Sublingual blood vessel lightweight segmentation method based on double-branch lateralization interaction

Abstract

The application discloses a sublingual blood vessel lightweight segmentation method based on double-branch lateralization interaction and belongs to the technical field of medical image segmentation. The collected sublingual blood vessel original image is preprocessed to obtain a to-be-segmented image; a lightweight segmentation network with an asymmetric double-branch structure is constructed, including a left branch, a right branch and a plurality of cross-branch feature fusion units, global context features and local detail features are extracted, and multi-level information interaction is realized according to the fusion units at the encoder, the bottleneck layer and the decoder; a multi-scale cyclic residual structure module is embedded in the network, and a plurality of depth separable convolutions and cyclic residual connections are combined to complete feature extraction and transmission; the application uses a composite loss function to complete network training, and realizes model iteration and application according to a special data set. The application effectively extracts sublingual blood vessel features of different scales, can balance segmentation accuracy and running efficiency, and is suitable for embedded and portable medical terminals to carry out automatic sublingual blood vessel segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a lightweight segmentation method for the hypolingual vessels based on bibranch lateralization interaction. Background Technology

[0002] The morphology and blood flow status of sublingual vessels are important references for clinical diagnosis of acute conditions such as shock and sepsis. Automatic segmentation of sublingual vessels using machine vision and deep learning technologies can promote the implementation of intelligent assisted diagnosis for related diseases. Sublingual vessels are characterized by their slender diameter, dense branching, and blurred edges. Furthermore, noise interference easily introduced during medical imaging significantly increases the technical difficulty of automatic segmentation. Currently, mainstream medical image segmentation models are divided into two main categories: heavy-duty networks and lightweight networks. Heavy-duty segmentation networks, such as U-Net and R2U-Net, rely on multi-layer convolutional structures and complex connection methods to ensure segmentation accuracy and can completely identify the main vessels and small branches. However, these models have a large number of parameters, high floating-point computation requirements, and consume a large amount of memory and computing power during inference, making them difficult to deploy in embedded or edge-based clinical devices such as bedside detection equipment and portable diagnostic terminals. While lightweight networks reduce model parameters and computational overhead through convolutional structure optimization, making them suitable for resource-constrained devices, they often suffer from over-compression of feature extraction capabilities. This leads to issues such as missed detections, segmentation fragmentation, and background misjudgment when processing small sublingual vessels or those with blurred vessel edges. The trade-off between accuracy and efficiency has long existed, making it impossible to simultaneously meet the dual requirements of real-time clinical detection and accurate segmentation. Therefore, developing deep learning networks that balance lightweight design with high segmentation accuracy for sublingual vessel segmentation has become a key research direction in the field of medical image processing.

[0003] In recent years, numerous domestic and international academic teams have conducted in-depth research on lightweight medical image segmentation networks, producing a series of classic theoretical and technological achievements. In 2017, the Google team published the paper "MobileNets: EfficientConvolutional Neural Networks for Mobile Vision Applications" (arXiv:1704.04861), which first proposed a depthwise separable convolutional structure. This structure decomposes the standard convolution into two independent operational units: depthwise convolution and pointwise convolution, significantly reducing the number of parameters and computational cost of convolutional layers. This structure has become the foundational component of most subsequent lightweight networks. In the same year, the Megvii Technology team published the paper "ShuffleNet: An Extremely EfficientConvolutional Neural Network for Mobile Devices" (arXiv:1707.01083), innovatively introducing pointwise group convolution and channel shuffling operations to solve the problem of feature flow obstruction caused by group convolution, further improving the inference speed of lightweight models, and making them widely used in mobile vision tasks. In addition, some studies have incorporated brain-like structures such as small-world networks and spiking neural networks into segmentation models. Networks such as LSWNet and Bio-Net have borrowed the connectivity characteristics of the biological brain to optimize information transmission logic, attempting to retain feature extraction capabilities while reducing computational consumption. However, the above-mentioned academic research results still have obvious shortcomings when applied to the sublingual vessel segmentation scenario: MobileNet and ShuffleNet simply focus on computational compression without making specific optimizations for the multi-scale morphology of blood vessels and blurred edges, resulting in a high rate of missed detection of small blood vessels; although brain-like lightweight networks simulate the topological characteristics of biological networks, they only use a single-branch structure for feature extraction, which cannot simultaneously take into account global contextual information and local detail information. Faced with the complex branching structure and low-contrast imaging area of ​​the sublingual vessels, the segmentation integrity and accuracy are difficult to meet the standards for clinical use, and the overall adaptability is insufficient.

