A 3D medical image segmentation system and method based on triaxial structure enhancement

Abstract

A three-dimensional medical image segmentation system and method based on three-dimensional spatial enhancement is disclosed, belonging to the field of medical image processing technology. It solves the technical problem of accurately depicting the three-dimensional morphological features of organs when the RWKV architecture is directly applied to three-dimensional medical image segmentation tasks. The system includes an encoder and a decoder. The encoder includes several downsampling layers containing three-dimensional spatial enhancement modules, and the decoder includes the same number of upsampling layers containing three-dimensional spatial enhancement modules as the encoder. The three-dimensional spatial enhancement modules in the encoder are connected to the corresponding three-dimensional spatial enhancement modules in the decoder via skip connections. The three-dimensional spatial enhancement modules are improved from existing RWKV modules, with improvements including: changing the RWKV module's serialization modeling operation from a single sequence direction to three sequence directions; and adding spatial shift operations at the data input ends of the two hybrid modules of the RWKV module.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a three-dimensional medical image segmentation system and method based on three-dimensional structure enhancement. Background Technology

[0002] Three-dimensional medical image segmentation is a core task in the field of medical image analysis. Its goal is to accurately identify and voxel-by-voxel delineate the boundaries of target organs, tissues, or lesions from three-dimensional medical images such as CT and MRI, thereby providing crucial quantitative information on shape, location, and volume. This technology, as an important prerequisite for image-guided diagnosis and computer-aided intervention, provides indispensable support for clinical diagnosis, surgical planning, radiotherapy, and disease progression monitoring, and has vital clinical application value. However, three-dimensional medical images typically exhibit characteristics such as large data volume, complex structure, variable target scale, and blurred boundaries, which places extremely high demands on the feature extraction capabilities, computational efficiency, and spatial structure modeling capabilities of deep learning models.

[0003] In recent years, the rapid development of deep learning has brought about a significant revolution in 3D medical image segmentation. Architectures based on convolutional neural networks (CNNs), such as 3DU-Net, can effectively extract multi-scale local features and recover spatial details from volumetric data through encoder-decoder structures and skip connections, providing a feasible solution for automatic segmentation. However, this technology has inherent limitations. Its core computation relies on local convolution operations, resulting in a limited receptive field and difficulty in establishing long-distance global dependencies between voxels. Although subsequent improvements have expanded the receptive field by increasing the convolution kernel size, this faces enormous computational and memory overhead when processing high-resolution 3D scans commonly found in clinical settings.

[0004] Many studies have explored effective methods for capturing long-range dependencies. Transformer-based methods, with their self-attention mechanism, possess powerful global contextual information capture capabilities. However, their computational complexity is proportional to the square of the sequence length, resulting in significant computational overhead when processing massive voxels in 3D medical images. State-space models, such as Mamba, excel in long sequence modeling with linear computational complexity, but their method of flattening 3D data into a one-dimensional sequence disrupts spatial topology, leading to spatial information loss. The emerging Receptance Weighted Key Value (RWKV) architecture combines the efficiency of RNNs with the expressive power of Transformers, but its direct application to 3D tasks has significant limitations: flattening 3D voxels into a sequence disrupts spatial neighborhood relationships, and unidirectional sequence modeling is insensitive to 3D spatial directionality, making it difficult to accurately characterize the three-dimensional morphological features of organs. Summary of the Invention

[0005] To address the technical problems that arise when the RWKV architecture is directly applied to 3D medical image segmentation tasks, such as the destruction of spatial neighborhood relationships due to 3D voxel flattening into a sequence, and the insensitivity of unidirectional sequence modeling to 3D spatial directionality, making it difficult to accurately characterize the three-dimensional morphological features of organs, this invention provides a 3D medical image segmentation system and method based on triaxial structure enhancement.

[0006] The system includes an encoder and a decoder. The encoder includes several downsampling layers containing triaxial spatial enhancement modules. The decoder includes the same number of upsampling layers containing triaxial spatial enhancement modules as the encoder. The triaxial spatial enhancement modules in the encoder are connected to the corresponding triaxial spatial enhancement modules in the decoder via skip connections.

