Lymphatic vessel segmentation method based on dynamic integrated structure perception 3D U-Net

Abstract

The application relates to the technical field of medical image processing, in particular to a lymphatic vessel segmentation method based on dynamic integrated structure perception 3D-Net, which comprises two parts of a training stage and a use stage. The training stage comprises construction, training and optimization of multiple 3D-UNet models, wherein each 3D-UNet model adopts different initialization and hyperparameter settings, so that the models have diversity and difference. The use stage comprises integrating the prediction results of the multiple 3D-UNet models, and improving the segmentation accuracy and robustness through a dynamic weight integration strategy. Finally, post-processing technology is adopted to optimize the integrated results, so that the accuracy and stability of the segmentation results are ensured. Through dynamic integration of multiple 3D-UNet models, the method combines a structure perception module and a multi-scale feature fusion strategy, significantly improves the accuracy and robustness of lymphatic vessel segmentation, and provides more reliable image data support for clinical diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a lymphatic vessel segmentation method based on Dynamic Integrated Structure-Aware 3DU-Net. Background Technology

[0002] Lymphatic vessels are part of the human lymphatic system, primarily responsible for transporting lymph fluid and participating in immune responses. They play a vital role in maintaining fluid homeostasis and transporting fats and metabolic waste. Distributed throughout the body, lymphatic vessels are particularly interconnected with other immune tissues via lymph nodes, forming a complex network. The health and function of lymphatic vessels directly affect the human immune system and play a significant role in the occurrence and development of various diseases.

[0003] Precise segmentation of lymphatic vessels is crucial in medical imaging, particularly in the diagnosis and treatment of cancer, lymphedema, and other diseases related to the lymphatic system. Traditional lymphatic vessel segmentation methods often rely on physician experience, resulting in low efficiency and a high risk of error. In recent years, advancements in medical imaging technology, especially the application of deep learning and artificial intelligence, have made automated lymphatic vessel segmentation possible.

[0004] With the rapid development of computer vision technology, especially the application of deep learning algorithms, the accuracy of lymphatic vessel segmentation has been continuously improving. Lymphatic vessel segmentation methods based on deep learning models such as 3D-UNet can efficiently and automatically process large amounts of medical image data and provide reliable segmentation results. However, although existing deep learning technologies (such as 3D-UNet) can improve segmentation accuracy, they still suffer from insufficient robustness due to the limitation of model homogeneity.

[0005] Therefore, a new technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of the prior art and provide a lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net, so as to solve the technical problem of insufficient robustness caused by the model singularity of the existing deep learning-based lymphatic vessel segmentation model.

[0007] The above objectives are achieved through the following technical solutions:

[0008] A lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net includes a training phase and an application phase, wherein:

[0009] Training phase

[0010] Multiple pre-training 3D-UNet models with different initialization and hyperparameter settings are constructed, and each pre-training 3D-UNet model adopts a different initialization strategy;

[0011] Preprocess the MRI image data to generate a training dataset;

[0012] Based on the training dataset, each of the pre-training 3D-UNet models is trained using a structure-aware module and a multi-scale feature fusion strategy to obtain the post-training 3D-UNet model.

[0013] The outputs of multiple post-3D-UNet models are fused using a dynamic ensemble learning strategy to obtain an ensemble segmentation result; (Usage phase)

[0014] The MRI image data to be segmented is preprocessed to generate an image dataset to be processed;

[0015] Multiple trained 3D-UNet models are used to predict the image dataset to be processed, and the corresponding prediction results are obtained.

[0016] The prediction results are fused using a dynamic weight fusion strategy to obtain the final segmentation result;

[0017] Morphological post-processing is performed on the final segmentation results to optimize the accuracy and stability of the segmentation results.

[0018] Furthermore, the initialization strategy for the 3D-UNet model before training includes at least one combination of the following:

[0019] Convolutional layer initialization: Xavier normal distribution, Kaiming orthogonal initialization, orthogonal initialization, sparse initialization, or prediction-level weight transfer;

[0020] Normalization layer initialization: uniform He distribution, zero-mean Gaussian distribution, fixed scaling factor, dynamic range scaling, or freezing the first 3 layers;

[0021] Learning rate strategies: cosine annealing, step descent, adaptive AdamW, cyclic learning rate, or linear warmup.

[0022] Furthermore, the pre-training 3D-UNet model comprises five models, including:

[0023] 3D-UNet model before training 1:

[0024] The convolutional layers are initialized using a Xavier normal distribution;

[0025] The normalization layer is initialized using a uniform He distribution;

[0026] The learning rate employs a cosine annealing strategy;

[0027] 3D-UNet model before training 2:

[0028] Convolutional layers undergo Kaiming orthogonal initialization;

[0029] The normalized layer parameters are initialized to a Gaussian distribution (σ = 0.1);

[0030] The learning rate decreases in a stepwise manner;

[0031] 3D-UNet model before training:

[0032] The convolutional layers use strictly orthogonal initialization, satisfying W T The constraint condition W = I;

[0033] The normalization layer scaling factor is initialized to a constant γ = 0.8;

[0034] The optimizer uses adaptive AdamW;

[0035] 3D-UNet model before training 4:

[0036] The convolutional layer is sparsely initialized, with 50% of the weights set to 0.

