Double-path fusion neural network for prostate precise segmentation and segmentation method

Abstract

The invention belongs to the technical field of medical image segmentation, and relates to a dual-path fusion neural network for prostate precise segmentation and a segmentation method, the neural network constructs a dual-path decoupling encoder architecture based on an nnU-Net framework, captures fine anatomical structure and local texture information through a context sensing residual encoder of a local path, and obtains a dual-path fusion neural network for prostate precise segmentation. A long-range dependency relationship is modeled with linear complexity through a visual state space module of a global path, a double-flow alignment gating module is designed to realize self-adaptive alignment and fusion of cross-path features, and model training is optimized in combination with a mixed loss function and a depth supervision strategy; according to the method, the limitation of an existing segmentation model in the aspect of local structure and global semantic integration is solved, and the segmentation precision, the boundary goodness of fit and the generalization performance of the focus and gland region in the prostate MRI image are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical image segmentation technology, specifically involving a dual-path fusion neural network and segmentation method for precise prostate segmentation, applicable to the automatic and precise segmentation of glandular and lesion regions in prostate cancer MRI images. Background Technology

[0002] Prostate cancer is the second most common malignant tumor among men worldwide, and its clinical diagnosis and treatment planning heavily rely on accurate lesion segmentation from MRI images. The development of deep learning technology has driven progress in medical image segmentation, with U-Net and its improved models becoming the mainstream paradigm due to their encoder-decoder structure and skip connection mechanism. However, the local receptive field characteristics of convolutional neural networks (CNNs) make it difficult to model long-range semantic dependencies, limiting their ability to handle MRI images with blurred prostate lesion boundaries and irregular shapes.

[0003] To overcome the locality limitation of CNNs, researchers introduced the Transformer architecture into the vision domain, achieving global context modeling through a self-attention mechanism. However, the computational complexity of self-attention increases quadratically with the number of image patches, limiting its application in high-resolution MRI images. Existing hybrid architectures (such as TransUNet and Swin-Unet) attempt to combine the local feature extraction advantages of CNNs with the global modeling advantages of Transformers, but they suffer from problems such as semantic fragmentation and insufficient feature alignment. While segmentation models based on visual Mamba (such as Swin-UMamba) capture long-range dependencies with linear complexity, they still need improvement in areas such as local texture feature modeling and generalization ability in multi-center data domains.

[0004] Furthermore, prostate segmentation tasks suffer from issues such as inconsistent dataset distribution, insufficient labeled data, and multimodal data registration errors, leading to insufficient robustness and generalization ability of existing models. The Dice coefficient of different methods in cross-validation can fluctuate by 6-12%, making it difficult to meet the segmentation accuracy requirements of clinical applications. Therefore, designing a segmentation model that balances local detail modeling with global dependency capture, achieves efficient feature fusion, and possesses strong generalization ability has become a critical problem urgently needing to be solved in the field of accurate prostate segmentation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dual-path fusion neural network and segmentation method for accurate prostate segmentation.

[0006] To achieve the objectives of this invention, the following technical solutions are adopted.

[0007] A dual-path fusion neural network for accurate prostate segmentation employs a U-shaped encoder and decoder structure, wherein: The encoder employs a dual-path decoupling architecture to collaboratively extract multi-scale feature representations of local details and global context from the input image; the dual-path decoupling architecture includes parallel local paths and global paths, wherein: The local path is a convolutional attention refinement encoder, which sequentially enhances the input features through residual blocks, performs at least two stages of feature encoding by concatenating at least two local kernel blocks, and fuses the original and encoded features after batch normalization to capture the local texture, edge information and region-level context of the input features, i.e., local features; wherein: the local kernel block contains at least three branch convolutions and selective kernel attention; The global path is a visual state space encoder with a pre-trained VMamba model as the backbone network, including a trunk module and at least two stacked visual state space stages for capturing long-range dependencies of input features. The trunk module converts the input image into embedded features through convolutional layers and instance normalization layers. Each visual state space stage adopts a visual state space layer with a pre-normalized residual structure, which contains layer normalization, a two-dimensional selective scanning module and a deep neural network. The two-dimensional selective scanning module captures the long-range dependencies of the embedded features, i.e., global features. The skip connection is a dual-stream alignment gating module that sequentially fuses local and global features through channel alignment, hybrid feature generation, adaptive gating weight calculation, grouped convolutional refinement, and residual fusion operations. The decoder uses cascaded upsampling operations to gradually restore spatial resolution in order to generate segmentation results.

[0008] Furthermore, the loss function of the dual-path fusion neural network adopts a hybrid loss function, including single-scale loss and deep supervision loss, wherein: The single-scale loss is a weighted sum of the Soft Dice loss and the cross-entropy loss; The depth supervision loss is a weighted sum of the single-scale losses at each resolution level.