[0007] The three-dimensional spatial enhancement module is an improvement on the existing RWKV module. The improvements include: changing the serialization modeling operation of the RWKV module from a single sequence direction to three sequence directions; adding a spatial shift operation at the data input end of the two hybrid modules of the RWKV module; and adding a 3D convolution module at the output end of the RWKV module.

[0008] Furthermore, the three sequence directions are the depth, height, and width directions of the input features.

[0009] Furthermore, the spatial shifting operation is specifically as follows: the input feature is divided into six parts according to the channels, shifted forward along the depth direction in one channel, shifted backward along the depth direction in another channel, shifted forward along the height direction in one channel, shifted backward along the height direction in another channel, shifted forward along the width direction in one channel, and shifted backward along the width direction in another channel.

[0010] Furthermore, the encoder also includes a block embedding layer, where features are first input and then enter the block embedding layer; the decoder also includes a final projection layer, where features from the decoder are output from the final projection layer.

[0011] Furthermore, in the system, the input features are added element-wise to the output features of the final projection layer through residual blocks.

[0012] Furthermore, the system also includes a segmentation head, which adds the input features and the output features of the final projection layer element by element, and then inputs the input features into the segmentation head, which outputs a segmented image.

[0013] Furthermore, in the encoder, the downsampling layer containing the three-dimensional spatial enhancement module is composed of the three-dimensional spatial enhancement module and the downsampling module in sequence along the data input direction; along the data input direction, the final output features of several downsampling layers containing the three-dimensional spatial enhancement module are input into the three-dimensional spatial enhancement module;

[0014] In the decoder, the upsampling layer containing the three-dimensional spatial enhancement module is composed of the three-dimensional spatial enhancement module and the upsampling module in sequence along the data input direction; along the data input direction, the final output features of several upsampling layers containing the three-dimensional spatial enhancement module are input into the three-dimensional spatial enhancement module.

[0015] Based on the above system, the present invention further provides a three-dimensional medical image segmentation method based on triaxial structure enhancement. The method uses a three-dimensional medical image segmentation system based on triaxial structure enhancement to segment the three-dimensional medical image to be processed and obtain the segmentation result.

[0016] Furthermore, the three-dimensional medical image to be processed is segmented using a three-dimensional medical image segmentation system based on triaxial structure enhancement. The system is trained, and after passing the training, the three-dimensional medical image to be processed is segmented.

[0017] Furthermore, the training of the system specifically involves:

[0018] S1. Collect a 3D medical image dataset, divide it into training and validation sets, and perform preprocessing.

[0019] S2. Input the preprocessed training set data into the system, perform end-to-end training using a hybrid loss function, and evaluate the system's segmentation performance using a validation set.

[0020] S3. Validate and evaluate the trained system using the validation set. If it fails, return to step S2 to continue training. If it succeeds, the validation is complete and the system can be applied.

[0021] The beneficial effects of the system described in this invention are as follows:

[0022] (1) Enhance the ability to model three-dimensional spatial structures and expand the receptive field: By constructing a three-dimensional spatial enhancement module (TSE-R Block), combined with a parameterless spatial shifting strategy and a three-dimensional orthogonal scanning mechanism, the spatial interaction relationship between local voxels is explicitly established while maintaining linear computational complexity. This effectively expands the model's receptive field and alleviates the problem of spatial topological structure destruction caused by sequential modeling, thereby significantly improving the modeling accuracy of complex organ volume morphology.

[0023] (2) Enhance multi-directional semantic perception and edge detail capture capabilities: By designing an independent sequence scanning and directional attention weighted fusion mechanism along the depth, height and width directions (x, y and z three orthogonal directions), the model can achieve collaborative modeling of multi-directional global dependence and local geometric details of three-dimensional images, enhance the model's segmentation consistency and boundary positioning accuracy in edge and structurally complex regions, and significantly improve the anatomical integrity of the segmentation results.

