Lateral ventricle broadening prediction method and device, program product and medium

Abstract

The invention belongs to the field of brain image processing, and particularly relates to a lateral ventricle widening prediction method and device, a program product and a medium, and the method comprises the steps: carrying out the three-dimensional segmentation of a magnetic resonance image through a three-dimensional multi-sequence segmentation model, and obtaining a three-dimensional segmentation result; the three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network, and a topology loss function is adopted to train the three-dimensional multi-sequence segmentation model; and inputting the three-dimensional segmentation result into a prediction model to obtain a classification result of lateral ventricle malfunction prognosis. According to the method, three-dimensional full-automatic segmentation of the lateral ventricle is carried out by adopting the nnUNet network, TopoLoss is adopted in the segmentation process, then a classification result of poor prognosis of the lateral ventricle is obtained, the continuity and accuracy of three-dimensional segmentation are improved, and the problem that the overall anatomical form of the lateral ventricle cannot be captured through traditional two-dimensional segmentation is solved.

Description

Technical Field

[0001] The present application relates to the field of brain image processing, and in particular to a method, device, program product and medium for predicting lateral ventriculomegaly. Background Art

[0002] Fetal lateral ventriculomegaly is an important early indicator of abnormal neurodevelopment, and its early detection is of great significance for the prevention and intervention of postnatal neurological diseases. Traditional prenatal examinations rely on ultrasound imaging, which, while real-time and safe, has significant limitations in resolution and deep structural display, especially in the assessment of deep fetal brain structures. Magnetic resonance imaging (MRI) technology, due to its high resolution and multi-sequence characteristics, has shown great advantages in the assessment of fetal brain structure. However, most current studies are still limited to the use of single sequences (such as T2-weighted) and two-dimensional views, which cannot fully capture the three-dimensional morphology and developmental dynamics of the fetal lateral ventricles.

[0003] Existing technologies struggle to capture the complete spatial structure of the lateral ventricles with two-dimensional segmentation models, leading to inaccurate volume and morphological measurements. Single-sequence features limit comprehensive characterization of pathological conditions. In particular, T1 / T2 sequences are widely used for rapid clinical imaging. While FLAIR sequences are sensitive to cerebral edema, they are affected by fetal motion artifacts and are therefore rarely used in practice. Therefore, developing a prediction system that integrates three-dimensional segmentation, multi-sequence fusion, and dynamic time series modeling could not only significantly improve the accuracy of lateral ventriculomegaly detection and classification but also provide important decision-making support for early clinical intervention. Summary of the Invention

[0004] In order to overcome the above-mentioned defects, this application realizes three-dimensional segmentation of brain images based on the nnUNet architecture, which is compatible with a variety of common clinical sequences. It proposes a method, device, program product and medium for predicting lateral ventricle widening based on a three-dimensional nnUNet network to achieve accurate reconstruction of the fetal lateral ventricle and avoid image distortion caused by excessive scanning time.

[0005] Based on the above ideas, a technical solution was proposed: The method for predicting lateral ventriculomegaly based on a three-dimensional nnUNet network includes the following steps: The magnetic resonance image is subjected to three-dimensional segmentation processing by a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; The three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network, and the three-dimensional multi-sequence segmentation model is trained using a topological loss function; The three-dimensional segmentation results are input into a prediction model to obtain a classification result of poor prognosis of the lateral ventricle.

[0006] Furthermore, cross-modal feature alignment is used in the three-dimensional segmentation process to achieve spatial alignment of features of multiple modalities.

[0007] Furthermore, the cross-modal feature alignment is implemented in the encoder of the three-dimensional nnUNet network. The multiple modal features in the encoder are mapped to the same projection space after convolution dimensionality reduction, thereby realizing spatial alignment of the multiple modal features.

[0008] Furthermore, the encoder implements convolution dimensionality reduction through several layers of convolution modules, and the steps adopted by each layer of the convolution module include: convolution processing, normalization processing, ReLU activation and downsampling.

[0009] Furthermore, the cross-modal feature alignment is implemented between two adjacent convolutional modules in the encoder.

[0010] Furthermore, cross-modal feature alignment also includes: using contrast loss to optimize the consistency of features of the same anatomical structure between multiple modalities.

[0011] Furthermore, the use of a topological loss function to train the three-dimensional multi-sequence segmentation model specifically includes: extracting a three-dimensional normal vector field from the boundary of the three-dimensional segmentation result output by the three-dimensional multi-sequence segmentation model, processing the sudden change in the angle between adjacent normal vectors in the three-dimensional normal vector field, and achieving a smooth structural transition of the prediction result.

[0012] Furthermore, the step also includes: resampling the magnetic resonance image in several directions with the same resolution to obtain resampled data, and inputting the resampled data into the three-dimensional nnUNet network to perform three-dimensional segmentation processing.

[0013] Furthermore, it also includes the step of optimizing the time series dynamic modeling: obtaining a dynamic change map based on the evolution trend of the ventricular structure in different time periods; inputting the time series indicators in the dynamic change map and the three-dimensional segmentation results into the prediction model together to obtain a classification result of poor prognosis of the lateral ventricle based on dynamic time series.