[0009] Furthermore, the convolutional attention refinement encoder includes residual blocks arranged in series, at least two feature refinement stages consisting of at least two local kernel blocks arranged in series, a batch normalization layer, and residual connections, wherein: Residual blocks extract initial features through convolutional structures with residual connections, while mitigating the gradient vanishing problem in deep networks and enhancing the expressive power of the initial features. At least two feature refinement stages are used to perform multi-scale modeling and channel attention enhancement on the initial features using at least two concatenated local kernel blocks to achieve refined representation of local features; wherein: each local kernel block includes at least three branch convolutions for capturing local multi-scale features with different receptive fields, a View layer for unifying feature dimensions, and a selective kernel attention mechanism for dynamically learning and strengthening local feature weights; Batch normalization layers are used to unify feature scales to enhance the consistency of feature representation; Residual connections fuse the original input features with the refined encoded features element-wise to output the enhanced encoded features.

[0010] Furthermore, the specific process of the dual-stream alignment gating module is as follows: Channel alignment: Adjust the number of channels in the global path output features using a 1×1 convolutional layer to make it consistent with the number of channels in the local path output features; Hybrid feature generation: The channel-aligned global path output features are added element-wise to the local path output features to obtain a hybrid feature map; Adaptive gating weight calculation: Global average pooling is performed on the mixed feature map, followed by two 1×1 convolutions and non-linear activation, and then grouped gating weights are generated by the Sigmoid function; Grouped convolutional refinement: The mixed feature map is divided into a preset number of feature groups along the channel dimension. Each group is refined locally through independent 3×3 group convolutions and multiplied with the corresponding gate weights. Residual fusion: All weighted group features are concatenated and added to the original mixed feature map through residual connections to obtain the fused output.

[0011] A method for precise prostate segmentation includes the following steps: S51. Obtain the prostate MRI dataset, resample the images, perform offset field correction and intensity normalization, and convert the three-dimensional MRI scan data into a two-dimensional slice sequence. S52. Divide the preprocessed two-dimensional slice sequence into a training set and a test set, configure the training parameters, and use the Adam optimizer to train the dual-path fusion neural network. During the iteration process, optimize the model parameters through a hybrid loss function until the model converges. S53. After preprocessing the prostate MRI image to be segmented in step S51, input it into the pre-trained dual-path fusion neural network in step S52. The encoder extracts local and global features, which are then fused by the dual-stream alignment gating module. Finally, the decoder outputs the segmentation results of the prostate gland and lesion area.

[0012] Furthermore, the intensity standardization is selected as Z-Score standardization or CT standardization based on the prostate MRI dataset type; data augmentation operations are used during the training process, including random rotation, cropping, flipping, sharpening, brightness adjustment, color jittering, elastic transformation, and mesh perturbation.

[0013] Furthermore, the training parameters are configured as follows: 300 iterations, initial learning rate 1e-4, weight decay coefficient 5e-2; image patch size is adjusted to 320×320 pixels or 384×384 pixels according to the dataset, and batch size is adjusted to 32 or 22 accordingly.

[0014] Furthermore, the prostate MRI dataset includes the Prostate158 dataset, the PICAI dataset, and the MSD-Task05 Prostate dataset. The training set and test set are divided using a patient-level random allocation strategy to ensure that the training set and test set are consistent in terms of lesion distribution and anatomical structure distribution.

[0015] Furthermore, the evaluation metrics for the segmentation results include the Dice coefficient, intersection-over-union ratio, sensitivity, relative volume error, and 95% Hausdorff distance. Compared with the prior art, the beneficial effects achieved by the present invention are as follows: High segmentation accuracy: On the Prostate158 dataset, the Dice coefficient reaches 64.25%, the IoU reaches 50.94%, and the 95HD is as low as 9.84mm; on the PICAI dataset, the Dice coefficient reaches 67.50%, the IoU reaches 50.95%, and the 95HD is as low as 3.54mm, all of which are better than mainstream models such as SwinUMamba and nnUNETR, especially in terms of boundary matching.

[0016] Strong generalization ability: The model is integrated into the nnU-Net framework and performs stably on multi-center and multi-modal datasets (Prostate158, PICAI, MSD-Task05). On the MSD-Task05 dataset, the average Dice coefficient reaches 94.99%, demonstrating excellent domain adaptability.

[0017] Excellent computational efficiency: The global path uses the VSSM module with linear complexity, avoiding the high computational overhead of Transformer. The total number of model parameters is 41.6M, which is lower than that of comparative models such as SwinUMamba (60.0M) and UNETR (116.1M), ensuring both performance and computational efficiency.