[0024] (3) Achieving efficient global-local feature fusion and computational resource optimization: The RWKV linear attention mechanism and 3D convolution operation are integrated into the three-dimensional spatial enhancement module. While constructing long-distance dependencies, the local correlation between voxels is preserved, achieving an efficient balance between global context and local details. Experiments show that the model reduces the number of parameters to 9.33M in multiple 3D medical image segmentation tasks, which is about 16.08 times lower than the existing state-of-the-art methods. It significantly improves computational efficiency while maintaining high accuracy and has practical value for deployment on edge devices. Attached Figure Description

[0025] Figure 1 This is a structural diagram of a three-dimensional medical image segmentation system based on three-dimensional structure enhancement in an embodiment of the present invention;

[0026] Figure 2 This is a structural diagram of the three-dimensional space enhancement module in an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram illustrating the serialization modeling operation of the Tri-wkv module in three sequence directions in an embodiment of the present invention;

[0028] Figure 4 This is a graph showing the qualitative comparison results of U-RWKV with other models on the VerSe2019 dataset in this embodiment of the invention.

[0029] Figure 5 This figure shows the comparison results of the accuracy and computational complexity of U-RWKV with other models on the Synapse dataset in this embodiment of the invention. Detailed Implementation

[0030] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only 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 protection scope of the present invention.

[0031] Example 1

[0032] This embodiment provides a three-dimensional medical image segmentation system based on triaxial structure enhancement. The structure diagram of the system is shown below. Figure 1 As shown, the system includes an encoder and a decoder. The encoder includes three downsampling layers containing three-dimensional spatial enhancement modules (TSE-RBlock), and the decoder includes three upsampling layers containing three-dimensional spatial enhancement modules. The three-dimensional spatial enhancement modules in the encoder are connected to the corresponding three-dimensional spatial enhancement modules in the decoder via skip connections.

[0033] In the encoder, the downsampling layer containing the three-dimensional spatial enhancement module is composed of the three-dimensional spatial enhancement module and the downsampling module in sequence along the data input direction; along the data input direction, the final output feature of the downsampling layer containing the three-dimensional spatial enhancement module is input to a three-dimensional spatial enhancement module;

[0034] In the decoder, the upsampling layer containing the three-dimensional spatial enhancement module is composed of the three-dimensional spatial enhancement module and the upsampling module in sequence along the data input direction; along the data input direction, the final output feature of the upsampling layer containing the three-dimensional spatial enhancement module is input into a three-dimensional spatial enhancement module.

[0035] The encoder also includes a block embedding layer, where features are first input and then enter the block embedding layer; the decoder also includes a final projection layer, where features from the decoder are output from the final projection layer.

[0036] The input features are added element-wise to the output features of the final projection layer through residual blocks.

[0037] The system also includes a segmentation head. The input features are added element-wise to the output features of the final projection layer and then input into the segmentation head, which outputs a segmented image.

[0038] Example 2

[0039] This embodiment further defines Embodiment 1. The structural diagram of the three-dimensional space enhancement module is shown below. Figure 2 As shown, the three-dimensional space enhancement module is an improvement on the existing RWKV module, and the improvements include:

[0040] Improvement 1: The serialization modeling operation of the RWKV module has been modified from a single sequence direction to three sequence directions. In the original RWKV module, the serialization modeling operation in a single sequence direction was implemented through the wkv module, such as... Figure 2 As shown, in the three-dimensional spatial enhancement module, the Tri-wkv module replaces the wkv module. The Tri-wkv module implements serialization modeling operations in three sequence directions. A schematic diagram of the Tri-wkv module performing serialization modeling operations in three sequence directions is shown below. Figure 3 As shown;

[0041] To ensure that WKV serialization modeling does not lose 3D spatial structure information, the sequence dimensions are first reshaped to their original spatial dimensions, that is, the flattened one-dimensional sequence is restored to its 3D form (depth, height, width). Then, it is flattened into a 1D sequence along the depth, height, and width directions of the feature, and then subjected to three-way orthogonal scanning processing. Finally, the WKV attention, which includes 3D spatial information, is obtained through weighted fusion in the three directions. The calculation process can be described by the formula:

[0042] ;

[0043] in, This represents the wkv attention containing three-dimensional spatial information obtained by performing serialization modeling operations in three sequence directions through the Tri-wkv module; This indicates that the Tri-wkv module handles input features and The operation performed; Indicates the kernel size as 3D convolution; This indicates a stacking operation along the channel dimension, which means splicing them together; , and They represent the input features respectively and The flattened one-dimensional sequence is restored to its three-dimensional form (depth, height, and width). Then, it is flattened into a one-dimensional sequence along the depth, height, and width directions of the feature, and then processed by three-way orthogonal scanning to obtain three feature sequences.