[0037] The normalization layer is initialized using dynamic range scaling;

[0038] The learning rate changes cyclically according to a triangular period.

[0039] 3D-UNet model before training 5:

[0040] The convolutional layers employ prediction-level weight transfer, inheriting convolutional kernel parameters from a pre-trained liver segmentation model;

[0041] Normalize the γ and β parameters of the three layers before freezing;

[0042] Learning rate linear warmup: (The first 5 epochs);

[0043] Furthermore, the preprocessing methods for both the preprocessing of the MRI image data and the preprocessing of the MRI image data to be segmented are the same, including:

[0044] Rigid registration of MRI images was performed, and alignment was achieved using a 6-DOF affine transformation.

[0045] Isotropic resampling based on B-spline interpolation was performed on MRI images of different sizes, with a unified network input size of 512×512×32 and a voxel spacing of 1mm×1mm×3mm.

[0046] Dynamic data augmentation strategies are employed, including adaptive window width and level adjustment, spatial transformation, and noise injection.

[0047] Furthermore, the structure-aware module includes edge-enhancing Dice loss and topology-preserving loss, with the loss function being:

[0048]

[0049] in To enhance the Dice loss at the edges, The topology-preserving loss function is expressed as follows:

[0050]

[0051] Where y edge Extracted using the Canny operator (threshold 0.2, σ 1.0);

[0052]

[0053] Where χ() is the Euler number characteristic, used to constrain the tubular topology of lymphatic vessels; P = 8 neighborhood connectivity.

[0054] Furthermore, the dynamic ensemble learning strategy includes:

[0055] The weights of each model are calculated based on the Dice scores on the validation set, and the prediction results of multiple models are combined using a weighted average, as shown in the following formula:

[0056]

[0057] The weights are calculated as follows:

[0058]

[0059] Among them, Dice m Let be the Dice score of the m-th model on the validation set.

[0060] Furthermore, the morphological post-processing includes:

[0061] Anisotropic closing operations are performed using a flattened circular structuring element with parameter I. post =I·S 3×3×1 ;

[0062] The connected component filtering criteria remove fragments and retain tubular structures. The retention criteria are:

[0063] V region >8voxels and

[0064] Further, it is characterized by comprising:

[0065] The data preprocessing module is used to preprocess the input MRI image data and generate a standardized three-dimensional matrix;

[0066] The data augmentation module is used to improve the model's robustness to differences in scan parameters;

[0067] The multi-model dynamic integrated inference module includes multiple 3D-UNet models for predicting MRI image data and outputting the corresponding prediction results.

[0068] The dynamic fusion module is used to perform weighted fusion of the output of the 3D-UNet model to obtain the final segmentation result;

[0069] The post-processing module is used to perform morphological post-processing on the final segmentation results to optimize the accuracy and stability of the segmentation results.

[0070] The lymphatic vessel segmentation method based on dynamically integrated structure-aware 3DU-Net provided by this invention significantly improves the accuracy and robustness of lymphatic vessel segmentation by dynamically integrating multiple 3D-UNet models and combining structure-aware modules and multi-scale feature fusion strategies, providing more reliable image data support for clinical diagnosis and treatment. Attached Figure Description

[0071] Figure 1 This is a flowchart of the lymphatic vessel segmentation method based on the dynamically integrated structure-aware 3DU-Net described in this invention;

[0072] Figure 2 This is a block diagram of the lymphatic vessel segmentation system based on the dynamically integrated structure-aware 3DU-Net described in this invention;

[0073] Figure 3 This is a structural diagram of the 3D-UNet model in the lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net described in this invention.

[0074] Figure 4 This is a training curve of the 3D-UNet dynamic ensemble model in the lymphatic vessel segmentation method based on dynamic ensemble structure perception 3DU-Net described in this invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. The described embodiments are merely some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] like Figure 1As shown, this scheme provides a lymphatic vessel segmentation method based on dynamically integrated structure-aware 3D-UNet, including a training phase and an application phase. The training phase includes the construction, training, and optimization of multiple 3D-UNet models, each with different initialization and hyperparameter settings to ensure model diversity and variability. The application phase involves integrating the prediction results of multiple 3D-UNet models, using a dynamic weight ensemble strategy to improve segmentation accuracy and robustness. Finally, post-processing techniques are used to optimize the ensemble results, ensuring the accuracy and stability of the segmentation results. Specifically:

[0077] During the training phase

[0078] Multiple pre-training 3D-UNet models with different initialization and hyperparameter settings are constructed, and each pre-training 3D-UNet model adopts a different initialization strategy;

[0079] Preprocess the MRI image data to generate a training dataset;

[0080] Based on the training dataset, the pre-training 3D-UNet models are trained using a structure-aware module and a multi-scale feature fusion strategy to obtain the post-training 3D-UNet models. Specifically, in this step, the structure-aware module uses edge enhancement Dice loss. and topology preservation loss (L Topo This process ensures that the model focuses not only on overall segmentation accuracy during training but also on learning lymphatic vessel boundaries (such as fine branches) and tubular topology (such as connectivity). Multi-scale feature fusion utilizes the multi-level convolutional structure of 3D-UNet to simultaneously extract local details (shallow features) and global context (deep features), addressing the issue of large differences in lymphatic vessel size (such as the difference between the main trunk and fine branches) in MRI images. Through this training step, each trained 3D-UNet model can independently capture different features of lymphatic vessels (e.g., Model 3 is sensitive to small-scale structures, while Model 4 is robust to noise). Due to differences in initialization, different models may produce different error patterns on the same image after training (e.g., Model 1 has more accurate boundary segmentation, while Model 5 is more robust to low-contrast regions), providing a diverse foundation for subsequent integration.

[0081] The outputs of multiple post-3D-UNet models are fused using a dynamic ensemble learning strategy to obtain ensemble segmentation results. This provides an optimization foundation and performance verification basis for dynamic weight ensemble in the subsequent usage phase, ensuring that the dynamic weight ensemble and post-processing in the usage phase can stably output segmentation results that meet clinical needs. This process embodies the closed-loop design concept of "training verification - feedback optimization". Specific functions include:

[0082] Reduce single-model bias: through weighted averaging By integrating the prediction results of multiple models, errors caused by randomness in model initialization or training can be reduced.

[0083] Adaptive weight allocation: based on the Dice score on the validation set. The model weights are dynamically adjusted so that the better-performing model contributes more to the final result (as shown in Table 3, the ensemble Dice score of -0.67156 is better than that of the single model).

[0084] The uses of segmentation results include:

[0085] Clinical diagnostic aids: Provides accurate three-dimensional segmentation results of lymphatic vessels to help doctors locate lesions (such as lymphedema or tumor metastasis pathways).

[0086] Quantitative analysis: used to measure the morphological parameters of lymphatic vessels (such as diameter and number of branches) to support disease progression assessment or surgical planning.

[0087] Research tools: Provide reliable data support for the study of the mechanisms of lymphatic system-related diseases (such as lymph node metastasis in cancer).

[0088] In brain MRI, ensemble models can avoid misidentifying vascular artifacts as lymphatic vessels by single models (by constraining the topology through a structure-aware module), while preserving small lymphatic vessels (through multi-scale feature fusion). The final output can be used to assess the severity of the condition in patients with lymphatic drainage disorders.

[0089] In this embodiment, the ensemble segmentation results generated through dynamic ensemble during the training phase essentially establish an "expert decision-making system," specifically:

[0090] During the training phase, the system learns the weight allocation rules of each sub-model (3D-UNet1-5) through validation set data (such as the private brain MRI dataset of the Affiliated Hospital of Jiangnan University);

[0091] During the usage phase, the system directly applies the weight knowledge learned during training to new data without recalculating the weights. As shown in Table 3, the integrated Dice score (-0.67156) is better than any sub-model, and this optimization capability has been fully transferred to the application scenario.

[0092] Furthermore, the ensemble results from the training phase (as shown in Table 3) and the output from the usage phase share the same fusion algorithm, ensuring that:

[0093] Preprocessing consistency: The B-spline interpolation and window width and level adjustments from the training phase are used in the usage phase;

[0094] Post-processing consistency: all processes employ anisotropic closing operations for morphological optimization.

[0095] This design allows the accuracy validated during the training phase (such as the Dice score) to be directly extrapolated to clinical applications.

[0096] When processing pediatric brain MRI images (where lymphatic vessels are smaller), Model 3 (with a fixed scaling factor γ = 0.8) is known to be sensitive to fine structures during the training phase, and the system automatically increases its weight. However, when processing images of elderly patients (where there are more artifacts), the noise suppression characteristics of Model 4 (with sparse initialization) receive higher weight. This dynamic adjustment capability is entirely inherited from the integration rules established during the training phase.

[0097] During the usage phase

[0098] The MRI image data to be segmented is preprocessed to generate an image dataset to be processed;

[0099] Multiple trained 3D-UNet models are used to predict the image dataset to be processed, and the corresponding prediction results are obtained.