[0018] Good clinical applicability: The segmentation results can accurately restore the anatomical structure of the prostate gland and lesions, which helps to reduce the burden of doctors reading images, reduce diagnostic discrepancies, and provide a reliable basis for quantitative tumor analysis and treatment planning. Attached Figure Description

[0019] Figure 1 The following are schematic diagrams of the structure of the DPF-UNet described in this invention: (a) is a schematic diagram of the structure of the DPF-Unet; (b) is a schematic diagram of the structure of the DSAG module; (c) is a schematic diagram of the structure of the VSSBlock module; and (d) is a schematic diagram of the structure of the UNetrUpBlock module. Figure 2 The following is a schematic diagram of the structure of the CARE module described in this invention; wherein: (a) is a schematic diagram of the structure of the context-aware residual encoder; (b) is a schematic diagram of the structure of the large kernel block; Figure 3 The images show a comparison of model convergence, with the right image showing the model convergence on the Prostate158 dataset and the left image showing the model convergence on the PICAI dataset. Figure 4 The image shows the segmentation results on the Prostate158 dataset; where the red areas represent the segmentation results predicted by different models, and the white outlines represent the actual lesion areas annotated by experts. Figure 5 The image shows the segmentation results visualized on the PICAI dataset; where the red areas represent the segmentation results predicted by different models, and the white outlines represent the actual lesion areas annotated by experts. Figure 6 The visualization results are shown on the Prostate158 dataset; where the colored filled areas represent the model's segmentation results, and the white outlines represent the gold standard of expert annotation. Figure 7 This is a visualization of the segmentation results on the MSD dataset; where: the colored filled areas represent the segmentation results of the model, and the white outlines represent the gold standard of expert annotation. Detailed Implementation

[0020] The present invention will be further described in conjunction with the accompanying drawings and embodiments.

[0021] As an embodiment 1 of the present invention, such as Figure 1 As shown, a dual-path fusion neural network for accurate prostate segmentation employs a U-shaped encoder and decoder structure, wherein: The encoder employs a dual-path decoupling architecture to collaboratively extract multi-scale feature representations of local details and global context from the input image; the dual-path decoupling architecture includes parallel local paths and global paths, wherein: The local path is a convolutional attention refinement encoder, which sequentially enhances the input features through residual blocks, performs at least two stages of feature refinement encoding by concatenating at least two local kernel blocks, and fuses the original and encoded features after batch normalization to capture the local texture, edge information and region-level context of the input features, i.e., local features; wherein: the local kernel block contains at least three branch convolutions and selective kernel attention; The global path is a visual state space encoder with a pre-trained VMamba model as the backbone network, including a trunk module and at least two stacked visual state space stages for capturing long-range dependencies of input features. The trunk module converts the input image into embedded features through convolutional layers and instance normalization layers. Each visual state space stage adopts a visual state space layer with a pre-normalized residual structure, which contains layer normalization, a two-dimensional selective scanning module and a deep neural network. The two-dimensional selective scanning module captures the long-range dependencies of the embedded features, i.e., global features. The skip connection is a dual-stream alignment gating module that sequentially fuses local and global features through channel alignment, hybrid feature generation, adaptive gating weight calculation, grouped convolutional refinement, and residual fusion operations. The decoder uses cascaded upsampling operations to gradually restore spatial resolution in order to generate segmentation results.

[0022] We will delve into the two core components that make up this framework: the Dual Stream Aligned Gating (DSAG) module for intelligent fusion of dual-path features and the Convolutional Attention Refinement Encoder (CARE) for capturing local and regional features.

[0023] Figure 1(a) illustrates the overall architecture of DPF-UNet. This model employs a classic U-shaped encoder and decoder structure, with its core innovation being the introduction of a parallel dual-path encoder, designed to process the input image X∈R. C×H×W The proposed dual-path design extracts multi-scale feature representations of local details and global context in a collaborative manner. Compared with the traditional U-Net which relies on single-path feature extraction, this invention can more comprehensively capture the multi-scale features of prostate lesions.

[0024] The CARE module is used in local paths, focusing on local texture, edge information and region-level context to support the fine depiction of lesion structures.

[0025] The visual state-space VSSM encoder, constructed using a pre-trained VMamba model as the backbone network in the global path, is shown in Figure 1(c). This path aims to overcome the inherent limitations of convolutional neural networks in modeling long-range spatial dependencies, efficiently establishing long-range pixel associations with near-linear computational complexity, and avoiding the quadratic computational overhead of the self-attention mechanism in the Vision Transformer. The VSSM encoder path begins with a trunk module, which converts the input image into embedded features through a 4×4 convolutional layer with a stride of 4 and an instance normalization layer. The encoder body consists of multiple stacked VSSM stages, with the core of each stage being a VSS layer employing a pre-normalized residual structure. ; Wherein: F SS2D This represents a two-dimensional selective scanning module, responsible for efficiently modeling long-range spatial dependencies; DropPath is a random depth operation, often used for regularization to prevent overfitting. It is the input feature map. The output feature map after passing through the module.

