An efficient ultrasound image segmentation method and system based on a two-path heterogeneous encoder

Abstract

The application discloses a kind of high-efficiency ultrasonic image segmentation method and system based on two-way heterogeneous encoder, it is related to computer vision technical field.The method comprises: constructing two-way heterogeneous coding network, and training using hybrid loss function, the ultrasonic image to be segmented is input into the two-way heterogeneous coding network trained to carry out feature processing, and the final segmentation prediction graph is obtained.Wherein, the two-way heterogeneous coding network includes parallel structural branch encoder and detail branch encoder, and the decoder connected with the structural branch encoder and the detail branch encoder, and the detail branch encoder is also connected with the feature enhancement module of detail guide, and the decoder is also connected with boundary guide module.The application can extract and fuse global structure information and local detail information in the segmentation process of ultrasonic image, and improve the segmentation accuracy while considering the efficiency of model deployment.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a high-efficiency ultrasonic image segmentation method and system based on a dual-channel heterogeneous encoder. Background Technology

[0002] Ultrasound imaging has become a routine tool in clinical diagnosis due to its advantages of real-time operation, radiation-free operation, and low cost. However, the inherent low contrast, speckle noise, duct shadow interference, and blurred tissue boundaries of ultrasound images make accurate and automated image segmentation extremely challenging. Currently, deep learning-based automatic segmentation methods have become the mainstream approach for medical image segmentation. Existing technologies mainly revolve around two major architectures: Convolutional Neural Network (CNN) and Transformer architecture.

[0003] Among them, the CNN architecture, represented by U-Net and its many variants, has established performance benchmarks in various medical image segmentation tasks due to its powerful local feature extraction capabilities. However, the receptive field of the convolutional operations in the CNN architecture is limited by the size of the convolutional kernel, making it difficult for the network to model long-distance contextual dependencies. For tissues or lesions of varying sizes and shapes in ultrasound images, a single-path encoder struggles to collaboratively capture the complete feature spectrum from global structure to local details. Furthermore, traditional encoder-decoder structures fuse features through skip connections, but this fusion is usually a simple concatenation or addition of multi-scale features, lacking a refined and adaptive fusion mechanism for features of different levels and properties, resulting in insufficient information exchange and limited improvement in boundary segmentation accuracy. The Transformer architecture overcomes the locality limitation of CNNs through its self-attention mechanism, effectively modeling global feature associations and showing potential in complex structure segmentation. However, the Transformer's self-attention mechanism has secondary computational complexity, leading to slow model training and inference speeds and high memory consumption. It also has weak local detail perception capabilities; the Transformer lacks the spatial inductive bias inherent in CNNs, resulting in relatively weaker ability to capture local textures, edges, and other detailed features. The diagnostic value of ultrasound images depends precisely on these subtle details and boundary information.

[0004] Existing hybrid models have failed to fundamentally solve the problem of balancing efficiency and accuracy: To combine the advantages of CNNs and Transformers, various hybrid models have been proposed in the industry. For example, DDTransUNet employs a dual-branch (Swin Transformer and CNN) encoder and a dual attention mechanism; TransUNet deeply integrates CNNs and Transformers. However, both CNN and Transformer modules contain a large number of parameters for model training, resulting in limited lightweighting and a single feature fusion method, with the fusion precision needing improvement.

[0005] Currently, whether it is pure CNN, pure Transformer, or existing hybrid models, none of them have achieved a satisfactory balance in the three dimensions of comprehensive feature extraction (coordinating global structure and local details), effectiveness of feature fusion (adaptive interaction of hierarchical and heterogeneous features), and feasibility of model efficiency (meeting the needs of real-time clinical deployment while maintaining high accuracy) when dealing with ultrasound image segmentation. Summary of the Invention

[0006] The purpose of this invention is to provide a high-efficiency ultrasonic image segmentation method and system based on a dual-channel heterogeneous encoder, aiming to solve or improve at least one of the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following solution: An efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder includes: Obtain the ultrasound image to be segmented; A dual-path heterogeneous coding network is constructed and trained using a hybrid loss function to obtain a trained dual-path heterogeneous coding network. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing on the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive high-level features from the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features from the decoder to refine the segmentation boundaries. The ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map.