[0014] Furthermore, the steps of prognostic risk prediction are also included, including: The classification results of the degree of lateral ventricle enlargement are input into a prognostic risk prediction model to output a prognostic classification level; the prognostic risk prediction model includes at least two classifiers, the classifiers are connected in parallel, and the results output by each classifier are weighted and summed before output.

[0015] Based on the same concept, a 3D nnUNet-based prediction device for lateral ventricle enlargement was also proposed, including: an acquisition module, for acquiring magnetic resonance images of the brain; The segmentation module performs three-dimensional segmentation processing on the magnetic resonance image through a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; the three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network, and a topological loss function is used to train the three-dimensional multi-sequence segmentation model; The prediction module is configured to input the three-dimensional segmentation result into a prediction model to obtain a classification result of the lateral ventricle widening degree.

[0016] Further, a cross-modal feature alignment module is further included, which is embedded in the segmentation module and is configured to align features of multiple modalities in space.

[0017] Further, a time sequence dynamic modeling module is further included, which outputs a dynamic change atlas according to the evolution trend of the ventricle structure at different gestational weeks; a time sequence index in the dynamic change atlas is configured to be input into the prediction model together with the three-dimensional segmentation result to obtain a classification result of the lateral ventricle widening degree based on dynamic time sequence.

[0018] Further, a prognosis risk prediction module is further included, which is configured to input the classification result of the lateral ventricle widening degree into a prognosis risk prediction model to output a prognosis classification level; the prognosis risk prediction model includes at least two classifiers, the classifiers are connected in parallel, and the results output by the classifiers are weighted and summed to output.

[0019] Based on the same concept, a program product for lateral ventricle widening prediction based on a three-dimensional nnUNet network is also proposed, which realizes the lateral ventricle widening prediction method based on the three-dimensional nnUNet network as described in any one of the above when the program product is run on a computer.

[0020] Based on the same concept, a medium having instructions executable by a processor stored thereon is also proposed, which causes the processor to execute the lateral ventricle widening prediction method based on the three-dimensional nnUNet network as described in any one of the above when the instructions are executed by the processor.

[0021] Compared with the prior art, the beneficial effects of the present application are: the present application adopts the nnUNet network for three-dimensional full-automatic segmentation of the lateral ventricle, and adopts the topological loss (TopoLoss) in the segmentation process, thereby obtaining the classification result of the lateral ventricle adverse prognosis, which improves the accuracy of the three-dimensional segmentation structure, and solves the problem that the traditional two-dimensional segmentation cannot capture the overall anatomical morphology of the lateral ventricle. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The flowchart of the lateral ventricle widening prediction method based on the three-dimensional nnUNet network in Embodiment 1 of the present application; Figure 2 This is a model architecture diagram of the three-dimensional multi-sequence segmentation model in Example 1 of the present application; Figure 3 This is a diagram of the 3D full-resolution image segmentation network architecture in Example 2 of this application; Figure 4 This is a flowchart of the prediction process in the prediction model in Example 2 of the present application; Figure 5 A flowchart of the steps for optimizing the dynamic modeling of time series in Example 3 of the present application; Figure 6 This is a structural diagram of the device for predicting lateral ventriculomegaly based on the three-dimensional nnUNet network in Example 5 of the present application; Figure 7 These are the experimental results of risk prediction in Example 5 of this application. DETAILED DESCRIPTION

[0023] The present application is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as limiting the scope of the above-mentioned subject matter of the present application to the following embodiments. All technologies implemented based on the content of the present application fall within the scope of protection of the present application.

[0024] Unless otherwise specified, in the description of the specific embodiments of this application, the terms indicating the orientation or position relationship such as "up", "down", "left", "right", "center", "inside", "outside", and "side" are based on the expression of the orientation or position relationship shown in the accompanying drawings, or the orientation or position relationship in which the product / device / apparatus is placed when it is usually used. These terms of orientation or position relationship are only for the convenience of describing the scheme of this application or simplifying the description in the specific embodiments to facilitate the technicians to quickly understand the scheme, and do not indicate or imply that a specific device / component / element must have a specific orientation, or be constructed and operated in a specific position relationship, and therefore should not be understood as limiting this application.

[0025] In the description of the embodiments of this application, the technical terms "first," "second," etc., merely distinguish one entity or operation from another and are not to be understood as indicating or implying relative importance or implicitly specifying the quantity, specific order, or primary-secondary relationship of the indicated technical features. In the description of the embodiments of this application, "plurality" means two or more, unless otherwise specifically defined.

[0026] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0027] Example 1 The prediction method of lateral ventricle enlargement based on three-dimensional nnUNet network is as follows: Figure 1 As shown, the following steps are included: The magnetic resonance image is subjected to three-dimensional segmentation processing by a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; The three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network, and a topological loss function is used to train the three-dimensional multi-sequence segmentation model. The three-dimensional segmentation results are input into a prediction model to obtain a classification result of poor prognosis of the lateral ventricle.