[0026] This module splits the input features into a main stream and a gated stream. The main stream is enhanced by local convolution and then enters a four-directional selective scanning mechanism. The feature map is flattened into a sequence in four directions, which is then processed by the selective scanning state-space model. , ; Where t is the time step or spatial location index of the sequence. and These represent the input and output of the sequence at time t, respectively. and Representing the current Time and the previous The hidden state at any given moment. This is the Skip Connection parameter, used to project input features directly onto the output. , and They are composed of continuous system parameters ( Discretization methods such as Zero-Order Preservation (ZOH) are used, combined with input-dependent step sizes. The calculated discrete system matrix ( It is an input-dependent step size parameter used to control the degree to which current information is preserved during discretization. Discretization parameters , and Dynamically generated from the input sequence to achieve adaptive information delivery. The outputs from the four directions are aggregated and fused with gated stream modulation. ; ; ; in: This represents the core features after aggregation. This originates from the initially separated gated flow. It is the gating signal after activation. This usually refers to the activation function. These are the backbone features after layer normalization. It is the output projection function (usually a linear projection layer) used to adjust channel dimensions or fuse features. It is a feature fusion operation (usually element-wise multiplication* or element-wise addition, but in gating structures it usually refers to the Hadamard product, which corresponds to element-wise multiplication).

[0027] To improve model performance, we use a VMamba model pre-trained on ImageNet as the backbone network for this path, thereby injecting rich general visual prior knowledge into the model.

[0028] The dynamic feature fusion stage is achieved through the DSAG module, such as... Figure 1 As shown in Figure (b), the core idea is to adaptively adjust the feature contributions from the two paths through a gating mechanism, thereby alleviating the potential feature inconsistency problem in the multi-path structure and constructing an enhanced skip connection.

[0029] The decoder stage adopts a UNETR-style design, such as... Figure 1 As shown in Figure (d), spatial resolution is gradually restored through cascaded upsampling operations, ultimately generating the segmentation result. A deep supervision strategy is employed during training to promote efficient gradient propagation, accelerate model convergence, and improve overall training stability.

[0030] 1. Dual Stream Alignment Gating Module (DSAG): The overall architecture of DSAG, such as Figure 1 As shown in Figure (b), it is responsible for intelligently fusing heterogeneous features from the CARE and VSSM encoders at each scale of the decoder. Its workflow can be broken down into the following steps: At the l-th scale, the DSAG module receives features from two encoding paths: features output from the CARE path... and the characteristics of VSSM path output. DSAG performs the following four steps sequentially to achieve deep fusion of cross-path features: First, in the feature alignment stage, a 1×1 convolutional layer ϕ is used to adjust the number of channels in the VSSM feature to align it with the CARE feature: ; in: ( ) here specifically refers to the part used for channel alignment. Convolution mapping function.

[0031] Next, in the hybrid feature generation stage, the channel-aligned VSSM features and CARE features are added element-wise and fused into a hybrid feature map that combines local details and global context. ; These represent the feature maps to be fused from the global visual state space path and the local convolutional attention refinement path, respectively.

[0032] Subsequently, in the adaptive gating stage, a lightweight gating mechanism is designed to dynamically calibrate the importance of different feature groups. This mechanism first applies the mixed features... Perform global average pooling (GAP) This represents the global average pooling operation, used to compress spatial information into channel descriptors. The compressed information is then passed through two consecutive 1×1 convolutional layers. and And non-linear activation, finally through the Sigmoid function Generate grouped gating weights Where: Ng is the preset number of groups: ; ; ; ; ; Finally, in the gating refinement and fusion stage, the hybrid features are first... Divide the data along the channel dimension into Ng groups, as defined below: ; Each group is processed by a separate 3×3 group convolution. Perform local structural refinement and correlate it with the corresponding gating weights. Multiplication is performed to adaptively enhance useful features. Subsequently, all weighted group features are concatenated and then added to the original mixed features via residual connections to obtain the final fusion output at the current scale. The calculation process is as follows: ; ; ; Through the above process, the DSAG module achieves intelligent fusion of local details and global dependencies at different scales, enabling the decoder to dynamically select and strengthen the most discriminative features based on semantic content, thereby providing highly information-rich jump connection features for accurate segmentation.

[0033] 2. Convolutional Attention Refined Encoder (CARE): In the deeper stages of the encoder, feature representation often faces problems such as loss of detail and semantic ambiguity. We designed a Convolutional Attention Refinement Encoder (CARE), the structure of which is as follows: Figure 2 As shown in (a) of the diagram. This module combines the stability of static convolution with the adaptability of dynamic convolution through a carefully designed path, enabling refined processing of complex features.

[0034] The main structure of the CARE module consists of three key parts: a standard residual block, two consecutive large kernel blocks (LKBlock), and a residual connection path. This hierarchical design ensures smooth gradient propagation and enhances the discriminative power of features. Its forward propagation process can be described as follows: ; ; ; ; in, The transformation function representing the residual block. This represents the operation of the k-th LKBlock, where k is a natural number from 1 to 10. LKBlock is the core of adaptive feature extraction. Its computation process first generates dynamic weights through the Large Kernel Path (LKP), then performs dynamic convolution through the Spatial Kernel Attention (SKA) module, and finally completes feature fusion through residual connections. This structure cleverly combines dynamic convolutional paths with identity mappings: the dynamic paths capture content-specific features, while residual connections ensure information integrity. This design not only enhances the model's expressive power but also ensures training stability.