[0008] Optionally, the ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map. The specific processing steps include: Step 1: Obtain the ultrasound image to be segmented; Step 2: Input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network, and extract features in parallel to obtain global structural features and local detail features respectively; Step 3: Input the output features of the first three layers of the detail branch encoder into the detail-guided feature enhancement module in sequence for feature enhancement and aggregation to obtain enhanced local detail features; Step 4: Input the third-layer output features of the structural branch encoder and the detail branch encoder into their respective subsequent encoders containing the Token_MLP module for further processing, and obtain the independent last two layers of features, namely the fourth-layer output features and the fifth-layer output features. Step 5: First, perform shallow fusion of the first three layers of output features of the structural branch encoder and the enhanced local detail features using an adaptive kernel selection fusion method to obtain the fused features from the first to the third layer. Then, perform deep fusion of the last two layers of features of the structural branch encoder and the last two layers of features of the detail branch encoder using a channel adaptive convolution fusion method to obtain the fused features of the fourth and fifth layers. Step 6: Input the fusion features of the fifth layer into the decoder for decoding, and optimize the features through the boundary guidance module. After upsampling, aggregate with the fusion features of the fourth layer. After decoding the aggregated features, further optimize them through the boundary guidance module. After upsampling, aggregate with the fusion features of the third layer again. Repeat the optimization and aggregation process, and process with the fusion features of the second and first layers in turn to complete image reconstruction and obtain the final segmentation prediction map. Step 7: Output the final segmentation prediction map.

[0009] Optionally, the specific processing steps of the detail-guided feature enhancement module include: Edge information extraction and multi-scale depth-separable convolution operation are performed on the first layer output features of the detail branch encoder to obtain the first enhanced feature; The first enhanced feature is downsampled and added to the second layer output feature of the detail branch encoder, and the same enhancement operation is performed on the addition result to obtain the second enhanced feature; The second enhanced feature is downsampled and added to the third layer output feature of the detail branch encoder, and the same enhancement operation is performed on the addition result to obtain the third enhanced feature; The first enhanced feature, the second enhanced feature, and the third enhanced feature are used as enhanced local detail features.

[0010] Optionally, the adaptive kernel selection fusion method is adopted, and the specific processing steps include: The first three layers of the output features of the structural branch encoder at the same level and the enhanced local detail features are used as independent inputs. The two branch features are added element by element and then global average pooling is performed. Selection weights corresponding to the number of channels of the two branches are generated through a multilayer perceptron network. Using the selection weights, the output features of the structural branch encoder and the enhanced local detail features are respectively channel-weighted, and then the weighted features are added together to obtain the fused features from the first layer to the third layer.

[0011] The channel-adaptive convolutional fusion method specifically includes the following processing steps: The output features of the last two layers of the structural branch encoder and the output features of the last two layers of the detail branch encoder at the same level are used as independent inputs and concatenated along the channel dimension at the same level to form joint features. The global average pooling vector and the global max pooling vector of the joint features are calculated respectively and processed separately through a shared multilayer perceptron. The two processed vectors are added together and passed through an activation function to generate a channel attention vector for reweighting the joint features. The channel attention vector is multiplied channel by channel with the joint feature to obtain the channel enhancement feature. The channel enhancement feature is then convolved to fuse spatial information and restore the number of channels, resulting in the fused features of the fourth and fifth layers.

[0012] Optionally, the structural branch encoder extracts features through parallel branches comprising a first standard convolutional layer, a dilated convolutional layer with a first dilation rate, and a standard convolutional layer with a first size, and fuses them via a channel and spatial attention mechanism to obtain global structural features; the detail branch encoder extracts features through parallel branches comprising a second standard convolutional layer, a dilated convolutional layer with a second dilation rate, and a standard convolutional layer with a second size, and fuses them via a channel and spatial attention mechanism to obtain local detail features; wherein, the first dilation rate is greater than the second dilation rate, and the first size is greater than the second size.