[0028] From the above steps, we can see that the method mainly includes three parts: image acquisition, three-dimensional segmentation processing and prediction of the degree of lateral ventricle enlargement. The model architecture of the three-dimensional multi-sequence segmentation model is as follows: Figure 2 As shown, the following describes the three main steps: 1. Image Acquisition Acquires magnetic resonance imaging sequences of the brain, giving priority to fast sequence images such as T1 / T2, and is compatible with a variety of other common clinical sequence images to avoid image distortion caused by long scanning time.

[0029] 2. 3D segmentation processing The magnetic resonance sequence images are input into a three-dimensional multi-sequence segmentation model to perform three-dimensional multi-sequence segmentation processing to obtain a three-dimensional segmentation result, which specifically includes the following contents: First, the acquired T1 or T2 sequence images are resampled in the sagittal, coronal, and transverse directions to achieve a uniform isotropic resolution (e.g., 1 mm³). N4 bias field correction (based on the N4 algorithm), z-score normalization (a form of intensity normalization), and brain region ROI cropping are then performed sequentially to generate multimodal data. Isotropic resolution refers to the uniform resolution of medical images in three dimensions (X, Y, and Z axes), meaning that the length, width, and height of the voxels are equal (e.g., 1 mm × 1 mm × 1 mm).

[0030] Next, the acquired multimodal data is fed into a 3D multi-sequence segmentation model, which outputs a 3D segmentation result. This model is based on the nnUNet network architecture, a 3D version of the 3D U-Net. The model consists of a four-layer encoder and a symmetrical decoder. Each layer utilizes a combination of convolution (Conv3D), normalization, ReLU activation, and downsampling (3D Max Pooling).

[0031] Furthermore, the three-dimensional multi-sequence segmentation process further includes: adding a topological loss function to constrain the topological structure of the three-dimensional multi-sequence segmentation model. When training the three-dimensional multi-sequence segmentation model, The (TopologicalLoss, topological loss) function constrains the topological structure of the three-dimensional multi-sequence segmentation model. The three-dimensional normal vector field is extracted from the predicted mask boundary, the sudden change of the angle between adjacent normal vectors is penalized, and the structure is smoothly transitioned to ensure the spatial continuity and integrity of the segmented lateral ventricle area.

[0032] Penalizing sudden changes in the angle between adjacent normal vectors mainly refers to the introduction of the loss function TopoLoss, which numerically penalizes drastic changes in the angle between the normal vectors of adjacent voxels in the segmentation results. That is, during network training, drastic changes in the angle between the normal vectors of adjacent voxels are assigned a larger loss value, driving the model to reduce such changes. Specifically, during training, for voxels on the boundary of the segmentation mask, the normal vector of the 3D boundary is calculated, and for the normal vectors of adjacent voxels, the angle θ between them is calculated. If the angle is large, the loss is large, prompting the network to learn smoother boundaries.

[0033] Unsmoothness means the segmented mask boundary appears jagged, broken, and has sharp corners. Achieving smooth structural transitions specifically means making the boundary continuous and rounded, the normal vectors changing slowly in space, and the normal vectors of adjacent voxels pointing in similar directions. Smooth boundaries more accurately reflect the integrity and continuity of anatomical structures, preventing false positives and false negatives caused by model prediction noise.

[0034] In image segmentation tasks, a mask is a matrix of the same size as the original image, typically a binary image (0 and 1) or a multi-channel label image. Its core function is to accurately annotate the spatial distribution of target objects. One approach involves binary classification tasks: a single-channel binary mask, where 1 represents the foreground (target) and 0 represents the background. Another approach involves multi-class segmentation tasks: a binary mask with each channel corresponding to a class, or color labels used to distinguish different objects (e.g., instance segmentation).

[0035] The formula of the topological loss function is shown in formula (1): (1) in, They represent the normal vectors between adjacent voxels, i represents the number of one type of voxel, j represents the number of another type of voxel, N is the total number of voxels, and voxel refers to a volume element, which is the smallest unit of digital data in three-dimensional space segmentation. Voxel is used in three-dimensional imaging, scientific data, medical imaging and other fields.

[0036] 3. Prediction of the degree of lateral ventriculomegaly The degree of lateral ventricle enlargement was predicted using a prediction model. The model input came from the three-dimensional nnUNet segmentation results, and the output was three categories: normal, mild enlargement, and moderate / severe enlargement.

[0037] The prediction process is as follows Figure 4 Shown, including: Based on the 3D segmentation results, the 3D features and multi-view radiomic features of the lateral ventricle and surrounding white matter are extracted. The extracted 3D features and multi-view radiomic features are fused to form a high-dimensional feature vector. The high-dimensional feature vector is selected to obtain the feature subset most relevant to the degree of lateral ventriculomegaly. This feature subset is input into the prediction model to output the classification result of the degree of lateral ventriculomegaly. The specific implementation process is as follows: 3.1. 3D Feature Extraction and Multimodal Fusion Based on the 3D segmentation results, three-dimensional features (such as expansion directionality and 3D texture distribution) and multi-view radiomic features (such as shape features and grayscale first-order statistical features) of the lateral ventricle and surrounding white matter regions were extracted. Specifically, principal component analysis (PCA) was first used to calculate the principal axes of the 3D shape, estimate the expansion direction and eccentricity, and determine the expansion directionality. Then, gray-level co-occurrence matrix (GLCM) and wavelet decomposition were applied to extract 3D texture distribution features (such as contrast, energy, and uniformity) from each voxel block. Finally, shape features (such as area, perimeter, and circularity) and grayscale first-order statistical features (such as mean, variance, and skewness) of the lateral ventricle were extracted from different viewpoints (sagittal, coronal, and transverse). Weighted averaging or fusion of multiple statistics (P25, P75, maximum, and minimum) were used to combine multiple views.