[0035] The Large Kernel Path (LKP) is key to achieving dynamic convolution, which generates spatially variable convolution kernels step by step through a three-stage process.

[0036] First, in the feature preprocessing stage, the input feature X undergoes a preliminary transformation through a standard convolution-normalization combination. This process begins with a weighted... and bias The defined convolution operation's output is then fed into a batch normalization layer. This layer utilizes the statistics (mean) of the current batch B. and variance The features are normalized using learnable affine parameters (gain γ and bias β). Finally, the output features X for this stage are generated using a non-linear activation function (e.g., ReLU). cv1 .

[0037] Next, to efficiently capture global contextual information, we employ a large-kernel depthwise convolution layer. The core of this operation lies in its use of the input feature X... cv1 Each channel c is assigned an independent L×L spatial convolution kernel. And convolution is performed only within that channel. This channel-by-channel processing effectively extracts spatial relationships while maintaining low computational complexity. This operation for the c-th channel can be represented as: ; Similar to the previous stage, the output Y of the depthwise convolution... dw Similarly, it undergoes batch normalization and nonlinear activation function sequentially. The processing ultimately generates a feature map X containing long-range dependency information. cv2 Then, in the kernel parameter generation stage, the original kernel parameters are generated by compressing the channel dimensions through 1×1 convolution: ; Finally, the spatially varying kernel weights W = View(W′) are obtained through tensor reshaping operations. Taking an input size of [B,C,H,W] as an example, the kernel tensor dimension output by LKP is [B,C,S2,H,W], where S is the dynamic convolution kernel size. The Spatial Kernel Attention (SKA) module performs the actual dynamic convolution operation. For each spatial location (h, w), it extracts the corresponding S×S convolution kernel from W and performs a weighted summation within the local neighborhood of the input features: ; Where δ=⌊S / 2⌋ ensures center alignment of the convolution operation. This mechanism gives the CARE module unique content-aware capabilities: in edge regions, it can enhance contour responses; in regions with complex textures, it can adaptively adjust the receptive field to focus on the most discriminative features.

[0038] Experiments show that this design performs exceptionally well in handling fine-grained visual tasks, demonstrating a significant improvement in detail preservation compared to traditional convolutional methods. Through this "static structure, dynamic parameters" design philosophy, the CARE module greatly enhances the model's feature representation capabilities while maintaining computational efficiency.

[0039] 3. Loss Function: To address the challenges of imbalanced samples and complex boundary conditions in prostate segmentation, we employ a strategy combining hybrid loss functions with deep supervision. The single-scale loss consists of Soft Dice and cross-entropy (CE): ; Among them, LDice focuses on optimizing the spatial overlap of small targets (such as lesions): ; In the formula: N and K represent the total number of voxels and classes, respectively; pi,k and gi,k represent the predicted probability and the true label, respectively. Complementary LCE is used to stabilize gradients and refine anatomical boundaries. By introducing deep supervision, the total loss is defined as the weighted sum of the losses at each resolution level: ; Where: weight It decays as resolution decreases in order to enhance structural reconstruction at high resolution.

[0040] Experiments demonstrate that this composite supervision strategy not only accelerates model convergence but also significantly improves segmentation robustness at different scales. Particularly when dealing with the complex boundary between the central and peripheral regions of the prostate, the model exhibits excellent anatomical structure recognition capabilities. 4. Experiment: Table 1: Quantitative results of lesion segmentation on the Prostate158 dataset (%)

[0041] Table 2: Quantitative results of lesion segmentation on the PICAI dataset (%)

[0042] 1. Dataset Description: To systematically evaluate the model's performance in prostate cancer lesion segmentation and differentiation of internal and external glands, three complementary multicenter prostate mp-MRI datasets were selected. A hierarchical evaluation system was constructed from three levels: annotation granularity, imaging protocol, and clinical applicability. The Prostate158 dataset contains 158 3T dual-parametric MRIs from multiple centers. This dataset was used for dual-task evaluation of lesion and internal / external gland segmentation. Patient-level randomization (seed=42) was used for partitioning, with 138 cases used for training and 28 for testing, ensuring no cross-patient data leakage and maintaining a balance in lesion distribution between the training and test sets. The MSD-Task05 Prostate dataset contains 48 prostate MRIs. We selected 30 valid cases from the official training set, with 25 used for training and 5 for testing. The partitioning also followed patient-level randomization (seed=42), maintaining consistency in anatomical structure distribution between the training and test sets. The PI-CAI dataset contains 1500 dual-parametric MRIs. This study randomly selected 300 cases (250 for training and 50 for testing) using a patient-level randomization strategy (seed=42) to ensure that the proportion of positive samples remained largely consistent between the training and testing sets. All data underwent standardized preprocessing before use, including resampling, bias field correction, and intensity normalization, as detailed in the next section.