[0013] Optionally, the hybrid loss function is specifically expressed as: in, For pixel-level cross-entropy loss, and Jointly represented as regional losses, These are the weighting coefficients for each item.

[0014] This invention also provides a high-efficiency ultrasound image segmentation system based on a dual-channel heterogeneous encoder, comprising: Image acquisition unit, used to acquire ultrasound images to be segmented; A network training unit is used to construct a dual-path heterogeneous coding network and train it using a hybrid loss function to obtain a trained dual-path heterogeneous coding network. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing of the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive high-level features from the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features from the decoder to refine the segmentation boundaries. The image segmentation unit is used to input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses an efficient ultrasound image segmentation method and system based on a dual-path heterogeneous encoder. The method includes acquiring an ultrasound image to be segmented; constructing a dual-path heterogeneous coding network and training the network using a hybrid loss function to obtain a trained network; wherein the dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the encoders, the detail branch encoder being connected to a detail-guided feature enhancement module, and the decoder to a boundary guidance module; the structural branch encoder extracts global structural features; the detail branch encoder extracts local detail features; the detail-guided feature enhancement module performs sequential processing of the first three layers of the detail branch encoder's output from shallow to deep layers; the boundary guidance module receives high-level features from the decoder and generates a boundary saliency map, and performs gated fusion with upsampled features from the decoder to refine the segmentation boundary; the ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain a final segmentation prediction map. This invention can selectively extract and fuse global structural information and local detail information during the segmentation of ultrasound images, thereby improving segmentation accuracy while taking into account model deployment efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the efficient ultrasonic image segmentation method based on a dual-channel heterogeneous encoder according to the present invention; Figure 2 This is a schematic diagram of the dual-path heterogeneous coding network model architecture in this embodiment; wherein, (a) is a schematic diagram of the coding stage process; and (b) is a schematic diagram of the decoding stage process. Figure 3 This is a schematic diagram of the visualization results on the public dataset in this embodiment; wherein, (a) is the encoder heatmap; and (b) is a schematic diagram of the segmentation results. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a high-efficiency ultrasonic image segmentation method and system based on a dual-channel heterogeneous encoder, aiming to solve or improve at least one of the above-mentioned technical problems.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] As a first aspect, such as Figure 1 As shown, this invention provides a highly efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder, comprising: A dual-path heterogeneous coding network is constructed and trained using a hybrid loss function. The ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing of the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive features from each layer of the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features of the decoder to refine the segmentation boundary.

[0022] As one specific implementation method, the process of the above steps is as follows: Step 1: Obtain the ultrasound image to be segmented.

[0023] Step 2: Input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network, and extract features in parallel to obtain global structural features and local detail features. Specifically, the structural branch encoder extracts features through parallel branches containing a first standard convolutional layer, a dilated convolutional layer with a first dilation rate, and a standard convolutional layer with a first size, and then refines the layer-by-layer features using channel and spatial attention mechanisms to obtain global structural features. The detail branch encoder extracts features through parallel branches containing a second standard convolutional layer, a dilated convolutional layer with a second dilation rate, and a standard convolutional layer with a second size, and then refines the layer-by-layer features using channel and spatial attention mechanisms to obtain local detail features. The first dilation rate is greater than the second dilation rate, and the first size is greater than the second size.

[0024] Step 3: The output features of the first three layers of the detail branch encoder are sequentially input into the detail-guided feature enhancement module for feature enhancement and aggregation to obtain enhanced local detail features. The specific processing steps of the detail-guided feature enhancement module include: Edge information extraction and multi-scale depth-separable convolution operations are performed on the first layer output features of the detail branch encoder to obtain a first enhanced feature; the first enhanced feature is downsampled and added to the second layer output features of the detail branch encoder, and the same enhancement operation is performed on the added result to obtain a second enhanced feature; the second enhanced feature is downsampled and added to the third layer output features of the detail branch encoder, and the same enhancement operation is performed on the added result to obtain a third enhanced feature; the first enhanced feature, the second enhanced feature, and the third enhanced feature are used as the enhanced local detail features.