[0004] In China, numerous invention patents have been filed focusing on medical vessel segmentation and lightweight image segmentation. Among them, Chinese invention patent CN115760872A, "A Lightweight Retinal Vessel Segmentation Method Based on Attention Mechanism," combines an attention mechanism with a lightweight network. By strengthening the weights of vessel features and suppressing background interference, it achieves rapid segmentation of retinal vessels. This solution employs a lightweight encoding / decoding architecture, effectively controlling the number of model parameters and adapting to mobile medical device operating scenarios. Chinese invention patent CN116433905B, "A Lightweight Medical Image Segmentation Method Based on Dynamic Convolution," designs a dual-path structure with semantic and spatial branches to extract semantic and spatial features of the image respectively. Dynamic convolution enhances the model's adaptability to different morphological targets, reducing model complexity while optimizing general medical image segmentation results. While these patented technologies have made beneficial explorations in lightweight design, multi-branch feature extraction, and anti-interference, their technical solutions still have limitations. On the one hand, both patents are designed for retinal and general medical image design, without taking into account the unique morphological characteristics of the small diameter and dense branches of the sublingual vessels. Direct application of these patents would result in a decrease in segmentation performance. On the other hand, their dual-branch structure only adopts a parallel feature extraction mode and does not introduce functional lateralization design. There is a lack of multi-level dynamic interaction and collaboration mechanism between the two branches, resulting in high functional overlap between the branches. It is impossible to achieve differentiated extraction of global features and local details. Faced with the complex topological structure of the sublingual vessels, it is still difficult to avoid the defects of missing small vessels and discontinuous segmentation of vessel edges.

[0005] Existing academic research and publicly available patented technologies reveal that there is currently no lightweight segmentation scheme specifically designed for sublingual vessel segmentation that integrates brain-like lateralization and collaborative mechanisms. Existing technologies either prioritize accuracy at the expense of operational efficiency, making them unsuitable for deployment in edge clinical devices; or they prioritize lightweight design at the expense of feature extraction capabilities, making it difficult to accurately identify tiny sublingual vessels. Some bi-branch segmentation schemes also fail to achieve differentiation in branch functions and multi-level interactive collaboration, resulting in a lack of full integration of global information and local details. Consequently, they remain unable to overcome the industry pain point of the mutual constraints between accuracy, parameter quantity, and inference speed. Summary of the Invention

[0006] To address the aforementioned technical problems, this application discloses a lightweight segmentation method for the hypolingual vessels based on bibranch lateralization interaction, comprising the following steps:

[0007] 101. Obtain the original image of the sublingual blood vessels, and perform image standardization preprocessing on the original image of the sublingual blood vessels to obtain the image to be segmented;

[0008] 102. Construct an asymmetric dual-branch lateralized lightweight segmentation network, which includes a parallel left branch network, a right branch network, and multiple sets of cross-branch feature fusion units. The left branch network and the right branch network each have an independently set encoder, bottleneck layer, and decoder.

[0009] 103. Input the image to be segmented into an asymmetric dual-branch lateralized lightweight segmentation network. Feature extraction is completed through the left branch network and the right branch network respectively. Relying on the cross-branch feature fusion unit, multi-level feature interaction between the two branch networks is realized in the encoder, bottleneck layer and decoder.

[0010] 104. The features that have completed interactive fusion are sequentially decoded and classified to finally output the sublingual blood vessel segmentation result image.

[0011] Preferably, the left branch network and the right branch network are functionally differentiated asymmetric branch structures, and both branches adopt a lightweight convolutional architecture and set a small-world topology connection form;

[0012] The left branch network internally constructs a global feature extraction link to extract global contextual features of the image to be segmented;

[0013] The right branch network internally constructs a local feature extraction link to extract local detailed features of the sublingual blood vessels in the image to be segmented.

[0014] Preferably, the encoder, bottleneck layer, and decoder of the left and right branch networks are each embedded with several multi-scale cyclic residual structure modules, which are the core lightweight feature extraction units of the network.