[0043] 2. Experimental setup: All experiments were conducted using Python 3.12.9 and PyTorch 2.3.0, with an NVIDIA RTX 4080 GPU equipped with 16GB of VRAM. Network training used the Adam optimizer for 300 iterations, with an initial learning rate of 1e-4 and a uniform weight decay of 5e-2. In experiments on different datasets, we employed a 2D slice-based analysis method, treating 3D MRI scans as a series of 2D images. During preprocessing, the in-plane resolution of all images was uniformly resampled: 0.40×0.40 mm for Prostate158, 0.625×0.625 mm for the MSD prostate task, and 0.5×0.5 mm for the PI-CAI dataset. The input modalities for each dataset were independently standardized: Z-Score standardization was used for Prostate158 and the MSD task, while PICAI additionally explored various schemes including Z-Score and CT standardization. During the training phase, Instance Normalization and LeakyReLU activation functions were used uniformly, but there were differences in image patch size and batch size: Prostate158 and MSD tasks both used 320×320 pixel image patches with a batch size of 32; while PI-CAI used 384×384 pixel image patches with a batch size of 22.

[0044] During training, images were preprocessed using the NUNET framework, including random rotation, cropping, flipping, sharpening, brightness adjustment, color jittering, elastic transformation, mesh perturbation, scaling, and normalization. Model performance was evaluated using Dice coefficients, Intersection over Union (IoU), 95% Hausdorff distance (HD95), Sens, and RVE. All experimental models used 2D datasets preprocessed using the NUNET framework.

[0045] 3. Experimental Results: To comprehensively evaluate model performance, we conducted systematic testing on four publicly available prostate MRI datasets, including Prostate158, PICAI, and the prostate segmentation task in the Medical Segmentation Decathlon (MSD). Experimental results show that our proposed method demonstrates significant advantages in several key dimensions—including segmentation accuracy, boundary fit, and cross-dataset generalization ability.

[0046] On the Prostate158 dataset, as shown in Table 1, our model achieved the best segmentation performance, with Dice coefficient and IoU reaching 64.25% and 50.94% respectively, ranking first in both metrics. Compared to the second-best performing SwinUMamba, the Dice value is improved by approximately 2.46%. More notably, the model excels in boundary delineation, with a 95% Hausdorff distance (95HD) as low as 9.84mm, indicating its superior ability to capture subtle lesion contours. Testing on the PICAI Challenge dataset further validates the competitiveness of our model, as shown in Table 2. It comprehensively leads in four key metrics: Dice (67.50%), IoU (50.95%), sensitivity (72.96%), and 95HD (3.54mm). Compared to SwinUMamba, Dice is improved by 2.06%, and the significant advantage of 95HD particularly demonstrates that the model can still achieve accurate boundary localization in complex anatomical regions.

[0047] We further evaluated the model's segmentation performance in different functional regions of the prostate—the central gland (CG) and the peripheral zone (PZ), as shown in Table 3. In whole gland segmentation, the model maintained its leading position with 83.74% Dice and 72.86% IoU. Particularly in the structurally regular CG region, our method ranked first on multiple metrics; while in the more blurred and morphologically complex PZ region, the model remained robust, achieving a 95HD of 4.73 mm, superior to all comparison models, demonstrating reliable boundary recognition capability even for difficult samples.

[0048] Finally, we tested the generalization performance of our model on the MSD prostate segmentation benchmark, as shown in Table 4. Our method achieved an average Dice score of 94.99% and an IoU score of 90.53%, outperforming several mainstream models, including nnUNet and SwinUNETR.

[0049] Specifically, in the CG segmentation task, we achieved the best Dice (93.58%) and 95HD (0.78mm) simultaneously; in the more challenging PZ region, the model also performed excellently, with a very small difference from the best results. These experiments consistently demonstrate that our method maintains high accuracy and strong robustness on data from different sources and with varying annotation specifications.

[0050] Figure 3 The curves showing the training loss of different models on four datasets as a function of epochs are presented. It can be seen that in the early stages of training (approximately the first 100 epochs), all models exhibit a rapid decline in loss, which then gradually levels off, eventually converging to a low and stable level. This convergence process reflects the robustness of the model training, corroborating the aforementioned quantitative performance evaluation and further supporting the reliability of our model in practical applications.

[0051] 4. Visual analysis: To more intuitively evaluate DPFUNet's performance in segmentation tasks, we visualized its output on multiple prostate MRI datasets, focusing on segmentation integrity and boundary recognition capabilities.

[0052] Figure 4 and Figure 5 The results show the segmentation performance of each model on the Prostate158 and PICAI datasets for lesion regions (marked in red).

[0053] It can be observed that DPFUNet, while maintaining region integrity, also restores edge details more accurately. Figure 4 Taking a lesion with an irregular concave structure as an example, the segmentation results generated by DPFUNET are highly consistent with the real contour, effectively alleviating the problems of over-smoothing or local omissions that are common in other methods, showing its good adaptability to complex morphologies.

[0054] Figure 6The performance of each model in simultaneously segmenting the entire gland (green) and lesions (red) on Prostate158 was further compared. UNETR showed significant undersegmentation in both representative cases, failing to cover the entire target region. In contrast, DPFUNet demonstrated stronger robustness, with its output closely matching the gold standard (white outline) for both the overall structure of the gland and the detailed contours of the lesions, indicating that the model has significant advantages in feature extraction and structural understanding.