[0025] Step 4: Input the third-layer output features of the structural branch encoder and the detail branch encoder into their respective subsequent encoders containing the Token_MLP module for further processing, and obtain the independent last two layers of features, namely the fourth-layer output features and the fifth-layer output features.

[0026] Step 5: First, perform shallow fusion of the first three layers of output features of the structural branch encoder and the enhanced local detail features using an adaptive kernel selection fusion method to obtain the fused features from the first to the third layer. Then, perform deep fusion of the last two layers of features of the structural branch encoder and the last two layers of features of the detail branch encoder using a channel adaptive convolution fusion method to obtain the fused features of the fourth and fifth layers.

[0027] The adaptive kernel selection fusion method specifically includes the following processing steps: The first three layers of the structural branch encoder at the same level and the enhanced local detail features are used as independent inputs. The two branch features are added element-wise and then global average pooling is performed. Selection weights corresponding to the number of channels in the two branches are generated through a multilayer perceptron network. Using the selection weights, the output features of the structural branch encoder and the enhanced local detail features are channel-weighted. The weighted features are then added together to obtain the fused features from the first to the third layer.

[0028] The channel-adaptive convolutional fusion method specifically includes the following processing steps: The output features of the last two layers of the structural branch encoder and the last two layers of the detail branch encoder at the same level are used as independent inputs and concatenated along the channel dimension at the same level to form joint features. The global average pooling vector and the global max pooling vector of the joint features are calculated separately and processed separately through a shared multilayer perceptron. The two processed vectors are added and passed through an activation function to generate a channel attention vector for reweighting the joint features. The channel attention vector is multiplied with the joint features channel by channel to obtain channel-enhanced features. The channel-enhanced features are then convolved to fuse spatial information and restore the number of channels, resulting in the fused features of the fourth and fifth layers.

[0029] Step Six: Input the fusion features of the fifth layer into the decoder for decoding, and optimize the features through the boundary guidance module. After upsampling, aggregate with the fusion features of the fourth layer. After decoding the aggregated features, further optimize them through the boundary guidance module. After upsampling, aggregate with the fusion features of the third layer again. Repeat the optimization and aggregation process, and process them sequentially with the fusion features of the second and first layers to complete image reconstruction and obtain the final segmentation prediction map.

[0030] Step 7: Output the final segmentation prediction map.

[0031] Based on the above technical solution, the following embodiments are provided.

[0032] This embodiment provides an efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder, and designs a set of collaborative feature enhancement, feature fusion, and boundary optimization mechanisms around it. Specific steps are as follows: Figure 2 As shown: Step 1: Acquire the ultrasound image to be segmented and preprocess it to a fixed size. In this embodiment, the input image is normalized to 256 pixels × 256 pixels with 3 channels. This input can be formally represented as... .

[0033] Step 2: Dual-path heterogeneous feature extraction. Extract the preprocessed image... The parallel inputs are fed to a dual-channel heterogeneous encoder. For the structure branch encoder G-DSHE and the detail branch encoder L-DSHE, the aim is to extract global structural features and local detail features. The encoding process of its first three layers can be represented as follows: in, , and The first, representing the structural branch and the detailed path respectively. Heterogeneous coding units of layers The structural branches employ 7×7 standard convolutions and 3×3 convolutions with a dilation rate of 4 (equivalent to a 9×9 receptive field), while the detail branches employ 3×3 standard convolutions and 3×3 convolutions with a dilation rate of 2 (equivalent to a 5×5 receptive field). Features extracted through standard and dilated convolutions are then concatenated. and Then, the feature map is output by combining channel and spatial attention mechanisms. The spatial dimensions decrease as the number of floors increases.

[0034] Step 3: Enhanced Feature Detail. This involves enhancing the output features of the first three layers of the detail branches. The input is fed into the detail-guided feature enhancement module DGBEM, which performs a sequential feature aggregation enhancement process from shallow to deep, as follows: in, This represents the core detail enhancement unit, which includes operations such as Sobel edge information extraction and multi-scale depthwise separable convolution; This represents a downsampling operation; This represents the element-wise addition of features. Ultimately, the enhanced detail features are obtained. .