[0044] Improvement point two: Add a space shift operation to the data input terminals of the two hybrid modules of the RWKV module, such as... Figure 2 As shown, in the two mixing modules of the RWKV module, namely the Spatial Mix module and the ChannelMix module, a spatial shift operation has been added to the data input end;

[0045] The spatial shift operation is a parameter-free spatial shift strategy designed in this invention to address the spatial topology disruption problem caused by serialization modeling. The specific operation is as follows: the sequence data is restored to a three-dimensional form, and shifted along the depth, height, and width directions respectively. The calculation formula is as follows:

[0046] ;

[0047] in, This indicates splicing along the channel; the characteristics of this operation typically include... aisle, depth, Height and The width and other dimensions are considered. Since the number of channels varies with the number of layers, they are divided proportionally. In this embodiment, the feature is divided into six parts according to the number of channels and processed separately. First, in channels 0 to 1 / 6, the voxel is shifted to the left along the width direction (…). In channels 1 / 6 to 2 / 6, the voxel moves to the right along the width direction ( This process is repeated six times, and then these six parts are spliced ​​together along the channel to obtain the final output. This operation explicitly establishes interactions between local neighborhoods, expands the receptive field, and enhances spatial correlation.

[0048] Improvement point 3: Add a 3D convolution module to the output of the RWKV module.

[0049] Example 3

[0050] This embodiment is a further limitation of embodiment 1. In the encoder, after the feature input, it first enters the patch embedding layer. Through a patch embedding layer, the input image is divided into several image patches of the same size that do not overlap.

[0051] The block embedding layer is the input preprocessing layer of the encoder. Its core responsibility is to perform structured dimensionality reduction and serialization on high-resolution 3D input medical images, transforming the continuous voxel space into a block sequence suitable for subsequent RWKV sequence model processing. The specific implementation is as follows: Given an input 3D medical image... This module first divides it into a series of 3D Patches of the same size that do not overlap. s (d represents the patch size), this operation is implemented through a three-dimensional convolution, the kernel size and stride of which are set to be the same as the image patch size. This convolutional layer performs two functions at the same time: the first is spatial segmentation, which divides the input image into spatial dimensions evenly through non-overlapping convolution; the second is channel projection, which projects all voxel values ​​in each image patch onto a new, higher feature channel dimension.

[0052] In the decoder, features are output from the final projection layer, which restores the image patches output by the decoder to the same dimension as the input. This step aims to recover the spatial details lost during encoder downsampling.

[0053] Subsequently, the data is fed into the Seg Head for fine-grained prediction: the Seg Head is a lightweight convolutional network, typically consisting of several standard 3D convolutional layers, normalization layers (such as InstanceNorm3d), and activation functions (such as ReLU). Its purpose is to perform a final non-linear transformation and channel mapping on the upsampled features, converting them into a probability map corresponding to the number of segmentation categories.

[0054] The final projection layer quickly recovers the global size through a simple upsampling operation, while the subsequent segmentation head uses its local receptive field and nonlinear capabilities to refine the global features generated by the sequence model, working together to generate high-quality segmentation results with clear boundaries and accurate semantics.

[0055] Example 4

[0056] This embodiment is a further explanation of Embodiment 1, introducing data processing in a three-dimensional medical image segmentation system based on triaxial structure enhancement.

[0057] First, a three-dimensional medical image is given as input. (Where C represents the number of channels, and D, H, and W represent the depth, height, and width of the 3D data, respectively), the input image is segmented into several non-overlapping patches of the same size through a block embedding layer, resulting in... (d represents the patch size). This feature is then fed into an encoder consisting of four layers of three-dimensional spatial augmentation modules, with the number of channels in each layer being [C, 2C, 4C, 8C]. The encoder downsamples using standard convolutional blocks, adjusting the resolution and number of channels of the feature map. The decoder uses transposed convolutions for upsampling and adds the output features of the corresponding encoder layer to the input features of the decoder element-wise through skip connections to preserve spatial detail. The final output features of the decoder are then restored to the input size through a final projection operation. The features are added to the original input after processing by the residual block, and finally the final segmentation result is generated by the segmentation head based on convolution operation.