[0038] The following are some examples of combining multiple views using weighted average or multi-statistics (P25, P75, maximum, minimum) fusion: The grayscale mean feature was extracted for each of the sagittal, coronal, and transverse planes. The grayscale mean values ​​are shown in Table 1: Table 1 Mean grayscale values ​​of sagittal, coronal and transverse planes

[0039] Weighted average fusion assigns a weight to each view: the sagittal plane (S) has a weight of 0.3, the coronal plane (C) has a weight of 0.4, and the transverse plane (A) has a weight of 0.3. The resulting grayscale value F is: F = 0.3 × 100 + 0.4 × 120 + 0.3 × 110 = 30 + 48 + 33 = 111. Multi-statistic fusion uses the grayscale means of the sagittal plane (S), coronal plane (C), and transverse plane (A) as samples and extracts the following statistical features: ① Maximum value (Max) = 120; ② Minimum value (Min) = 100; ③ Quartiles (P25 ≈ 105, P75 ≈ 115) (here simplified as median estimates). The resulting feature vector might be: [Max = 120, Min = 100, P25 = 105, P75 = 115].

[0040] Furthermore, the extracted 3D features were fused with multi-view radiomics features to form a high-dimensional feature vector. A mutual information algorithm was then used to select features and identify the subset of features most relevant to the grade of lateral ventriculomegaly, reducing the model input dimensionality while retaining key information.

[0041] An example of generating a feature subset is as follows: The following 10-dimensional original feature vectors are extracted from the segmentation results. The mutual information algorithm is used to quantify the correlation between each feature and the label. The calculation results are shown in Table 2: Table 2 Calculation results of using mutual information algorithm to quantify the correlation between each feature and label

[0042] A threshold can be set (e.g. greater than 0.1) to ultimately select the feature subsets: F1, F3, F4, F5, F7, and F10. This is the feature subset selection in the dimensionality reduction process.

[0043] Model Prediction A multi-class prediction model for lateral ventriculomegaly was constructed using a variety of classic and advanced machine learning methods. First, the structural feature vectors, filtered by mutual information, were input into the classification network. In the implementation, various classification algorithms, including support vector machines (SVM), random forests, logistic regression, K-nearest neighbors (KNN), and multi-layer perceptrons (MLP), were employed for performance comparison and model integration. During model training, 5-fold cross-validation was used to assess generalization performance, and grid search was used to systematically optimize each model's hyperparameters. Multi-class cross entropy was used as the primary optimization objective to minimize the difference between the predicted class probability distribution and the true label. Furthermore, during model selection, multiple metrics, including accuracy, area under the curve (AUC), recall, and F1 score, were comprehensively considered. The model architecture and configuration that performed best on the validation set were ultimately selected for system integration deployment. AUC, commonly referred to as the "area under the curve" in medicine, is primarily used to assess diagnostic accuracy, model performance, and the pros and cons of different detection methods. AUC values ​​closer to 1 indicate higher diagnostic accuracy and better model performance.

[0044] The advantage of the three-dimensional feature-driven widening degree prediction model is that it extracts the morphological and texture features of the lateral ventricles and their surrounding white matter areas based on the three-dimensional segmentation results, combines multi-view statistics, constructs high-dimensional radiomics feature expression, and accurately predicts the three types of ventricular states: normal, mild widening, and moderate / severe widening through a classification model.

[0045] Example 2 Based on the lateral ventricle widening prediction method based on the three-dimensional nnUNet network in Example 1, in order to establish an alignment relationship between feature maps of different modalities, a cross-modal feature alignment method is further proposed. By comparing the loss function, the feature distributions of different sequences in the same anatomical region are made closer.

[0046] The cross-modal feature alignment method is implemented in the encoder of the nnUNet architecture. The modal feature maps in the encoder of the nnUNet architecture are mapped by convolution dimensionality reduction and uniformly enter the shared projection space to achieve cross-modal feature alignment. The cross-modal feature alignment method is usually implemented in a cross-modal feature alignment module (CM-Align module). The CM-Align module uses contrast loss to bring the feature distributions of different sequences in the same anatomical region closer. For feature vectors extracted from the same anatomical region under different modalities, the contrast loss between the feature vectors is calculated, and the network parameters are optimized through back propagation to make the feature distributions of different modalities closer, thereby achieving effective fusion of cross-modal features. The CM-Align module processing process adopts a three-dimensional full-resolution image segmentation network architecture. The three-dimensional full-resolution image segmentation network architecture is as follows: Figure 3 shown.