[0035] Step 4: Deep Feature Encoding. The output features of the third layer of the dual-branch system are then encoded. and These are input into the encoders of their respective branches, which contain Token-MLP modules, for global context modeling. The output features of the fourth-layer encoder are: in, This indicates a deep encoder that independently processes deep features in dual-branch systems.

[0036] Step 5: Hierarchical Feature Fusion Strategy. For the heterogeneous information extracted from the dual-branch approach, a differentiated fusion strategy is used to aggregate the dual-branch features. This includes shallow fusion (…). The adaptive kernel selection fusion method ASKF is adopted. For the first... The fusion process of the layers is as follows: The ASKF operation first generates an adaptive channel attention weight vector. Then perform channel-weighted summation: in, This indicates multiplication by channel.

[0037] For deep integration ( ), using the channel adaptive convolutional fusion method CACF, for the , The layer fusion process is as follows: Specifically, the deep features of the dual encoders are first spliced ​​together: Channel statistical descriptors are obtained through global average pooling (GAP) and GMP, and channel attention maps are generated via a shared MLP network. Finally, weighted fusion is performed, and spatial information is fused through 3x3 convolution and then reduced in dimensionality. Step Six: Boundary-Guided Decoding and Reconstruction. The fused features are input into the decoder for feature reconstruction. The decoder integrates a boundary-guided module (BD).

[0038] For deep fusion features, firstly, the previous layer fusion features are... The input is decoded by a decoder containing the Token-MLP module, and the features are optimized by the boundary guidance module, upsampled, and aggregated with the fused features of the current layer. For shallow fusion features, firstly, the previous layer fusion features are... The input is decoded by a decoder containing convolutional modules, and the features are optimized by the boundary guidance module, upsampled, and then aggregated with the fused features of the current layer. in, This represents an upsampling operation. This indicates a feature addition operation. Indicates the boundary guidance module. This represents a 3×3 convolution. First, it uses the difference feature function... With energy function The salient responses at each feature location are calculated, and then fused with the input features through a gating mechanism: in, This represents element-wise multiplication. This is the sigmoid function.

[0039] Step 7: Reconstruct the final feature map from the decoder. Convolutional operations and nonlinear activations are performed to generate segmentation prediction maps. ,in This represents a 1×1 convolution, which maps the number of channels to the number of segmentation categories (2). This represents the activation function, and the output class probability map for each pixel is as follows: .

[0040] Model training employs the hybrid loss function described in claim 7 during the training process. in, For pixel-level cross-entropy loss, and Jointly represented as regional losses, The weight coefficients for each item are determined through cross-validation optimization to synergistically improve the segmentation performance of the model.

[0041] The method proposed in this invention has achieved significant results in ultrasound image segmentation tasks, such as... Figure 3 The layer-by-layer encoder heatmap shown in section (a) clearly demonstrates that the structural branches effectively focus on the overall outline of the target as the encoding process progresses, while the detail branches focus on details such as edges and textures as the encoding process progresses; for example... Figure 3 The visualization of the segmentation results shown in section (b) further confirms the effectiveness of the dual-path heterogeneous coding design. Quantitative evaluation on the public dataset DatasetB shows that the proposed method achieves good segmentation accuracy with an average Intersection over Union (IoU) of 79.39% and a Dice coefficient of 88.25%. Simultaneously, the model maintains its lightweight and efficient characteristics, processing 220 ultrasound images per second. These results validate that the proposed method, through collaborative structural and detailed features and the introduction of targeted enhancement and optimization mechanisms, can effectively address the challenges of low contrast and blurred boundaries in ultrasound images, achieving a balance between accuracy and efficiency.

[0042] As a second aspect, the present invention also provides a high-efficiency ultrasound image segmentation system based on a dual-channel heterogeneous encoder, comprising: Image acquisition unit, used to acquire ultrasound images to be segmented; A network training unit is used to construct a dual-path heterogeneous coding network and train it using a hybrid loss function to obtain a trained dual-path heterogeneous coding network. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing of the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive high-level features from the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features from the decoder to refine the segmentation boundaries. The image segmentation unit is used to input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map.