[0015] The multi-scale cyclic residual structure module integrates three types of depthwise separable convolutions, specifically including 3×3 depthwise separable convolution, 1×1 depthwise separable convolution, and dilated depthwise separable convolution.

[0016] Preferably, the multi-scale cyclic residual structure module performs channel-dimensional concatenation and fusion operations on the feature maps output by the three types of depthwise separable convolutions within it. The calculation formula for the fused features is as follows:

[0017]

[0018] in, To input the feature map of the multi-scale cyclic residual structure module, The output features are from a 3×3 depthwise separable convolution. The output features are those of a 1×1 depthwise separable convolution. The output features of the depth-separable convolution are the dilated features. This represents the concatenation operation of the channel dimension.

[0019] Preferably, in addition to the convolution operation unit, the multi-scale cyclic residual structure module is also equipped with a cyclic residual connection structure to realize cross-layer transmission of features within the module;

[0020] The cyclic residual connection structure establishes cross-layer direct connections between different convolutional layers and feature operation layers within the module, directly transmitting the feature maps generated in the shallow layers of the module to the deep operation nodes of the module.

[0021] Preferably, the right branch network adopts an encoder-decoder architecture, with the encoder and decoder respectively paired with multi-scale cyclic residual structure modules and skip connection paths at corresponding levels;

[0022] The encoder of the right branch network is composed of multiple downsampling layers connected in series, and each downsampling layer is connected to a multi-scale cyclic residual structure module at the same level.

[0023] The decoder of the right branch network consists of multiple fine-grained upsampling layers. Each upsampling layer has an independent skip connection path, which connects the corresponding layers of the right branch network encoder and decoder.

[0024] Preferably, the left branch network also adopts an encoder-decoder architecture, with an overall hierarchical structure corresponding to the right branch network, and independently configured with downsampling layers, upsampling layers, and skip connection structures;

[0025] The encoder of the left branch network consists of a global modeling path composed of continuous multi-level downsampling layers, with downsampling layers and multi-scale cyclic residual structure modules arranged alternately.

[0026] The decoder of the left branch network includes a multi-level upsampling module, and the skip connection structure in the decoder connects the shallow feature map of the encoder of the left branch network to the corresponding operation layer of the decoder.

[0027] Preferably, the cross-branch feature fusion unit realizes information interaction between branches by dividing it into three categories according to the network hierarchy, which correspond to different functional areas of the network to complete feature interoperability.

[0028] The cross-branch feature fusion unit includes an encoding area fusion unit, a bottleneck area fusion unit, and a decoding area fusion unit;

[0029] The encoding area fusion unit connects the same level nodes of the left branch network encoder and the right branch network encoder;

[0030] The bottleneck fusion unit connects the bottleneck layer of the left branch network and the bottleneck layer of the right branch network.

[0031] The decoding area fusion unit connects the same level nodes of the left branch network decoder and the right branch network decoder.

[0032] Preferably, when training the dual-branch lateralized lightweight segmentation network, a composite loss function is used as a constraint condition for parameter iteration. This composite loss function is composed of a combination of the Dice loss function and the cross-entropy loss function, specifically including:

[0033] The Dice loss function formula is: ;

[0034] The formula for the cross-entropy loss function is: ;

[0035] The formula for the composite loss function is: ;

[0036] in, These are the network pixel prediction values. The actual pixel label value. These are the weight coefficients; during training, the gradient is backpropagated based on the composite loss function to correct the weight parameters of each convolutional layer and connection structure within the network.

[0037] Preferably, the dataset completes network training and segmentation operations, and the dataset is divided into different subsets according to the use scenario and used separately;

[0038] Construct a dedicated dataset for sublingual blood vessels containing multiple images and corresponding labeled images;

[0039] The dataset is divided into a training subset and a test subset;

[0040] Input the training subset into the network to perform iterative parameter training, and input the test subset into the trained network to perform sublingual vessel segmentation.

