete maturity category probability distribution of the fetal brain based on 64-dimensional features. The gestational age regression head outputs the continuous gestational age estimate based on 64-dimensional features.

6. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 5, characterized in that, For the j-th center point of the l-th layer, the local grouping radius is calculated using the aforementioned curvature adaptive radius grouping mechanism. The formula is: ； in, The preset base radius, It is the Sigmoid activation function. For learnable adjustment of weight coefficients, Let be the normalized curvature at the center point j.

7. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 5, characterized in that, Using a curvature attention module, the formula for calculating the attention weight coefficient of neighborhood point i within the adaptive neighborhood is as follows: ； in, , The projection matrix is ​​learnable. For feature dimension, This is a hyperparameter for controlling curvature contrast. Let i be a feature of point i in layer l. The features of center point j in layer l, To traverse the features at layer l, In order to be in Points traversed within the inner loop; For the l-th layer, the set of local neighborhood points of the center point j is obtained by adaptive radius grouping. Let i be the average curvature of point i. The average curvature at point j; The average curvature of the point traversal; From attention weight coefficient The updated local features are as follows: ； in, It is a learnable projection matrix.

8. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 1, characterized in that, In step S3, the formula for the learnable gating weights is: ； in, It is a learnable gating weight. It is the global multi-scale feature vector obtained by fusion. It is the layer point feature of the improved output of layer l. It is an MLP specific to layer l.

9. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 1, characterized in that, In step S4, the formula for the output layer consistency loss function is: ； in, Preset maturity category range The midpoint value, To predict successive gestational ages, To predict the probability of discrete maturity categories, C is the number of auxiliary categories, and C is 3.

10. The method for assessing fetal brain maturity based on the fusion of three-dimensional point cloud geometric multi-scale features according to claim 9, characterized in that, In step S5, the formula for the overall optimization objective loss function is: ； in, and These are the true gestational age and the true discrete maturity labels, respectively. For the prediction class, y is the actual maturity level analogy. For Huber's return loss, For weighted cross-entropy loss, It is 0.

35. It is 0.08.

Images