[0100] The prediction results are fused using a dynamic weight fusion strategy to obtain the final segmentation result;

[0101] Morphological post-processing is performed on the final segmentation results to optimize the accuracy and stability of the segmentation results.

[0102] This embodiment includes the following in the training phase:

[0103] Initial 3D-UNet model construction stage: Using medical imaging principles and deep learning technology, multiple 3D-UNet models with different initialization and hyperparameter settings were constructed by preprocessing and enhancing MRI data.

[0104] Specifically, a differentiated model initialization strategy was adopted, with the five sub-models using the initialization combinations shown in Table 1. Each model processed different lymphatic vessel image features to ensure that the network had a good learning ability for the diversity of data, thereby constructing the initial 3D-UNet model;

[0105] Integrated model training phase: By training multiple 3D-UNet models with different initializations and hyperparameters, the model parameters are optimized using metrics such as structure-aware loss to ensure that each model performs well in the lymphatic vessel segmentation task and effectively improves the segmentation accuracy;

[0106] Mean ensemble strategy stage: The prediction results of multiple trained 3D-UNet models are fused using the mean ensemble method to obtain the final ensemble segmentation result.

[0107] As shown in Table 1, this is a model initialization strategy table. The initialization strategy of the 3D-UNet model before training in this embodiment includes at least one combination of the following:

[0108] Convolutional layer initialization: Xavier normal distribution, Kaiming orthogonal initialization, orthogonal initialization, sparse initialization, or prediction-level weight transfer;

[0109] Normalization layer initialization: uniform He distribution, zero-mean Gaussian distribution, fixed scaling factor, dynamic range scaling, or freezing the first 3 layers;

[0110] Learning rate strategies: cosine annealing, step descent, adaptive AdamW, cyclic learning rate, or linear warmup.

[0111] Table 1 Model Initialization Strategy

[0112] Model Number Convolutional layer initialization Normalization layer initialization Learning rate strategy 1 Xavier normality He is uniformly distributed Cosine annealing 2 Kaiming orthogonal Zero-mean Gaussian (σ = 0.1) descending stairs 3 Orthogonal initialization Fixed scaling factor (γ = 0.8) Adaptive AdamW 4 Sparse initialization Dynamic range scaling Loop learning rate 5 Pre-trained weight transfer 3 layers before freezing linear warmup

[0113] As shown in Table 1, there are 5 3D-UNet models before training, including:

[0114] 3D-UNet model before training 1:

[0115] The convolutional layers are initialized using a Xavier normal distribution, mathematically expressed as follows: Where n in and n out These represent the number of input and output channels, respectively.

[0116] The normalization layer is initialized using a uniform He distribution and follows the rules of normalization. distributed;

[0117] The learning rate uses a cosine annealing strategy: Where η max =1e -3 η min =1e -5 ;

[0118] Technical effect: It maintains stable variance of activation values ​​in each layer, making it suitable for treating standard lymphatic nodes;

[0119] 3D-UNet model before training 2:

[0120] The convolutional layer implements Kaiming orthogonal initialization, and generates an orthogonal matrix W = QR through QR decomposition, where Q is an orthogonal basis;

[0121] The normalized layer parameters are initialized to a Gaussian distribution (σ = 0.1);

[0122] The learning rate decreases in a stepwise manner: multiply by a decay factor of 0.5 every 20 epochs;

[0123] Technical effect: Orthogonality constraints make gradient update directions more stable, which is suitable for long-range lymphatic vessel tracking;

[0124] 3D-UNet model before training:

[0125] The convolutional layers use strictly orthogonal initialization, satisfying W T The constraint condition W = I;

[0126] The normalization layer scaling factor is initialized to a constant γ = 0.8;

[0127] The optimizer uses adaptive AdamW, satisfying AdamW(β1=0.9, β2=0.99);

[0128] Technical effect: Forced low activation intensity is suitable for micro-lymphatic vessel segmentation;

[0129] 3D-UNet model before training 4:

[0130] The convolutional layer is sparsely initialized, with 50% of the weights set to 0.

[0131] The normalization layer is initialized using dynamic range scaling;

[0132] The learning rate changes cyclically according to a triangular period: Period T = 100 steps;

[0133] Technical effect: It simulates the sparse connectivity characteristics of biological neurons, enhancing the recognition of broken lymphatic vessels;

[0134] 3D-UNet model before training 5:

[0135] The convolutional layers employ prediction-level weight transfer, inheriting the convolutional kernel parameters W from a pre-trained liver segmentation model. tpretrained ;

[0136] Normalize the γ and β parameters of the three layers before freezing;

[0137] Learning rate linear warmup: (The first 5 epochs);

[0138] Technical effect: Transfer learning significantly improves generalization ability under small sample sizes.