[0041] Compared with the prior art, the technical solution of this application has the following technical effects:

[0042] This invention employs an asymmetric dual-branch network architecture to segment sublingual blood vessel images. Based on the differentiated functional design of the left and right branches, it can extract global contextual features and local detail features respectively. Combined with a multi-level cross-branch feature fusion structure, it can achieve information exchange between branches throughout the entire process of network encoding, bottleneck, and decoding, allowing features of different dimensions to be fully combined, completely restoring the overall distribution and fine structure of sublingual blood vessels, and ensuring the comprehensive utilization of features during image segmentation.

[0043] This invention embeds multi-scale cyclic residual structure modules at each layer of the network, and completes feature acquisition at different scales based on separable convolutions of various depths. At the same time, it builds cross-layer feature transmission pathways by using cyclic residual connections within the modules. This structure can stably transmit shallow and deep features, optimize the information flow state within the network, adapt to the diverse morphology and complex edges of the sublingual blood vessels, and improve the network's ability to express features of various vascular structures.

[0044] This invention adopts a lightweight convolutional structure design and combines the inherent characteristics of depthwise separable convolution to simplify network computing units, effectively reducing the overall network size and computing overhead. The overall network topology is simple and efficient, and it occupies less hardware resources during operation. It can be adapted to resource-constrained operating environments such as embedded devices and portable terminals, and meets the needs of real-time sublingual vessel segmentation processing on site.

[0045] This invention employs a composite loss function to optimize network training. By combining the characteristics of different loss functions to constrain the network parameter update process, the model can gradually optimize pixel classification logic during the training phase. Based on standardized dataset partitioning and iterative training methods, the learning ability of the network can be steadily improved, enabling the segmentation results output by the model to have good stability and consistency, and adapting to long-term, routine medical image segmentation applications.

[0046] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0047] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0049] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0050] Figure 1This is a schematic diagram of the overall process framework of a lightweight sublingual vessel segmentation method based on bibranch lateralization interaction;

[0051] Figure 2 A complete hierarchical architecture diagram of a lightweight split network with asymmetric dual-branch structure and lateralized topological attributes;

[0052] Figure 3 A block diagram of the internal convolution and residual connection structure of the multi-scale cyclic residual structure fusion module;

[0053] Figure 4 Comparison of the segmentation results of various mainstream medical segmentation networks with the proposed DualSI-Net sublingual vessel segmentation scheme;

[0054] Figure 5 This is a scatter plot comparing the pixel prediction numerical distribution of our proposed DualSI-Net solution with that of the contrasting model.

[0055] Figure 6 A heatmap showing the correlation of segmentation prediction results for DualSI-Net and various comparative network models. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0057] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0058] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0059] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0060] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0061] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0062] Example 1 mainly describes a lightweight segmentation method for the hypolingual vessels based on bibranch lateralization interaction, such as... Figure 1 As shown, it includes the following steps:

[0063] 101. Obtain the original image of the sublingual blood vessels, and perform image standardization preprocessing on the original image of the sublingual blood vessels to obtain the image to be segmented;

[0064] 102. Construct an asymmetric dual-branch lateralized lightweight segmentation network, which includes a parallel left branch network, a right branch network, and multiple sets of cross-branch feature fusion units. The left branch network and the right branch network each have an independently set encoder, bottleneck layer, and decoder.

[0065] 103. Input the image to be segmented into an asymmetric dual-branch lateralized lightweight segmentation network. Feature extraction is completed through the left branch network and the right branch network respectively. Relying on the cross-branch feature fusion unit, multi-level feature interaction between the two branch networks is realized in the encoder, bottleneck layer and decoder.

[0066] 104. The features that have completed interactive fusion are sequentially decoded and classified to finally output the sublingual blood vessel segmentation result image.

[0067] This application is designed based on the lateralization and collaborative working mechanism of the brain's nervous system and the topological characteristics of small-world networks. Specifically addressing the segmentation challenges of sublingual blood vessels—characterized by their small diameter, complex branches, varied morphology, blurred imaging edges, and susceptibility to noise interference—an asymmetric, dual-branch, lateralized, lightweight segmentation network is constructed. The overall architecture is shown in Figure 2. The entire network uses parallel dual branches as its core, with the two branches arranged side-by-side in physical structure and exhibiting clear functional differentiation, fully simulating the division of labor in visual processing between the left and right hemispheres of the human brain. Both branches are fully divided into three functional sections: encoder, bottleneck layer, and decoder. Each section connects multiple layers of basic computational units according to the network's operational logic, ensuring the layer-by-layer extraction, transformation, and reconstruction of features.