[0058] The three-dimensional spatial enhancement module is the core module of the three-dimensional medical image segmentation system with three-dimensional structural enhancement. It aims to achieve efficient global modeling and local structure preservation. This module contains two sub-modules: spatial mixing and channel mixing, and introduces a 3D convolutional block at the end to enhance local correlation.

[0059] The spatial blending part first involves input features. It is rearranged into a sequence, and then subjected to a spatial shift operation, the formula of which is:

[0060] ; Indicates length, and equal.

[0061] Then, After processing through three independent linear layers, its calculation formula is as follows:

[0062] .

[0063] then and The inputs are fed into the Tri-wkv module to compute global attention, and then processed by the Sigmoid algorithm. Multiplication controls the strength of the features, resulting in the output features. Finally, after normalization, the final output is obtained. Input channel mixing module. Its processing formula is:

[0064] ;in, This represents the Sigmoid function. Representation layer normalization.

[0065] The channel mixing part also undergoes spatial shifting, and then passes through two parallel linear layers to obtain... and The calculation formula is as follows:

[0066] ;

[0067] ;

[0068] at last, First, it is mapped to a hidden dimension, then subjected to Square ReLU, and finally mapped back to the original dimension. Then, with the sigmoid-processed Multiply to obtain the output features , After normalization, the final output is obtained. The calculation process is as follows:

[0069] ;

[0070] ;

[0071] in, Represents the squared ReLU function. and These are two hidden dimensions with opposite mapping directions.

[0072] A standard 3D convolution module is introduced at the end of the three-dimensional spatial enhancement module to enhance the correlation between local voxels and prevent the loss of details.

[0073] Example 5

[0074] This embodiment provides a three-dimensional medical image segmentation method based on triaxial structure enhancement. The method uses a three-dimensional medical image segmentation system based on triaxial structure enhancement to segment the three-dimensional medical image to be processed.

[0075] Includes the following steps:

[0076] S1. Data Preparation and Preprocessing:

[0077] We collected a 3D medical image dataset, including multimodal data such as abdominal CT and brain MRI. We divided the training set and validation set in a 7:3 ratio and performed uniform voxel spacing adjustment and intensity normalization on all data.

[0078] S2. Construct a three-dimensional medical image segmentation system based on three-dimensional structure enhancement. The system adopts a U-shaped encoder-decoder structure, specifically as follows:

[0079] S21. Construct an encoder, which consists of four downsampling stages. The first three stages each contain a three-dimensional spatial enhancement module and a 3D convolutional downsampling layer with a stride of 2. The fourth stage contains only a three-dimensional spatial enhancement module, which is used to gradually extract deep features.

[0080] S22. Construct a decoder, which also consists of four upsampling stages. The first three stages each contain a three-dimensional spatial enhancement module and a 3D transposed convolutional upsampling layer with a stride of 2. The fourth stage contains only a three-dimensional spatial enhancement module. All stages are fused with the corresponding layer features of the encoder through skip connections.

[0081] S23. Construct a segmentation head, which consists of two 3D convolutional layers and is used to map the features output by the decoder to the segmentation results of the target class.

[0082] S3. Model Training and Validation: The preprocessed training set data is input into the three-dimensional medical image segmentation system based on triaxial structure enhancement. End-to-end training is performed using a hybrid loss function, and the system segmentation performance is evaluated through a validation set.

[0083] S4. Model Inference Application: Input the 3D medical image to be segmented into the trained 3D medical image segmentation system based on triaxial structure enhancement, and directly output the corresponding voxel-level segmentation results.

[0084] Example 6

[0085] This embodiment uses specific experiments to systematically verify the effectiveness, versatility, and efficiency of the system described in Examples 1-4 (represented by U-RWKV) in three-dimensional medical image segmentation tasks.