[0047] By optimizing the neural network through a contrastive loss function, the distribution of feature vectors extracted from the same anatomical region under different modalities T1 and T2 in the high-dimensional embedding space becomes more similar, that is, the Euclidean distance is smaller. By minimizing the distribution distance of features from different modalities, the feature representations learned by the network are made consistent across modalities, achieving cross-modal consistency and promoting effective fusion.

[0048] The reason for using contrast loss to bring the feature distributions of different sequences in the same anatomical region closer together is explained by the fact that different MRI sequences, such as T1 and T2, have different mechanisms for imaging tissues, resulting in differences in the distribution of image feature expressions for the same structure in different sequences. If features from different modalities are directly concatenated or fused, modal differences may lead to a mix of feature spaces, affecting the model's understanding of the structural nature and fusion efficiency. The CM-Align module is used to maximize the similarity of features from different modalities in the same region. It can suppress irrelevant information specific to the modality and highlight the commonalities of the structure itself, thereby improving the performance of subsequent fusion, classification, or segmentation tasks.

[0049] Specifically, the CM-Align module is embedded between two adjacent convolutional layers of stage3 of the encoder backbone in the nnUnet network structure. Preferably, the CM-Align module is embedded between stage3 (the third convolutional layer) and stage4 (the fourth convolutional layer). Each modality feature map undergoes 1×1×1 convolution dimensionality reduction mapping and is uniformly entered into a shared projection space. Contrast loss is used to optimize the feature consistency of the same anatomical structure between modalities, encouraging T1 sequence images or T2 sequence images to output similar embedding vectors at the same structural position (if the sagittal T2 and transverse T1 are significantly activated in the left ventricular area, the CM-Align module forces the alignment of their feature responses to form a modality-invariant representation).

[0050] The advantage of cross-modal alignment and multi-view fusion is that a cross-modal feature alignment module (CM-Align) is proposed to achieve spatial alignment between multi-modal MRIs such as T1 / T2, integrate multi-directional information of sagittal, coronal and transverse planes, break through the limitations of multiple views but single sequence, and enhance the model's ability to express complex structural features.

[0051] Example 3 Existing predictions for fetal lateral ventricle enlargement ignore the dynamic characteristics of the fetus during rapid development in the second trimester (18-28 weeks), and static models cannot provide reliable prediction support. Therefore, given that the development of the fetal lateral ventricles has obvious temporal characteristics, especially the rapid growth rate and large structural changes in the second trimester (18-28 weeks), a process for optimizing temporal dynamic modeling is further proposed based on the prediction methods in Examples 1 and 2: the evolution trend of the ventricular structure in different time periods is modeled and a dynamic change map is output; the temporal indicators and three-dimensional segmentation results in the dynamic change map are input into the prediction model to obtain an optimized classification result of the degree of lateral ventricle enlargement.

[0052] The steps of optimizing time series dynamic modeling are integrated into the time series dynamic modeling module. The time series dynamic modeling module is integrated with the main prediction network to obtain an optimized fusion prediction model. The output of this model can reflect the changes during pregnancy. The main goal of the time series dynamic modeling module is to use existing or simulated pregnancy scan sequence modeling structure change trends, that is, through the time series modeling module, the morphological evolution trend of the structure during pregnancy is "numerically modeled and embedded expressed", and then used as the input of the prediction model to supplement the limitations of the static prediction model and improve the system's judgment on time-related indicators such as "expansion rate" and "abnormal growth direction". The overall design consists of three parts: adaptive grouping of gestational weeks, time series modeling structure, and generation of dynamic change maps. The steps of optimizing time series dynamic modeling are as follows: Figure 5 shown.

[0053] 1. Gestational age-adaptive grouping: Training samples are grouped into 2-week intervals, such as GW18-20 and GW20-22. Data are grouped based on the developmental characteristics of the fetal lateral ventricles at different gestational ages. Time-segmented sub-models are constructed or time embedding is set to allow the model to focus on the developmental morphology of the ventricles in that gestational age period, thereby improving the model's adaptability to data from different gestational ages.

[0054] 2. Time series modeling structure: Constructing a 3D feature input sequence of the lateral ventricles for consecutive gestational weeks , Indicates Volume of Zhou V , surface area S , curvature C Then use BiLSTM to encode the sequence and output the dynamic feature embedding for each gestational week. , bidirectional LSTM modeling utilizes both past and future time series information to better capture the temporal changes of the lateral ventricles, thereby capturing the structural changes before and after time. The dynamic characteristics of each gestational week are expressed as shown in formula (2): (2) where X represents the dynamic feature, t represents the gestational week, H t represents the dynamic feature coding sequence.