[0068] To balance lightweight design and efficient information transmission, this network adopts a brain-like small-world topology. This structure combines high clustering coefficients and short average path lengths, enabling high-density interaction of features within local regions while ensuring rapid information transmission globally. In addition to the two main branches, the network is equipped with multiple sets of cross-branch feature fusion units. These units are categorized and deployed according to network operation segments, corresponding to encoding, bottleneck, and decoding regions, and are integrated throughout the entire network computation process. After preprocessing, the sublingual vessel image to be segmented is simultaneously input into the left and right branch networks. The two branches independently perform feature extraction, while the cross-branch feature fusion units at each level communicate features in real time. This allows global features and local detail features to be shared and collaboratively processed across the entire network. Finally, the fused features are decoded and classified to output an accurate sublingual vessel segmentation image.

[0069] Furthermore, the dual-branch network includes a left-branch network, a right-branch network, and a cross-branch feature fusion unit, specifically:

[0070] The left branch is responsible for global feature extraction and context modeling, simulating the human brain's global perception and background reasoning working mode. This branch adopts a standard encoder-decoder architecture. The encoder consists of multiple downsampling layers connected in series to form a dedicated global modeling path. The downsampling layers and multi-scale cyclic residual structure modules are arranged alternately in a fixed order. Continuous downsampling operations gradually compress the image spatial size, continuously expanding the receptive field, thereby extracting overall background information, the overall distribution structure of blood vessels, and long-distance dependencies between pixels. It possesses complete information modeling capabilities for areas with blurred boundaries and low contrast in sublingual vascular imaging.

[0071] The bottleneck layer of the left branch receives the deep global features output by the encoder, performs in-depth feature processing, and then transmits them to the decoder. The decoder consists of multi-level upsampling modules, responsible for progressively restoring the original spatial resolution of the image. Simultaneously, the decoder internally configures multiple skip connection structures. Each skip connection directly connects the shallow edge and texture features corresponding to the left branch encoder to the same-level computation node in the decoder. This design preserves shallow basic features during image size restoration, optimizes the integrity of the overall blood vessel contour, and improves the stability of the network's overall prediction results. Throughout the process, the left branch provides global contextual information to the other branch as guidance, achieving functional complementarity between the two branches.

[0072] The right branch focuses on extracting local details and microvascular structures, simulating the human brain's ability for fine visual analysis. It is specifically designed for recognizing fine filamentous microvessels and complex intersecting vascular morphologies in the sublingual region. This branch also employs an encoder-decoder architecture, with the encoder consisting of multiple downsampling layers connected in series. Each downsampling layer is paired with a multi-scale cyclic residual structure module at the same level, continuously mining fine-grained textures and microvascular features while compressing feature sizes.

[0073] The right branch decoder is built using multi-level fine-grained upsampling layers. Unlike conventional upsampling methods, fine-grained upsampling can preserve the spatial resolution of the image to the greatest extent and avoid the loss of microvascular structures during reconstruction. Each upsampling layer of the decoder is independently configured with a skip connection path, with the two ends of the path connecting the corresponding layers of the right branch encoder and decoder, respectively. This ensures that the microvascular features and edge features captured in the encoding stage are completely transmitted to the decoding stage, effectively maintaining the structural continuity and spatial distribution integrity of microvascular structures. The right branch finally outputs a vascular feature map with high spatial accuracy and high structural consistency, providing sufficient local detail support for the overall segmentation result.

[0074] The cross-branch feature fusion unit is the core component for realizing the lateralized interaction between the two branches. Based on its deployment location, it is divided into three main types: encoding area fusion unit, bottleneck area fusion unit, and decoding area fusion unit. The encoding area fusion unit connects to nodes of the same level in both the left and right branch encoders, enabling bidirectional interaction of shallow texture and basic morphological features. The bottleneck area fusion unit directly connects to the bottleneck layers of the two branches, achieving the sharing of deep semantic features. The decoding area fusion unit connects to nodes of the same level in both the left and right branch decoders, enabling the fusion of contour and detail features during the decoding stage. Through full-level feature interaction, semantic information, edge information, global information, and local information are cross-fused, ensuring that the two branches always maintain a collaborative operational state.