[0139] This design enables integrated systems to possess:

[0140] Feature diversity: Each model focuses on lymphatic vessel features at different scales;

[0141] Error decorrelation: Misclassification by a single model can be corrected by other models;

[0142] Scene adaptability: Different initialization combinations can address various imaging quality issues.

[0143] Experimental data (Table 2) shows that the Dice scores of the five models on the test set ranged from -0.6570 to -0.6769, verifying the effectiveness of the initialization strategy. After dynamic weight integration (Table 3), the final Dice score improved to -0.67156, proving that the technical solution achieved an integration effect of 1+1>2.

[0144] This embodiment includes the following usage phase:

[0145] Acquire MRI image data to be segmented, and generate initial features of the image to be processed using the same preprocessing techniques as in the training phase① (such as denoising, standardization, etc.);

[0146] Next, multiple trained 3D-UNet models were used to predict the MRI image, generating segmentation results from multiple models.

[0147] Then, the prediction results of each model are fused using the mean ensemble strategy to obtain the final lymphatic vessel segmentation result;

[0148] Finally, morphological post-processing techniques are used to further optimize the segmentation results, remove artifacts, and enhance the boundary accuracy of lymphatic vessels, ensuring the accuracy and stability of the final output segmentation results.

[0149] MRI images are digital images generated by scanning human tissue and contain a large amount of pixel information. Deep convolutional neural networks cannot directly process raw image data, so MRI images need to be preprocessed to convert them into a numerical form that the network can accept.

[0150] In this embodiment, the preprocessing methods for MRI image data and the preprocessing of MRI image data to be segmented are the same, including:

[0151] Rigid registration of MRI images was performed using 6-DOF affine transformation alignment;

[0152] Isotropic resampling based on B-spline interpolation was performed on MRI images of different sizes, with a unified network input size of 512×512×32 and a voxel spacing of 1mm×1mm×3mm.

[0153] Dynamic data augmentation strategies are employed, including adaptive adjustment of window width and window level, spatial transformation (such as Z-axis sampling and random flipping), and noise injection (such as Gaussian noise).

[0154] Specifically, the original DCM data is first synthesized into a patient-level MRI dataset. The MRI images are then rigidly registered (using a 6-DOF affine transformation) and aligned at the data center level to ensure that the datasets in all centers are oriented in the same direction.

[0155] Next, for MRI images of different sizes, isotropic resampling based on B-spline interpolation was performed to unify them to a network input size of 512×512×32 (voxel spacing 1mm×1mm×3mm). Specifically, third-order B-spline interpolation was performed on the z-axis:

[0156]

[0157] in, The resolution error of the interpolated image is controlled within 0.5mm.

[0158] To improve the model's generalization ability, a dynamic data augmentation strategy is adopted. Adaptive window width and level adjustment: W = μ + 3σ, L = μ - σ. Where μ is the image mean, σ is the standard deviation, covering 99.7% of the soft tissue signal range; Spatial transformation: z-axis resampling (±15% jitter), random rotation (range ±25°), and flipping (probability 50%). Noise injection: Gaussian noise (σ∈[0,0.1]) and Rayleigh noise (intensity ratio 1:3) are added to simulate MRI scan artifacts.

[0159] While the aforementioned MRI image preprocessing methods can help extract some basic features, the sheer volume of information contained in images means that a single preprocessing method cannot fully uncover all valuable features. Therefore, employing more complex feature extraction techniques can further enhance the model's expressiveness. Compared to traditional two-dimensional features, three-dimensional features can more comprehensively capture the spatial information of images, thereby improving segmentation performance.

[0160] MRI image data contains rich spatial structural information, which is particularly important for lymphatic vessel segmentation. Therefore, this method employs a 3D convolutional neural network (3D-UNet) to extract features from MRI images, fully utilizing the three-dimensional structure of the images. Unlike conventional two-dimensional convolutional networks, 3D convolution can directly process the depth, width, and height of the image, thus capturing spatial relationships more accurately. Through 3D convolutional layers, the network can extract depth spatial features, which are more expressive for lymphatic vessel segmentation tasks.

[0161] Due to the differences in resolution and target region size between MRI images, single-scale feature extraction may not be able to cover all details. Therefore, this method introduces a multi-scale feature fusion strategy in 3D-UNet. Through convolutional operations at different scales, the network can extract detailed information from the image at different levels and fuse this information together to further improve segmentation accuracy. Multi-scale feature fusion can effectively capture lymphatic vessels of different sizes and shapes, avoiding segmentation errors caused by scale inconsistencies.

[0162] In addition to spatial information and multi-scale features, the morphological structure of lymphatic vessels is equally important. Therefore, this method also performs post-processing on the output of the 3D convolutional network through morphological operations (such as dilation and erosion) to further enhance the morphological information of lymphatic vessels. These operations help remove artifacts, strengthen the boundaries of lymphatic vessels, and ensure more accurate segmentation results.