[0055] 3. Dynamic change atlas generation: the LSTM output predicts the structure growth rate, such as , refers to the change amount of lateral ventricle volume in two consecutive gestational weeks, is the interval between two scans. And further extrapolate future structure development trend, such as lateral ventricle expansion rate, morphological evolution trajectory and other time series indicators. The model automatically associates the baseline features of GW20, quantifies the volume change rate within two weeks. Based on bidirectional LSTM, construct the dynamic change atlas in the second trimester, realize the individualized tracking of lateral ventricle development trajectory, and provide dynamic features in time series for the prediction model. The time series modeling results are input into the widening prediction model in the form of features, and the static structure features are jointly input, and the prediction results optimized by dynamic modeling are output. Among them, the mean square error is used to construct the auxiliary loss function, and the auxiliary loss function is represented by formula (3): (3) wherein, is the loss function after adding the time series dynamic modeling module. is the volume change amount predicted by the model, is the volume change amount calculated according to the label.

[0056] The time series indicators in the dynamic change atlas generated by the time series dynamic modeling optimization process are input into the prediction model as auxiliary features, combined with three-dimensional features and multi-view radiomics features, to jointly realize the prediction of fetal lateral ventricle widening state. Through time series dynamic modeling analysis of the dynamic change rule of lateral ventricle, the prediction accuracy is improved, so that the prediction model can more accurately capture the dynamic evolution process of lateral ventricle widening, thereby improving the accuracy and reliability of prediction.

[0057] This embodiment mainly embodies the time series dynamic modeling mechanism. In view of the rapid development characteristics of the fetal brain in the second trimester, the BiLSTM network is introduced to model the evolution trend of the ventricle structure at different gestational weeks, and the individual dynamic change atlas is output as an auxiliary feature to be fused into the classifier, thereby improving the sensitivity and prediction accuracy of the model to developmental trend abnormalities. It has high automation, interpretability and good clinical adaptability, and has practical application value.

[0058] Further, the schemes of embodiments 1, 2 and 3 can be combined together, and the prediction model is combined through structured feature extraction, time series modeling enhancement and multi-model classifier, to realize accurate prediction of fetal lateral ventricle widening grade at different gestational weeks.

[0059] Embodiment 4 Building on the methods of Examples 1-3 above, a prognostic risk prediction method is further proposed. This method aims to further predict the risk of postnatal neurodevelopmental outcomes for the fetus, based on the classification of lateral ventriculomegaly, and assist physicians in developing more appropriate prenatal intervention and management strategies. Analysis of statistical data related to adverse prognosis revealed that even isolated mild ventriculomegaly carries a 7–10% probability of postnatal neurodevelopmental abnormalities; for mild to moderate ventriculomegaly, the probability of concurrent abnormalities approaches 50%; and for severe ventriculomegaly (>15 mm), the combined abnormality rate is as high as 58–65%, and is closely associated with severe structural abnormalities such as abscess of the corpus callosum, Dandy-Walker syndrome, cerebral aqueduct stenosis, and spina bifida. Accurate prognostic prediction is crucial for optimizing intervention timing and guiding postnatal follow-up.

[0060] Based on the aforementioned broadened classification and segmentation results, a comprehensive feature construction strategy based on structure + time features was further introduced. The three-dimensional structural features, developmental trend indicators and classification results were output to a pre-built prognostic prediction model, and the risk contribution ranking of each key predictive factor was output. The prognostic prediction model included an explanatory evaluation part.

[0061] Furthermore, developmental trend indicators include developmental velocity features, radiomic features, and auxiliary labels derived from historical data. Developmental velocity features include the rate of change in ventricular volume, the rate of change in surface area, and the amplitude of temporal fluctuation in asymmetry. Radiomic features include texture complexity, shape deviation, left-right difference index, and periventricular gray matter distribution. Auxiliary labels derived from historical data include gestational age correction, widening classification results, and structural change curve embedding. By introducing developmental trend indicators, we can effectively capture subtle developmental deviations during the progression from mild widening to moderate or severe widening.

[0062] The prognostic prediction model utilizes an ensemble classifier strategy, combining multiple models such as LightGBM, XGBoost, and support vector machines in parallel. A weighted fusion approach is used to generate the final prognostic classification (good, intermediate, poor). The prognostic prediction model is trained based on real postnatal follow-up data, including neurodevelopmental scale scores (such as the Bayley or Gesell scales), for label fitting. This improves the ability to identify mild risk and occult structural abnormalities. If follow-up score labels are unavailable, a supplementary scoring system is constructed using a graphical model based on the co-occurrence of structural abnormalities.

[0063] The advantage of intelligent prediction of postnatal prognostic risk lies in the construction of an individual-level prognostic prediction model based on three-dimensional structural characteristics, developmental trend indicators and classification results, combined with multi-model integration and explanatory evaluation, to output good, medium, poor and other prognostic grades, realizing full-process intelligent prediction from image structure to functional risk, significantly enhancing the clinical implementation capability of this invention.

[0064] Example 5 The device for predicting lateral ventricle enlargement based on the three-dimensional nnUNet network is shown in the figure below. Figure 6 Shown, including: an acquisition module, for acquiring magnetic resonance images of the brain; a segmentation module that performs three-dimensional segmentation processing on the magnetic resonance image using a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; the three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network and is trained using a topological loss function; The prediction module is used to input the three-dimensional segmentation results into the prediction model to obtain the classification results of the poor prognosis of the lateral ventricle.