[0075] Furthermore, numerous multi-scale recurrent residual structure modules are embedded within the encoders, bottleneck layers, and decoders of both the left and right branch networks. These modules are the core units for achieving network lightweighting, multi-scale feature extraction, and stable training of deep networks. The specific structure is as follows: Figure 3 As shown, the module integrates a multi-scale feature fusion mechanism and a recurrent residual structure. Internally, it integrates three types of depthwise separable convolutional units with different functions: 3×3 depthwise separable convolutional units, 1×1 depthwise separable convolutional units, and dilated depthwise separable convolutional units. All three types of convolutional units are lightweight convolutional structures, capable of extracting features from different dimensions while significantly reducing computational and parameter requirements.

[0076] Among them, the 3×3 depthwise separable convolution first performs spatial convolution on a single channel, and then completes channel fusion through 1×1 pointwise convolution, mainly used to capture local details and texture information of blood vessels; the 1×1 depthwise separable convolution further simplifies the operation logic, continuously compresses the network computational overhead, and optimizes the overall lightweight effect; the dilated depthwise separable convolution expands the receptive field by relying on the dilation parameter, and can complete long-distance feature association modeling without increasing the convolution kernel size, integrating global context information. The three types of convolutional units simultaneously receive the input feature maps from the module. Each operation sequentially performs convolution, batch normalization (BN) processing, and activation function operation. The basic convolution branch is paired with the LeakyReLU activation function, while the fusion branch is paired with the ReLU activation function.

[0077] After the three types of convolutional units complete their operations, the output feature maps are concatenated and fused along the channel dimension. The fusion operation follows the calculation formula below:

[0078]

[0079] In the formula, To input the feature map of the multi-scale cyclic residual structure module, The output features are from a 3×3 depthwise separable convolution. The output features are 1×1 depthwise separable convolutions. Separable convolution output features for hole depth This represents the concatenation operation of the channel dimension.

[0080] In addition to the multi-scale convolutional fusion branch, the module also incorporates a cyclic residual connection structure. This structure establishes multiple cross-layer direct connections between various convolutional layers and feature operation layers within the module. These cross-layer paths directly transmit feature maps generated in the shallow layers of the module to deeper operation nodes, enabling the cyclic reuse of low-level and high-level features. During the operation of deep neural networks, feature decay and gradient vanishing issues easily arise as the network depth increases. Cyclic residual connections can construct stable gradient propagation paths, effectively alleviating these problems and reducing the training difficulty of deep networks. Simultaneously, this structure enhances the network's feature memory capability, allowing the module to accurately capture the edges of fine blood vessels and complex vascular intersections, comprehensively improving the network's robustness to complex sublingual vascular images. Multiple multi-scale cyclic residual structure modules can be cascaded and stacked, maintaining the cross-layer transmission logic of cyclic residuals even after stacking, further enhancing the network's feature extraction capabilities.

[0081] Furthermore, this application employs a composite loss function as a constraint during the network training phase to iteratively update the weight parameters of all convolutional layers, residual connection structures, and fusion units within the network. The composite loss function is a combination of the Dice loss function and the cross-entropy loss function. These two loss functions play different roles and work together to optimize the network training effect. The Dice loss function focuses on optimizing region overlap, adapting to the training requirements of target region integrity in image segmentation tasks; the cross-entropy loss function focuses on pixel-by-pixel classification optimization, stabilizing gradient output during training.

[0082] The complete calculation formulas for the three types of functions are as follows:

[0083] The Dice loss function formula is:

[0084]

[0085] The formula for the cross-entropy loss function is:

[0086]

[0087] The formula for the composite loss function is: ;

[0088] in: Represents the network's response to the image. The predicted output value of each pixel. This represents the actual human-labeled value corresponding to that pixel. These are preset weight coefficients used to balance the proportions of the two types of loss functions in the overall loss. During network training, image segmentation prediction results are first obtained through forward propagation, and the composite loss function value is calculated based on the labeled data. Then, the gradient is calculated layer by layer according to the backpropagation algorithm, and all weight parameters in the network are adjusted according to the gradient direction. The iterative process of forward propagation, loss calculation, backpropagation, and parameter update is repeatedly executed until the network parameters converge, completing the model training. This entire training logic can continuously optimize the network's pixel classification and region segmentation capabilities.