[0086] To comprehensively validate the universality and effectiveness of U-RWKV, a systematic evaluation was conducted on three representative 3D medical image segmentation datasets: the Synapse abdominal multi-organ segmentation dataset, the Verse2019 vertebral segmentation dataset, and the BraTS2021 brain tumor segmentation dataset. Furthermore, to further analyze the contributions of each module, detailed ablation experiments were performed on the Spatial Mix, Channel Mix, Spatial Shift, and Tri-wkv scanning strategies in the TSE-R Block. All comparative methods were trained and evaluated under the same experimental settings to ensure the fairness of the comparison.

[0087] The Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (HD95) were used as core evaluation metrics. These two metrics quantify the segmentation results from two dimensions: regional overlap and boundary fit. The Dice coefficient reflects segmentation accuracy by calculating the degree of overlap between the predicted segmented region and the gold standard (Ground Truth), and its definition is as follows: ;

[0088] Where A represents the segmentation region predicted by the model, and B represents the Ground Truth. The range of DSC values ​​is... The higher the percentage, the higher the overlap between the segmentation result and the ground truth (GT), and the better the segmentation effect.

[0089] However, it is worth noting that this metric is sensitive to the completeness of the internal filling of the segmented target, but may not be sensitive to boundary detail deviations, which is insufficient for medical image segmentation tasks with extremely high boundary accuracy requirements.

[0090] Therefore, to evaluate the consistency between the segmentation boundary and the true contour, HD95 is introduced to measure the distance between the predicted segmentation boundary and the voxel of the ground truth (GT). Its definition is as follows: ;

[0091] Where A represents the segmentation region predicted by the model, and B represents the Ground Truth. This represents the maximum distance between surface voxel points A and B, with the 95th percentile value taken as the result, in mm.

[0092] The experiments were implemented using PyTorch 2.1.2 and trained on an NVIDIA GeForce RTX 4090 24GB GPU. For the Synapse dataset, [the following was used]. , and Training with a specific input size. For the VerSe dataset, using... Slice size. For the BraTS dataset, the following is used: Training was performed using an input size of 2. All experiments used a batch size of 2, an SGD optimizer with Nesterov momentum (coefficient μ=0.99), and a weight decay of 3e-5. The loss function was a weighted sum of cross-entropy loss and Dice loss, and the data preprocessing workflow was consistent with nnUNet.

[0093] To fully verify the performance of the model and method of this invention, experiments were conducted on a three-dimensional medical image dataset. Tables 1 to 5 report the quantitative comparison results with state-of-the-art models on three datasets. Figure 4 This corresponds to the qualitative comparison results in Table 3. Figure 4 L1-L5 in the text refers to the first to fifth lumbar vertebrae. Figure 5 This is a comparison chart of the accuracy and computational complexity of the present invention on the Synapse dataset, corresponding to Table 5. Figure 5 The unit M in the table is the unit of the evaluation index Param, which represents millions. U-RWKV-96 corresponds to the cutting size [96×96×96] in Table 5, U-RWKV-128 corresponds to the cutting size [64×128×128] in Table 5, and U-RWKV-192 corresponds to the cutting size [64×192×192] in Table 5.

[0094] Table 1 shows a quantitative comparison with other methods on the Synapse dataset of abdominal organs, using the Dice similarity coefficient for evaluation. Another evaluation index, HD95, was also introduced (Aor: aorta, Gal: gallbladder, L Kid: left kidney, R Kid: right kidney, Liv: liver, Pan: pancreas, Spl: spleen, Sto: stomach).

[0095] Table 1:

[0096]

[0097] Table 2 uses larger slice sizes for quantitative comparison with the SOTA method. Compared with other methods, the Dice similarity coefficient is used for evaluation, and another evaluation index, HD95, is also introduced.

[0098] Table 2:

[0099]

[0100] Table 3 presents a quantitative comparison on the VerSe2019 dataset. Compared with other methods, the Dice similarity coefficient was used for evaluation, and another evaluation index, HD95, was also introduced. In Table 3, L1-L5 represent the first to fifth lumbar vertebrae.