[0065] Furthermore, the segmentation module includes a cross-modal feature alignment unit CM-Align module, which is embedded between stage3 and stage4 of the nnUnet encoder backbone. The CM-Align module uses contrastive loss to bring the feature distributions of different sequences in the same anatomical region closer. Specifically, for the feature vectors extracted from the same anatomical region under different modalities, the contrastive loss between them is calculated, and the network parameters are optimized through back-propagation to make the feature distributions of different modalities closer, thereby achieving effective fusion of cross-modal features.

[0066] Preferably, the prediction device also includes a temporal dynamic modeling module. Given the distinct temporal characteristics of fetal lateral ventricle development, particularly during the second trimester (18-28 weeks), when growth is rapid and structural changes are significant, a specially constructed temporal dynamic modeling module is integrated with the main prediction network. Gestational development trend information from the dynamic modeling module serves as an additional input feature, combined with static structural features for prediction. By introducing temporal trends, the prediction model improves sensitivity to the dynamic evolution of the ventricles, enabling it to not only reflect structural status but also capture the rate of structural development and the trajectory of morphological evolution.

[0067] Furthermore, it also includes a cross-modal feature alignment module, which is embedded in the segmentation module to achieve spatial alignment of features of multiple modalities.

[0068] Preferably, the system also includes a prognostic risk prediction module, which integrates interpretability technologies such as SHAP to output a risk contribution ranking of each key predictor. It also supports the generation of prognostic probability distributions, structural change trajectory maps, and personalized prognostic assessment reports, facilitating physicians' judgment and follow-up recommendations based on the results. Overall, this module completes the entire closed loop from anatomical quantification to functional prognosis, enhancing the clinical practical value and translational capabilities of the present invention.

[0069] The method and device of the present invention combine three-dimensional structural modeling, cross-modal alignment and dynamic learning of gestational age, which are significantly superior to traditional two-dimensional segmentation and static judgment methods in the early identification of lateral ventriculomegaly. They show higher sensitivity and discrimination for mild and potential widening cases during the 18-22 weeks of pregnancy, and can provide more forward-looking diagnostic basis for clinicians. Its high-precision segmentation capability, structural continuity optimization and developmental trend prediction demonstrate its leading potential in the field of AI-assisted diagnosis of prenatal imaging. At the same time, the present invention proposes a prognostic risk prediction module, and the experimental results of risk prediction are as follows: Figure 7 As shown, based on structural typing, the possibility of neurodevelopmental abnormalities can be further evaluated, and fetal individuals with a higher risk of structural abnormalities or developmental delays after birth can be identified, providing a decision-making basis for early precision intervention, postpartum management and family counseling, which has good medical translation and clinical practice value.

[0070] The beneficial effects of the present invention include: Three-dimensional segmentation and multi-sequence integration: This paper uses the nnUNet network for fully automatic three-dimensional segmentation of the fetal lateral ventricles, combining it with a topological loss (TopoLoss) to improve the continuity and accuracy of the three-dimensional structure, addressing the problem that traditional two-dimensional segmentation cannot capture the overall anatomical morphology of the lateral ventricles. Cross-modal alignment and multi-view fusion: A cross-modal feature alignment module (CM-Align) is proposed to achieve spatial alignment between multi-modal MRIs such as T1 / T2, integrating multi-directional information from sagittal, coronal, and transverse planes, breaking through the limitations of multiple views but a single sequence and enhancing the model's ability to represent complex structural features. A three-dimensional feature-driven widening prediction model: Based on the three-dimensional segmentation results, the morphological and texture features of the lateral ventricles and their surrounding white matter regions are extracted. Combined with multi-view statistics, a high-dimensional radiomics feature expression is constructed. A classification model is then used to accurately predict the three types of ventricular states: normal, mildly widened, and moderately / severely widened. Temporal Dynamic Modeling Mechanism: Targeting the rapid fetal brain development during the second trimester, a BiLSTM network is introduced to model the evolutionary trends of ventricular structure at different gestational ages. This outputs individual dynamic change maps, which are then integrated into the classifier as auxiliary features. This improves the model's sensitivity and prediction accuracy for developmental trend abnormalities. The model is highly automated, interpretable, and clinically applicable, possessing practical translational application value. Intelligent Prediction of Postnatal Prognosis Risk: An individual-level prognostic prediction model is constructed based on three-dimensional structural features, developmental trend indicators, and classification results. Combined with multi-model integration and interpretable evaluation, it outputs prognostic grades such as good, moderate, and poor, enabling intelligent prediction of the entire process from image structure to functional risk, significantly enhancing the clinical application capabilities of this invention. The above-described embodiments are intended only to illustrate the technical solutions of this application and are not intended to limit them. Although this application has been described in detail with reference to the aforementioned embodiments, those skilled in the art will appreciate that modifications to the technical solutions described in the aforementioned embodiments may be made, or some of the technical features therein may be replaced by equivalents. Such modifications or replacements do not deviate from the spirit and scope of the technical solutions of the various embodiments of this application and are therefore intended to be included within the scope of protection of this application.