[0089] Furthermore, this application utilizes a specially constructed dataset for sublingual vessels. This dataset consists of a massive number of raw sublingual vessel images collected in real-world scenarios, along with corresponding pixel-level annotated images. The annotated images accurately distinguish between the vessel region and the background region, providing a standard data source for network training and testing. Before formal use, the complete dataset is divided into two independent, non-overlapping subsets: a training subset and a testing subset. The number of samples in each subset is allocated according to application requirements, serving the model training and performance testing phases respectively.

[0090] Before the images enter the network for computation, a unified image standardization preprocessing operation is performed. The preprocessing includes a series of conventional processing steps such as uniform image size scaling, pixel value normalization, and noise filtering to eliminate size differences, brightness differences, and imaging noise between the original images, and finally generate images to be segmented with uniform format and stable state.

[0091] During the model training phase, all preprocessed images to be segmented from the training subset are batch-input into the untrained dual-branch lateralized lightweight segmentation network. Multiple rounds of iterative training are performed using the composite loss function described earlier, continuously optimizing the network parameters until the model performance reaches a stable state. In the actual segmentation application phase, raw images of the sublingual vessels are acquired on-site. After standardization preprocessing to obtain the images to be segmented, they are input into the trained network. The images sequentially undergo dual-branch feature extraction, full-level cross-branch feature fusion, multiple rounds of decoding operations, and pixel classification processing. The feature maps output from the two branches are merged, ultimately outputting a complete segmented image of the sublingual vessels. The entire process, from image acquisition, preprocessing, network computation to result output, forms a closed loop, enabling fully automated segmentation of the sublingual vessels independently.

[0092] This embodiment uses an asymmetric bibranch lateralization network combined with a multi-scale recurrent residual module to segment the sublingual vessels. The bibranch differential extraction extracts global and local vascular features, multi-level fusion enables information exchange, lightweight convolution reduces computational overhead, and a composite loss function stabilizes the training process, fully preserving the microvascular structure and adapting to portable medical devices to complete accurate segmentation calculations.

[0093] Example 2, based on Example 1, aims to evaluate the brain-like topological characteristics of DualSI-Net and its subnetworks. The network topology attributes of this model are compared with several classic segmentation models (U-Net, U-Net++, and SegNet) and LSWNet, which is inspired by small-world networks. Table 1 compares two graph-theory-based structural metrics: average clustering coefficient and average path length. The model structure is defined as an undirected graph G=(V,E), where V represents the set of neuron nodes, E represents the set of connecting edges, and the average clustering coefficient is defined as:

[0094]

[0095] in, represents the clustering coefficient of node v, used to measure the connectivity between its neighbors. The model achieves a clustering coefficient of 0.68, significantly higher than other models, indicating a tighter local connectivity structure; the average path length L represents the average shortest path length between any pair of nodes, defined as:

[0096]

[0097] Where, d It is a node and nodes The shortest path length between them. In the model, the average path length is 2.62, which remains at a low level, indicating that the model has a strong ability to transmit global information.

[0098] Table 1. Evaluation results of brain-inspired topological properties

[0099] 0.680 0.750 0.703 0.620 0.190 0.440 0.000 2.62 1.98 2.17 2.36 2.11 2.10 2.330

[0100] Based on the topological index quantification results in Table 1, the model in this application shows a balanced performance in both the average clustering coefficient (0.680) and average path length (2.62), exhibiting both high local connection density and short global information paths, displaying small-world characteristics similar to biological brain networks. This result indicates that the network not only enhances the extraction of subtle features through dense local connections but also maintains efficiency in cross-layer information transmission, thus supporting the synergy between local perception and global reasoning. Further observation of its internal sub-networks reveals structural differences: the left sub-network has the highest clustering coefficient (0.750) and the shortest path (1.98), indicating that its structure is highly focused on dense information interaction in local regions; the right sub-network has a similar clustering coefficient (0.703) but a slightly longer path (2.17), suggesting a structural basis for integrating broader contextual information. This structural division of labor is consistent with the design intent of simulating brain functional lateralization. Compared to other models, traditional U-Net and its variants, while having shorter paths, exhibit significantly lower clustering coefficients, reflecting their shortcomings in constructing local loops within their chain-like backbones. Furthermore, LSWNet, specifically optimized for small-world structures, also failed to surpass this model in topological performance. These comparisons demonstrate that by drawing upon the functional organization principles of brain networks for structural design, a more fundamental and efficient balance between local interactions and global propagation can be achieved in the network.