[0101] Table 3:

[0102]

[0103] Table 4 shows a quantitative comparison with the SOTA method on the BraTS2021 dataset. Compared with other methods, the Dice similarity coefficient was used for evaluation. Another evaluation index, HD95, was also introduced (WT: whole tumor, TC: tumor core, ET: enhancing tumor).

[0104] Table 4:

[0105]

[0106] Table 5 compares the computational efficiency of the SOTA method on the Synapse dataset. Compared with other methods, the Dice similarity coefficient is used for evaluation, and another evaluation metric, HD95, is also introduced.

[0107] Table 5:

[0108]

[0109] The results show that the method of the present invention has good universality for three-dimensional medical image segmentation tasks and can be used as a general model and method for three-dimensional medical image segmentation.

[0110] Table 6 shows the ablation experiments conducted on the Synapse dataset with a slice size of 96×96×96 to verify the effectiveness of each module in U-RWKV. The aim is to verify the contribution of each internal component (Spatial Mix, Channel Mix, SpatialShift) to the final segmentation performance. Among them, (1) Model 1 is an extremely simplified baseline, replacing TSE-RBlock with a regular 3D convolutional block and completely removing the Spatial Mix, Channel Mix and SpatialShift operations. This configuration only has basic local feature extraction capabilities and cannot perform effective long-distance dependency modeling, resulting in severely low performance. The Dice similarity coefficient is only 70.08%, and the HD95 is as high as 62.53mm. This fully demonstrates that the lack of an efficient global modeling mechanism is the key bottleneck limiting the performance of the three-dimensional medical image segmentation model.

[0111] (2) Model 2 retains Channel Mix and Spatial Shift operations based on Model 1, but removes the Spatial Mix part responsible for global spatial information modeling. Its Dice score is improved to 81.84%, and HD95 is significantly reduced to 14.78mm, representing a huge leap in performance compared to Model 1. This is mainly due to the effective modeling of complex relationships between channels by Channel Mix and the expansion of local receptive fields by Spatial Shift. However, due to the lack of the Tri-wkv module in Spatial Mix, the model's ability to capture long-distance contextual information between voxels is insufficient, resulting in a performance gap compared to the complete model.

[0112] (3) Model 3 retains Spatial Mix and Spatial Shift, but removes the Channel Mix component. This configuration achieves a Dice score of 77.34%, even lower than Model 2, indicating that the Channel Mix module is crucial for model performance. Channel Mix, through the squared ReLU activation function and linear transformation, is specifically responsible for modeling the interactions and dependencies between different feature channels. Removing this module severely disrupts the network's ability to integrate and utilize channel information. Even with strong spatial global modeling capabilities, it cannot focus on key features through effective channel recalibration, resulting in a significant performance drop.

[0113] (4) Model 4 forms a strong baseline, containing the two core mixing modules, Spatial Mix and Channel Mix, but removing the Spatial Shift operation. This model achieved a Dice score of 83.96%, demonstrating the effectiveness of RWKV-based global spatial and channel modeling. However, its performance is still insufficient compared to the complete model. This is because RWKV's serialization modeling is inherently spatially independent, and simple flattening operations disrupt the local neighborhood topology between voxels. The absence of the Spatial Shift operation prevents the model from explicitly introducing local contextual information and expanding the receptive field before building global dependencies, thus limiting the expressive power of features to some extent.

[0114] (5) U-RWKV, as the complete system of this invention, integrates all components of the TSE-R Block and achieves the optimal Dice score of 85.61% and HD95 of 12.74mm. This result indicates that the Spatial Mix, Channel Mix, and SpatialShift modules form a complementary and collaborative organic whole: Spatial Shift effectively preserves the local spatial structure and expands the receptive field before serialization modeling through parameterless displacement operations; Spatial Mix further captures global spatial context dependencies based on its Tri-wkv mechanism; and Channel Mix achieves adaptive recalibration of feature channels through dynamic weighting of channel relationships, highlighting important features. The combined effect of these three modules enables the model to achieve accurate and efficient segmentation of three-dimensional medical images while maintaining linear computational complexity.