[0055] exist Figure 7 In the MSD dataset partitioning results shown, DPFUNet also performed stably. For the elongated and curved central gland (CG, green), the model was able to accurately restore its anatomical shape, avoiding the "boundary bloat" phenomenon that occurs in other models; and for the lesion areas in the peripheral zone (PZ, red), DPFUNet also achieved high spatial coverage and boundary fit.

[0056] Based on visualization results from multiple datasets and segmentation tasks, DPFUNet demonstrates excellent generalization ability across data with different anatomical structures and imaging characteristics, especially in complex boundary reconstruction and structural integrity preservation, where it exhibits significant advantages.

[0057] Table 3: Quantitative results (%) of partitioning (CG & PZ) on the Prostate158 dataset

[0058] Table 4: Quantitative results (%) of prostate region segmentation (CG&PZ)

[0059] 5. Ablation test: A series of ablation experiments were conducted on the three datasets mentioned above to analyze the impact of different factors on model performance and to verify the design rationality of DPFUNet. The contributions of the CARE module, VSSM module, and DSAG module to model performance were evaluated. Furthermore, ablation studies were performed on the Prostate158 and PICAI datasets with different configurations of the CARE module. Each experiment ran for 150 epochs. Tables 5 and 6 show the contributions of the pre-trained weights (PW), MSGA module, VSSM module, and DSAG module to model performance. The results show that removing any one of the pre-trained weights (PW), MSGA module, VSSM module, or DSAG module significantly reduces the performance of DPF-UNet. The performance drop was greatest when the VSSM module was removed. These results indicate that the PW, MSGA, VSSM, and DSAG modules all have a positive effect on performance improvement, validating the rationality of the design of these key modules.

[0060] Table 5: Ablation test results of major components (%)

[0061] Table 6: Ablation test results of major components (%)

[0062] 6. Conclusion: Despite achieving competitive segmentation performance on multiple public datasets, this study still has several noteworthy limitations. First, the current model is built on 2D slices, which significantly improves computational efficiency but struggles to fully model the structural continuity and invasive characteristics of lesions in 3D space. Existing research has shown that 3D modeling can improve the Dice coefficient of prostate cancer segmentation by 3-5%. Future work will extend the dual-path architecture to 3D to enhance cross-slice context modeling capabilities, which is expected to further improve segmentation accuracy in tasks such as small lesion detection. Second, the current method is primarily based on mp-MRI single-modal data, failing to fully utilize the complementary information provided by multimodal images such as ultrasound and CT. This limits the model's performance in cases with poor image quality or atypical lesion features. Future research will introduce a cross-modal interaction module to enhance boundary recognition capabilities through feature-level fusion, thereby improving the model's adaptability in complex clinical scenarios.

[0063] Furthermore, the training data used in this study still has room for improvement in terms of sample size and annotation consistency. Future research will focus on expanding the data scale through multi-center collaboration and introducing semi-supervised learning strategies to alleviate the problem of insufficient labeled data.

[0064] The proposed DPF-UNet model, through a dual-path encoder and gating fusion mechanism, effectively coordinates local details and global semantic dependencies, achieving accurate segmentation of prostate cancer MRI. Experiments on public datasets such as Prostate158, PICAI, and MSD-Task05 demonstrate that this method outperforms other methods in key metrics such as Dice coefficient, intersection-over-union ratio, and 95% Hausdorff distance. Specifically, the Dice coefficient on the Prostate158 dataset reaches 91.2%, surpassing baseline methods such as SwinUNETR and nnU-Net. Figure 5 , 6As shown in Figure 7, the high-precision segmentation achieved by this method helps reduce the burden on doctors in reading images, decreases diagnostic discrepancies, and provides a reliable basis for quantitative tumor analysis and treatment planning. The implementation based on the nnU-Net framework ensures the clinical applicability and reproducibility of the model. Overall, DPFUNet promotes the development of intelligent segmentation for prostate cancer and provides a solution with both theoretical innovation and practical value for related medical image analysis tasks. Future work will focus on exploring 3D architecture design and multimodal feature fusion to further improve the model's performance in complex clinical scenarios.