[0043] Therefore, compared with the prior art, the beneficial effects of the present invention are as follows: 1) This invention achieves deep and specialized extraction of heterogeneous features from ultrasound images through the design of a dual-channel heterogeneous encoder. The network can extract two types of key features in parallel and fully. This solves the fundamental problem of the difficulty in coordinating the capture of global structure and local details.

[0044] 2) This invention constructs a segmentation optimization mechanism for blurred edges by introducing a detail-guided feature enhancement module and joint loss function supervision, which significantly improves the model's accuracy in locating and segmenting target boundaries under the interference of vocal tract shadows and speckle noise, making the segmentation contour more in line with clinical standards.

[0045] 3) This invention proposes a hierarchical feature fusion strategy, which adopts a differentiated fusion method for shallow high-resolution detail information and deep high-level semantic information, realizing refined information interaction across paths and scales, effectively improving feature utilization efficiency and the completeness of segmentation results.

[0046] 4) This invention constructs a two-stage boundary optimization system of "encoder-decoder," forming a complete boundary-aware link that runs through the feature extraction and reconstruction process: the former enhances local edge responses in the feature extraction stage, while the latter uses contextual semantics to perform global boundary correction in the feature reconstruction stage. This design enables boundary information to form a bidirectional enhancement loop between the network's encoding and decoding layers, significantly improving the model's ability to model complex and irregular target boundaries and its segmentation consistency.

[0047] 5) This invention introduces a tokenized MLP module deep into the encoder, which realizes global context modeling of high-level features with extremely low computational overhead, effectively making up for the limited receptive field of traditional convolution. This design, combined with a shallow multi-scale convolutional heterogeneous encoder, forms an efficient feature extraction mode of "local perception-global integration", which significantly enhances the expressiveness and discriminativeness of deep semantic features.

[0048] 6) The present invention is constructed using lightweight modules, which completely avoids the high computational overhead caused by complex global attention mechanisms. It achieves an advanced balance between accuracy and speed on multiple public datasets, and its efficient inference performance ensures the feasibility of deploying the model in real-time clinical diagnosis and resource-constrained devices.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0050] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A highly efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder, characterized in that, include: Acquire the ultrasound image to be segmented; A dual-path heterogeneous coding network is constructed and trained using a hybrid loss function to obtain a trained dual-path heterogeneous coding network. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing on the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive features from each layer of the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features of the decoder to refine the segmentation boundaries. The ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map.

2. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 1, characterized in that, The ultrasound image to be segmented is input into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map. The specific processing steps include: Step 1: Obtain the ultrasound image to be segmented; Step 2: Input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network, and extract features in parallel to obtain global structural features and local detail features respectively; Step 3: Input the output features of the first three layers of the detail branch encoder into the detail-guided feature enhancement module in sequence for feature enhancement and aggregation to obtain enhanced local detail features; Step 4: Input the third-layer output features of the structural branch encoder and the detail branch encoder into their respective subsequent encoders containing the Token_MLP module for further processing, and obtain the independent last two layers of features, namely the fourth-layer output features and the fifth-layer output features. Step 5: First, perform shallow fusion of the first three layers of output features of the structural branch encoder and the enhanced local detail features using an adaptive kernel selection fusion method to obtain the fused features from the first to the third layer. Then, perform deep fusion of the last two layers of features of the structural branch encoder and the last two layers of features of the detail branch encoder using a channel adaptive convolution fusion method to obtain the fused features of the fourth and fifth layers. Step 6: Input the fusion features of the fifth layer into the decoder for decoding, and optimize the features through the boundary guidance module. After upsampling, aggregate with the fusion features of the fourth layer. After decoding the aggregated features, further optimize them through the boundary guidance module. After upsampling, aggregate with the fusion features of the third layer again. Repeat the optimization and aggregation process, and process with the fusion features of the second and first layers in turn to complete image reconstruction and obtain the final segmentation prediction map. Step 7: Output the final segmentation prediction map.

3. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 2, characterized in that, The specific processing steps of the detail-guided feature enhancement module include: Edge information extraction and multi-scale depth-separable convolution operation are performed on the first layer output features of the detail branch encoder to obtain the first enhanced feature; The first enhanced feature is downsampled and added to the second layer output feature of the detail branch encoder, and the same enhancement operation is performed on the addition result to obtain the second enhanced feature; The second enhanced feature is downsampled and added to the third layer output feature of the detail branch encoder, and the same enhancement operation is performed on the addition result to obtain the third enhanced feature; The first enhanced feature, the second enhanced feature, and the third enhanced feature are used as enhanced local detail features.

4. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 2, characterized in that, The adaptive kernel selection fusion method is adopted, and the specific processing steps include: The first three layers of the output features of the structural branch encoder at the same level and the enhanced local detail features are used as independent inputs. The two branch features are added element by element and then global average pooling is performed. Selection weights corresponding to the number of channels of the two branches are generated through a multilayer perceptron network. Using the selection weights, the output features of the structural branch encoder and the enhanced local detail features are respectively channel-weighted, and then the weighted features are added together to obtain the fused features from the first layer to the third layer.

5. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 2, characterized in that, The channel-adaptive convolutional fusion method specifically includes the following processing steps: The output features of the last two layers of the structural branch encoder and the output features of the last two layers of the detail branch encoder at the same level are used as independent inputs and concatenated along the channel dimension at the same level to form joint features. The global average pooling vector and the global max pooling vector of the joint features are calculated respectively and processed separately through a shared multilayer perceptron. The two processed vectors are added together and passed through an activation function to generate a channel attention vector for reweighting the joint features. The channel attention vector is multiplied channel by channel with the joint feature to obtain the channel enhancement feature. The channel enhancement feature is then convolved to fuse spatial information and restore the number of channels, resulting in the fused features of the fourth and fifth layers.

6. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 1, characterized in that, The structural branch encoder extracts features through parallel branches comprising a first standard convolutional layer, a dilated convolutional layer with a first dilation rate, and a standard convolutional layer with a first size, and fuses them through a channel and spatial attention mechanism to obtain global structural features; the detail branch encoder extracts features through parallel branches comprising a second standard convolutional layer, a dilated convolutional layer with a second dilation rate, and a standard convolutional layer with a second size, and fuses them through a channel and spatial attention mechanism to obtain local detail features; wherein, the first dilation rate is greater than the second dilation rate, and the first size is greater than the second size.

7. The efficient ultrasound image segmentation method based on a dual-channel heterogeneous encoder according to claim 1, characterized in that, The hybrid loss function is specifically expressed as follows: in, For pixel-level cross-entropy loss, and Jointly represented as regional losses, These are the weighting coefficients for each item.

8. A high-efficiency ultrasonic image segmentation system based on a dual-channel heterogeneous encoder, characterized in that, include: Image acquisition unit, used to acquire ultrasound images to be segmented; A network training unit is used to construct a dual-path heterogeneous coding network and train it using a hybrid loss function to obtain a trained dual-path heterogeneous coding network. The dual-path heterogeneous coding network includes a parallel structural branch encoder and a detail branch encoder, and a decoder connected to the structural branch encoder and the detail branch encoder. The detail branch encoder is also connected to a detail-guided feature enhancement module, and the decoder is also connected to a boundary guidance module. The structural branch encoder is used to extract global structural features; the detail branch encoder is used to extract local detail features; the detail-guided feature enhancement module is used to perform interactive serial processing of the first three layers of the detail branch encoder output from shallow to deep; the boundary guidance module is used to receive high-level features from the decoder and generate a boundary saliency map, and to perform gated fusion with the upsampled features from the decoder to refine the segmentation boundaries. The image segmentation unit is used to input the ultrasound image to be segmented into the trained dual-path heterogeneous coding network for feature processing to obtain the final segmentation prediction map.

Images