[0115] Table 6:

[0116]

[0117] Table 7 presents the ablation experimental results for different scanning directions in the wkv module on the Synapse dataset. By comparing unidirectional (Model 5), bidirectional (Model 6), and tridirectional (U-RWKV) scanning strategies, the importance of multidirectional sequence modeling for spatial structure preservation in 3D medical image segmentation is verified. The results show that the tridirectional scanning strategy achieves the best results in both Dice coefficient and boundary segmentation accuracy (HD95), at 85.61% and 12.74 mm, respectively.

[0118] Table 7:

[0119]

Claims

1. A three-dimensional medical image segmentation system based on three-way structure enhancement, characterized by, The system comprises an encoder and a decoder, the encoder comprises a plurality of down-sampling layers comprising a three-way spatial enhancement module, the decoder comprises the same number of up-sampling layers comprising a three-way spatial enhancement module as the encoder, and the three-way spatial enhancement modules in the encoder are connected to the corresponding three-way spatial enhancement modules in the decoder through a skip connection; The three-way spatial enhancement module is improved from an existing RWKV module, and the improvement points include: modifying the RWKV module sequence modeling operation from a single sequence direction to three sequence directions, the three sequence directions being a depth direction, a height direction and a width direction of the input feature; adding a spatial shift operation at the data input end of the two mixing modules of the RWKV module; A 3D convolution module is added at the output end of the RWKV module; The spatial shift operation specifically includes: dividing the input feature into six parts according to the channel, shifting forward along the depth direction in one channel, shifting backward along the depth direction in one channel, shifting forward along the height direction in one channel, shifting backward along the height direction in one channel, shifting forward along the width direction in one channel, and shifting backward along the width direction in one channel.

2. The three-dimensional medical image segmentation system based on three-way structure enhancement of claim 1, wherein, In the encoder, a block embedding layer is further included, and after the input feature, the input feature first enters the block embedding layer; in the decoder, a final projection layer is further included, and the feature in the decoder is output from the final projection layer.

3. The three-dimensional medical image segmentation system based on three-way structure enhancement of claim 2, wherein, In the system, the input feature is added to the output feature of the final projection layer element by element through a residual block.

4. The three-dimensional medical image segmentation system based on three-way structure enhancement of claim 3, wherein, In the system, a segmentation head is further included, and after the input feature is added to the output feature of the final projection layer element by element, the input feature is input into the segmentation head, and the segmentation head outputs a segmentation image.

5. The three-dimensional medical image segmentation system based on three-way structure enhancement of claim 4, wherein, In the encoder, the down-sampling layer comprising the three-way spatial enhancement module is sequentially composed of the three-way spatial enhancement module and the down-sampling module along the data input direction; along the data input direction, the final output feature of the plurality of down-sampling layers comprising the three-way spatial enhancement module is input into the three-way spatial enhancement module. In the decoder, the up-sampling layer comprising the three-way spatial enhancement module is sequentially composed of the three-way spatial enhancement module and the up-sampling module along the data input direction; along the data input direction, the final output feature of the plurality of up-sampling layers comprising the three-way spatial enhancement module is input into the three-way spatial enhancement module.

6. A three-dimensional medical image segmentation method based on three-way structure enhancement, characterized by, The method uses the three-dimensional medical image segmentation system based on three-way structure enhancement according to any one of claims 1-5 to segment a three-dimensional medical image to be processed, and obtains a segmentation result.

7. The three-dimensional medical image segmentation method based on three-way structure enhancement according to claim 6, characterized in that, In the use of the three-dimensional medical image segmentation system based on three-way structure enhancement to segment a three-dimensional medical image to be processed, the system is trained, and after the training is qualified, the three-dimensional medical image to be processed is segmented.

8. The three-dimensional medical image segmentation method based on three-way structure enhancement according to claim 7, characterized in that, Training the system specifically includes: S1, collecting a three-dimensional medical image dataset, dividing a training set and a validation set, and preprocessing; S2, inputting the preprocessed training set data into the system, using a hybrid loss function for end-to-end training, and evaluating the system segmentation performance through the validation set; S3, verifying and evaluating the trained system through the validation set, if unqualified, returning to step S2 for further training, if qualified, completing the verification, and applying.

Images