Claims

1. A method for predicting lateral ventriculomegaly, characterized in that: The following steps are involved: The magnetic resonance image is subjected to three-dimensional segmentation processing by a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; The three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network, and the three-dimensional multi-sequence segmentation model is trained using a topological loss function; The three-dimensional segmentation results are input into a prediction model to obtain a classification result of poor prognosis of the lateral ventricle.

2. The method for predicting lateral ventriculomegaly according to claim 1, wherein: Cross-modal feature alignment is used in the three-dimensional segmentation process to achieve spatial alignment of features of multiple modalities.

3. The method for predicting lateral ventriculomegaly according to claim 2, wherein: The cross-modal feature alignment is implemented in the encoder of the three-dimensional nnUNet network. The multiple modal features in the encoder are mapped to the same projection space after convolution dimensionality reduction, thereby realizing spatial alignment of the multiple modal features.

4. A method for predicting lateral ventriculomegaly according to claim 3, characterized in that: The encoder implements convolution dimensionality reduction through several layers of convolution modules, and the steps adopted by each layer of the convolution module include: convolution processing, normalization processing, ReLU activation and downsampling.

5. The method for predicting lateral ventriculomegaly according to claim 3, wherein: The cross-modal feature alignment is implemented between two adjacent convolutional modules in the encoder.

6. The method for predicting lateral ventriculomegaly according to claim 2, wherein: Cross-modal feature alignment also includes: using contrast loss to optimize the consistency of features of the same anatomical structure between multiple modalities.

7. The method for predicting lateral ventriculomegaly according to claim 1, wherein: The use of a topological loss function to train the three-dimensional multi-sequence segmentation model specifically includes: extracting a three-dimensional normal vector field from the boundary of the three-dimensional segmentation result output by the three-dimensional multi-sequence segmentation model, processing sudden changes in the angles between adjacent normal vectors in the three-dimensional normal vector field, and achieving a smooth structural transition of the prediction results.

8. The method for predicting lateral ventriculomegaly according to any one of claims 1 to 7, wherein: The steps also include: resampling the magnetic resonance image in several directions with the same resolution to obtain resampled data, and inputting the resampled data into the three-dimensional nnUNet network to perform three-dimensional segmentation processing.

9. The method for predicting lateral ventriculomegaly according to any one of claims 1 to 7, wherein: It also includes the steps of optimizing the time series dynamic modeling: obtaining a dynamic change map based on the evolution trend of the ventricular structure in different time periods; inputting the time series indicators in the dynamic change map and the three-dimensional segmentation results into the prediction model together to obtain a classification result of poor prognosis of the lateral ventricle based on dynamic time series.

10. The method for predicting lateral ventriculomegaly according to any one of claims 1 to 7, characterized in that: It also includes steps for prognostic risk prediction, including: The classification results of the degree of lateral ventricle enlargement are input into a prognostic risk prediction model to output a prognostic classification level; the prognostic risk prediction model includes at least two classifiers, the classifiers are connected in parallel, and the results output by each classifier are weighted and summed before output.

11. A device for predicting lateral ventriculomegaly, characterized in that: include: an acquisition module, for acquiring magnetic resonance images of the brain; a segmentation module that performs three-dimensional segmentation processing on the magnetic resonance image using a three-dimensional multi-sequence segmentation model to obtain a three-dimensional segmentation result; the three-dimensional multi-sequence segmentation model is constructed based on a three-dimensional nnUNet network and is trained using a topological loss function; The prediction module is used to input the three-dimensional segmentation results into the prediction model to obtain the classification results of the poor prognosis of the lateral ventricle.

12. The device for predicting lateral ventriculomegaly according to claim 11, wherein: It also includes a cross-modal feature alignment module, which is embedded in the segmentation module and is used to achieve spatial alignment of features of multiple modalities.

13. The device for predicting lateral ventriculomegaly according to claim 11, wherein: It also includes a time series dynamic modeling module, which outputs a dynamic change map based on the evolution trend of the ventricular structure at different gestational weeks; the time series indicators in the dynamic change map are used to input into the prediction model together with the three-dimensional segmentation results to obtain a classification result of poor prognosis of the lateral ventricle based on dynamic time series.

14. The device for predicting lateral ventriculomegaly according to claim 11, wherein: It also includes a prognostic risk prediction module, which is used to input the classification results of the degree of lateral ventricle widening into a prognostic risk prediction model and output a prognostic classification level; the prognostic risk prediction model includes at least two classifiers, the classifiers are connected in parallel, and the results output by each classifier are weighted and summed before output.

15. A program product for predicting lateral ventriculomegaly, characterized in that: When the program product is run on a computer, it implements the method for predicting lateral ventriculomegaly based on a three-dimensional nnUNet network as described in any one of claims 1 to 10.

16. A medium, characterized in that Instructions executable by a processor are stored thereon, and when the instructions are executed by the processor, the processor executes a method for predicting lateral ventriculomegaly according to any one of claims 1 to 10.

Images