[0163] The morphological post-processing includes:

[0164] Anisotropic closing operations are performed using a flattened circular structuring element with parameter I. post =I·S 3×3×1 ;

[0165] The connected component filtering criteria remove fragments and retain tubular structures. The retention criteria are: retain tubular structures with a volume greater than 8 pixels and a table body ratio less than 2.7.

[0166] V region >8voxels and

[0167] Specifically, using anisotropic closing operation with parameter I post =I·S 3×3×1 Because its axial continuity is stronger than that of the coronal plane, a flattened circular structural element S is used. 3×3×1 Simultaneously, a connected component filtering condition is used, where the retention condition is: V region >8voxels and

[0168] This allows for the removal of debris while preserving tubular structures (volume threshold of 8 voxels, surface-to-volume ratio threshold of 2.7).

[0169] By combining the advantages of 3D convolution and multi-scale feature extraction, this method can effectively improve the segmentation accuracy of lymphatic vessels in MRI images and provide more accurate image data for subsequent medical analysis.

[0170] As a specific embodiment of this solution, it includes:

[0171] Step 1: Use the preprocessing matrix of the original MRI image data as the initial features, and perform operations such as denoising and standardization so that the network can accept and effectively extract image features.

[0172] Step 2: Using the principles of medical imaging and deep learning technology, the processed MRI image data is input into multiple 3D-UNet models, and lymphatic vessel features in the images are extracted through convolutional neural networks; each model is trained under different initialization and hyperparameter settings to ensure the model's performance under different image features.

[0173] Step 3: Employ a dynamic ensemble learning strategy to fuse the outputs of multiple trained 3D-UNet models, obtaining the final segmentation result through a mean ensemble strategy. The advantage of ensemble learning is that it can reduce the bias of a single model, improving the overall segmentation accuracy and robustness.

[0174] like Figure 3 The diagram shown is a structural diagram of one of the five integrated 3D-UNet models. Light blue and light red represent the input and output of the 3D MRI image, respectively.

[0175] Assuming the shape of the input is (1,32,512,512), this means that the input received by the model is a 3D tensor, where 1 is the batch size, 32 is the number of input channels, and 512x512 is the spatial dimension.

[0176] The input is processed by DoubleConv3D (consisting of two convolutional layers, batch normalization, and ReLU activation). This layer transforms the 32 channels of the input into 64 channels, and the output shape is (1, 64, 512, 512).

[0177] Downsampling phase:

[0178] First downsampling: Downsampling is performed using Down3D. First, MaxPool3D reduces the dimensionality of the input tensor to (1,64,256,256), and then DoubleConv3D (increasing the number of channels from 64 to 128) is used to process it, and the output shape is (1,128,256,256).

[0179] The second downsampling: similar to the first downsampling, the number of channels is increased from 128 to 256, and the output shape is (1,256,128,128).

[0180] The third downsampling: Similarly, the number of channels is increased from 256 to 512, and the output shape is (1,512,64,64).

[0181] Fourth downsampling: Dimensionality reduction continues, the number of channels increases from 512 to 1024, and the output shape is (1,1024,32,32).

[0182] Upsampling phase:

[0183] First upsampling: Use Up3D to reduce the number of channels from 1024 to 512, perform upsampling, and the shape becomes (1,512,64,64). Then concatenate it with the output of the previous layer (shape (1,512,64,64)) to get the shape (1,1024,64,64). Next, process it with DoubleConv3D to reduce the number of channels to 512, and the output shape is (1,512,64,64).

[0184] The second upsampling: the number of channels is reduced from 512 to 256, the shape becomes (1,256,128,128), and after splicing and processing by DoubleConv3D, the output is (1,256,128,128).

[0185] The third upsampling: the number of channels is reduced from 256 to 128, the shape becomes (1,128,256,256), and after splicing and processing by DoubleConv3D, the output is (1,128,256,256).

[0186] The fourth upsampling: the number of channels is reduced from 128 to 64, the shape becomes (1,64,512,512), and after splicing and processing by DoubleConv3D, the output is (1,64,512,512).

[0187] Finally, the OutConv3D layer reduces the number of output channels from 64 to the required number of classes (e.g., 1 for a binary classification task), resulting in an output shape of (1, n_classes, 512, 512).

[0188] The convolution operation mentioned above is immediately followed by batch normalization, the function of which is expressed as follows:

[0189]

[0190] The ReLU function is then used as the activation function, expressed as follows:

[0191] ReLU(x) = max(0,x)

[0192] This scheme employs cross-entropy and uses structure-aware weighting as the loss function, which is defined as follows:

[0193]

[0194] in To enhance the Dice loss at the edges, The topology preservation loss function is expressed as follows:

[0195]

[0196] Where y edge Extracted using the Canny operator (threshold 0.2, σ 1.0);

[0197]

[0198] Where χ() is the Euler number characteristic, used to constrain the tubular topology of lymphatic vessels; P = 8 neighborhood connectivity.