[0065] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A dual-path fusion neural network for precise prostate segmentation, employing a U-shaped encoder and decoder structure, characterized in that: The encoder employs a dual-path decoupling architecture to collaboratively extract multi-scale feature representations of local details and global context from the input image; the dual-path decoupling architecture includes parallel local paths and global paths, wherein: The local path is a convolutional attention refinement encoder, which sequentially enhances the input features through residual blocks, performs at least two stages of feature encoding by concatenating at least two local kernel blocks, and fuses the original and encoded features after batch normalization to capture the local texture, edge information and region-level context of the input features, i.e., local features; wherein: the local kernel block contains at least three branch convolutions and selective kernel attention; The global path is a visual state space encoder with a pre-trained VMamba model as the backbone network, including a trunk module and at least two stacked visual state space stages for capturing long-range dependencies of input features. The trunk module converts the input image into embedded features through convolutional layers and instance normalization layers. Each visual state space stage adopts a visual state space layer with a pre-normalized residual structure, which contains layer normalization, a two-dimensional selective scanning module and a deep neural network. The two-dimensional selective scanning module captures the long-range dependencies of the embedded features, i.e., global features. The skip connection is a dual-stream alignment gating module that sequentially fuses local and global features through channel alignment, hybrid feature generation, adaptive gating weight calculation, grouped convolutional refinement, and residual fusion operations. The decoder uses cascaded upsampling operations to gradually restore spatial resolution in order to generate segmentation results.

2. The dual-path fusion neural network for precise prostate segmentation according to claim 1, characterized in that: The loss function of the dual-path fusion neural network adopts a hybrid loss function, including single-scale loss and deep supervision loss, wherein: The single-scale loss is a weighted sum of the Soft Dice loss and the cross-entropy loss; The depth supervision loss is a weighted sum of the single-scale losses at each resolution level.

3. The dual-path fusion neural network for precise prostate segmentation according to claim 1, characterized in that: The convolutional attention refinement encoder includes residual blocks arranged in series, at least two feature refinement stages consisting of at least two local kernel blocks arranged in series, a batch normalization layer, and residual connections, wherein: Residual blocks extract initial features through convolutional structures with residual connections, while mitigating the gradient vanishing problem in deep networks and enhancing the expressive power of initial features. At least two feature refinement stages are used to perform multi-scale modeling and channel attention enhancement on the initial features using at least two concatenated local kernel blocks to achieve refined representation of local features; wherein: each local kernel block includes at least three branch convolutions for capturing local multi-scale features with different receptive fields, a View layer for unifying feature dimensions, and a selective kernel attention mechanism for dynamically learning and strengthening local feature weights; Batch normalization layers are used to unify feature scales to enhance the consistency of feature representation; Residual connections fuse the original input features with the refined encoded features element-wise to output the enhanced encoded features.

4. The dual-path fusion neural network for precise prostate segmentation according to claim 3, characterized in that: The specific process of the dual-stream alignment gating module is as follows: Channel alignment: Adjust the number of channels in the global path output features using a 1×1 convolutional layer to make it consistent with the number of channels in the local path output features; Hybrid feature generation: The channel-aligned global path output features are added element-wise to the local path output features to obtain a hybrid feature map; Adaptive gating weight calculation: Global average pooling is performed on the mixed feature map, followed by two 1×1 convolutions and non-linear activation, and then grouped gating weights are generated by the Sigmoid function; Grouped convolutional refinement: The mixed feature map is divided into a preset number of feature groups along the channel dimension. Each group is refined locally through independent 3×3 group convolutions and multiplied with the corresponding gate weights. Residual fusion: All weighted group features are concatenated and added to the original mixed feature map through residual connections to obtain the fused output.

5. A method for precise prostate segmentation, characterized in that: Includes the following steps: S51. Obtain the prostate MRI dataset, resample the images, perform offset field correction and intensity normalization, and convert the three-dimensional MRI scan data into a two-dimensional slice sequence. S52. Divide the preprocessed two-dimensional slice sequence into a training set and a test set, configure training parameters, and use the Adam optimizer to train the dual-path fusion neural network as described in any one of claims 1 to 4. During the iteration process, optimize the model parameters through a hybrid loss function until the model converges. S53. After preprocessing the prostate MRI image to be segmented in step S51, input it into the pre-trained dual-path fusion neural network in step S52. The encoder extracts local and global features, which are then fused by the dual-stream alignment gating module. Finally, the decoder outputs the segmentation results of the prostate gland and lesion area.

6. The method for precise prostate segmentation according to claim 5, characterized in that: The intensity standardization is selected as Z-Score standardization or CT standardization based on the prostate MRI dataset type; data augmentation operations are used during the training process, including random rotation, cropping, flipping, sharpening, brightness adjustment, color jittering, elastic transformation, and mesh perturbation.

7. The method for precise prostate segmentation according to claim 6, characterized in that: The training parameters are configured as follows: 300 iterations, initial learning rate 1e-4, weight decay coefficient 5e-2; image patch size is adjusted to 320×320 pixels or 384×384 pixels according to the dataset, and batch size is adjusted to 32 or 22 accordingly.

8. The method for precise prostate segmentation according to claim 7, characterized in that: The prostate MRI datasets include the Prostate158 dataset, the PICAI dataset, and the MSD-Task05 Prostate dataset. The training and test sets are divided using a patient-level random allocation strategy to ensure that the training and test sets are consistent in terms of lesion distribution and anatomical structure distribution.

9. A method for precise prostate segmentation according to claim 8, characterized in that: The evaluation metrics for the segmentation results include Dice coefficient, crossover-union ratio, sensitivity, relative volume error, and 95% Hausdorff distance.

Images