[0199] Finally, the Dice loss function is used as the evaluation metric for the segmentation network. The function is defined as follows:

[0200]

[0201] Where y represents the actual lymphatic vessel segmentation annotation, which is a three-dimensional binary matrix, where 1 (or 255) represents a lymphatic vessel voxel and 0 represents the background region, which is obtained by the operator manually annotating on the MRI image. The segmentation result predicted by the model is a three-dimensional probability matrix. After sigmoid activation, the value is ∈ [0,1]. The closer it is to 1, the higher the probability that the voxel belongs to the lymphatic vessel. Usually, a threshold of 0.5 is used for binarization (if >0.5, it is considered positive). This represents the calculation of the number of overlapping voxels between the actual annotations and the predicted results, implemented by adjusting y and The summation of element-wise multiplications reflects the lymphatic vessel regions correctly identified by the model; |y| represents the total number of lymphatic vessel voxels in the ground truth annotations. This represents the total number of voxels predicted by the model as lymphatic vessels. The denominator is essentially the sum of the number of voxels in the real and predicted regions.

[0202] This evaluation index is sensitive to imbalanced data (such as lymphatic vessels accounting for only 1-5% of the image), and its value ranges from [0,1]. The larger the value, the higher the segmentation accuracy. If it is 1, it is a perfect match, and if it is 0, it is a complete mismatch.

[0203] This scheme trains five 3D-UNet models using different hyperparameters, primarily differing in their learning rate curves, batch sizes, and loss functions. Finally, the prediction results from the five models are used to perform a voting decision through dynamic ensemble learning. The dynamic ensemble learning strategy includes:

[0204] The weights of each model are calculated based on the Dice scores on the validation set, and the prediction results of multiple models are combined using a weighted average, as shown in the following formula:

[0205]

[0206] The weights are calculated as follows:

[0207]

[0208] Among them, Dice m Let be the Dice score of the m-th model on the validation set.

[0209] Specifically, weight calculation is based on the Dice score of the validation set; a dynamic weight allocation function, based on a variant of the Softmax function, achieves weighted fusion of prediction results from multiple models; and adaptive weighting in model ensemble is achieved through exponential amplification and normalization. Its core advantage lies in adjusting the weight distribution through temperature parameters, making the ensemble result more aligned with the high-performance model. In practical applications, attention must be paid to the selection of temperature parameters and the reliability verification of Dice values, such as... Figure 4 As shown.

[0210] like Figure 2 As shown, this solution also provides a lymphatic vessel segmentation system based on Dynamic Integrated Structure-Aware 3DU-Net, used to implement the lymphatic vessel segmentation method based on Dynamic Integrated Structure-Aware 3DU-Net, including:

[0211] The data preprocessing module is used to preprocess the input MRI image data and generate a standardized three-dimensional matrix;

[0212] The data augmentation module is used to improve the model's robustness to differences in scan parameters;

[0213] The multi-model dynamic integrated inference module includes multiple 3D-UNet models for predicting MRI image data and outputting the corresponding prediction results.

[0214] The dynamic fusion module is used to perform weighted fusion of the output of the 3D-UNet model to obtain the final segmentation result;

[0215] The post-processing module is used to perform morphological post-processing on the final segmentation results to optimize the accuracy and stability of the segmentation results.

[0216] Data flow during operation: Raw MRI image data → Orientation alignment → Z-axis resampling → Window width adjustment → [Gaussian noise / random flipping] → 3D-UNet1~5 → Probabilistic fusion → Binarization → Morphological optimization → Lymphatic vessel mask.

[0217] A dynamic weighting mechanism is adopted to adjust the fusion weights in real time based on the Dice performance of each model on the validation set (Table 3);

[0218] Structure-aware training is used, through and Loss function constraints on tubular morphology;

[0219] A multi-scale collaborative approach is adopted, with models 1 and 3 focusing on backbone segmentation, models 2 and 4 handling noise / fragmentation, and model 5 ensuring stability for small samples.

[0220] The input size of 512×512×32 is compatible with the resolution of common MRI equipment; noise injection simulates motion artifacts in cancer patients, and rotational augmentation covers the morphological variations of lymphedema.

[0221] By using the technical path of differential initialization → collaborative training → dynamic ensemble, this system can achieve an optimization effect of -0.67156 Dice score on the test set (Table 3), which is about 12% higher than the traditional single-model method.

[0222] As a specific embodiment of this scheme, the implementation method of the training phase was carried out using the private brain MRI dataset of the Affiliated Hospital of Jiangnan University. The final training results are shown in Table 2 (the following are the scores on the test set).

[0223] Table 2 Training Results Before Voting

[0224] Model ID Cross-entropy loss Dice Score 1 0.7289 -0.6617 2 0.7056 -0.6769 3 0.7326 -0.6713 4 0.7302 -0.6759 5 0.7068 -0.6570

[0225] As another specific embodiment of this scheme, the mean ensemble algorithm is used to vote on the above 5 3D-UNets to obtain the final model training results in Table 3 (the following are the scores on the validation set).

[0226] Table 3 Final Prediction Results

[0227] Cross-entropy loss Dice Score 0.72082 -0.67156

[0228] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For 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 protection scope of the present invention.

Claims

1. A lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net, characterized in that, It includes a training phase and a usage phase, in which: Training phase Multiple pre-training 3D-UNet models with different initialization and hyperparameter settings are constructed, and each pre-training 3D-UNet model adopts a different initialization strategy; Preprocess the MRI image data to generate a training dataset; Based on the training dataset, each of the pre-training 3D-UNet models is trained using a structure-aware module and a multi-scale feature fusion strategy to obtain the post-training 3D-UNet model. The outputs of multiple post-3D-UNet models are fused using a dynamic ensemble learning strategy to obtain an ensemble segmentation result; Use phase The MRI image data to be segmented is preprocessed to generate an image dataset to be processed; Multiple trained 3D-UNet models are used to predict the image dataset to be processed, and the corresponding prediction results are obtained. The prediction results are fused using a dynamic weight fusion strategy to obtain the final segmentation result; Morphological post-processing is performed on the final segmentation results to optimize the accuracy and stability of the segmentation results; The pre-training 3D-UNet model consists of 5 models, including: 3D-UNet model before training 1: The convolutional layers are initialized using a Xavier normal distribution; The normalization layer is initialized using a uniform He distribution; The learning rate employs a cosine annealing strategy; 3D-UNet model before training 2: Convolutional layers undergo Kaiming orthogonal initialization; Normalization layer parameters are initialized to standard deviation Gaussian distribution; The learning rate decreases in a stepwise manner; 3D-UNet model before training: The convolutional layers use strictly orthogonal initialization, satisfying Constraints; The normalization layer scaling factor is fixed at 1 Constant initialization; The optimizer uses adaptive AdamW; 3D-UNet model before training 4: The convolutional layer is sparsely initialized, with 50% of the weights set to 0. The normalization layer is initialized using dynamic range scaling; The learning rate changes cyclically according to a triangular period. 3D-UNet model before training 5: The convolutional layers employ prediction-level weight transfer, inheriting convolutional kernel parameters from a pre-trained liver segmentation model; Normalize the γ and β parameters of the three layers before freezing; Learning rate linear warmup: .

2. The lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net according to claim 1, characterized in that, The initialization strategy for the 3D-UNet model before training includes at least one combination of the following: Convolutional layer initialization: Xavier normal distribution, Kaiming orthogonal initialization, orthogonal initialization, sparse initialization, or prediction-level weight transfer; Normalization layer initialization: uniform He distribution, zero-mean Gaussian distribution, fixed scaling factor, dynamic range scaling, or freezing the first 3 layers; Learning rate strategies: cosine annealing, step descent, adaptive AdamW, cyclic learning rate, or linear warmup.

3. The lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net according to claim 1, characterized in that, The preprocessing methods for the MRI image data and the preprocessing of the MRI image data to be segmented are the same, including: Rigid registration of MRI images was performed, and alignment was achieved using a 6-DOF affine transformation. Isotropic resampling based on B-spline interpolation was performed on MRI images of different sizes, with a unified network input size of 512×512×32 and a voxel spacing of 1mm×1mm×3mm. Dynamic data augmentation strategies are employed, including adaptive window width and level adjustment, spatial transformation, and noise injection.

4. The lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net according to claim 1, characterized in that, The structure-aware module includes edge enhancement Dice loss and topology preservation loss, with the loss function being: ， in To enhance the Dice loss at the edges, The topology preservation loss function is expressed as follows: ， in Extracted using the Canny operator; ， in, The Euler number characteristic is used to constrain the tubular topology of lymphatic vessels; P refers to 8-neighborhood connectivity.

5. The lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net according to claim 1, characterized in that, The dynamic ensemble learning strategy includes: The weights of each model are calculated based on the Dice scores on the validation set, and the prediction results of multiple models are combined using a weighted average, as shown in the following formula: ， The weights are calculated as follows: ， in, For the first The models on the validation set e-score.

6. The lymphatic vessel segmentation method based on dynamic integrated structure-aware 3DU-Net according to claim 5, characterized in that, The morphological post-processing includes: Anisotropic closing operations are performed using a flattened oval structuring element, with parameters as follows: ; The connected component filtering criteria remove fragments and retain tubular structures. The retention criteria are: 。

Images