[0101] DualSI-Net employs a lightweight design, resulting in excellent performance in terms of model complexity and computational efficiency. Table 1 lists the key lightweight metrics of this model: In terms of parameters, DualSI-Net has a total of 6,271,809 parameters, significantly lower than similar networks such as U-Net (approximately 32 million parameters). This makes DualSI-Net more suitable for deployment in resource-constrained environments. In terms of computational cost (FLOPS), DualSI-Net requires approximately 38.6 billion floating-point operations per image during inference, significantly lower than more complex models, demonstrating its fast inference speed and high efficiency. Inference time (ms / image) tests, based on cuDNN version 8100, show that DualSI-Net's average inference time is 42.91 milliseconds per image, meeting the requirements of real-time or near real-time segmentation tasks. In terms of memory usage (MB), the model's memory footprint is approximately 23.93 MB, making it suitable for deployment on devices with limited memory resources. Therefore, DualSI-Net has significant advantages in edge computing and embedded systems.

[0102] Furthermore, to comprehensively evaluate the stability and generalization ability of DualSI-Net in medical image segmentation tasks, this application adopted a five-fold cross-validation experimental strategy for performance testing. As shown in Table 2, the entire dataset was divided into five non-overlapping subsets. Each time, one subset was selected as the test set, and the remaining four subsets were used as the training set. This process was repeated five times, ensuring that each subset could be used as a test set for model validation. Each fold experiment involved complete training and testing of the model under the same training settings, ensuring the fairness and comparability of the experimental results. This cross-validation method not only improves data utilization efficiency but also effectively reduces random errors introduced by different sample partitions, thus more realistically reflecting the model's performance under different data distributions. The performance metrics obtained through multiple rounds of experiments have stronger statistical robustness and persuasiveness, providing a reliable basis for model stability analysis and practical application promotion.

[0103] Table 2 lists the performance metrics and statistical results of DualSI-Net at each fold under the five-fold cross-validation setting. Overall, the model achieves an average Dice coefficient of 0.791, mIoU of 0.658, and AUC of 0.859 in the five-fold experiment, indicating that the proposed DualSI-Net can achieve relatively stable and reliable segmentation performance in the sublingual vessel segmentation task and has good pixel-level discrimination ability.

[0104] Looking at the fluctuations in results across different folds, the standard deviations of all metrics remained at low levels: Dice had a standard deviation of ±0.005, Precision ±0.011, and mIoU ±0.007. These small fluctuations indicate that the model has a low dependence on data splitting and can maintain relatively stable performance under different training-test splits, demonstrating good model stability and a certain degree of generalization ability.

[0105] Despite significant class imbalance (72.48% background pixels) in this dataset, the model still achieved an average precision of 0.821, demonstrating its effective suppression of false positives during prediction. Simultaneously, the Dice metric consistently remained above 0.79, indicating good integrity in identifying small vascular structures and a significant reduction in vascular rupture or missed detections. This, to some extent, validates the effectiveness of the proposed bibranch structure in mitigating class imbalance and enhancing vascular structure expression.

[0106] Table 2. Segmentation results of five-fold cross-validation

[0107] 1 0.785 0.894 0.805 0.65 0.86 2 0.798 0.896 0.819 0.667 0.866 3 0.795 0.894 0.837 0.663 0.857 4 0.793 0.890 0.822 0.66 0.86 5 0.786 0.885 0.822 0.652 0.853 average value 0.791 0.892 0.821 0.658 0.859 Standard deviation ±0.005 ±0.004 ±0.011 ±0.007 ±0.004

[0108] To comprehensively evaluate the performance of the proposed model, this application selected several representative deep learning segmentation models for comparative experiments, including classic convolutional neural network models, lightweight networks, and the improved U-Net structure proposed in recent years. Evaluation metrics included Dice coefficient, accuracy (ACC), precision, mean intersection-over-union ratio (mIoU), and AUC. The number of parameters for each model was also calculated to measure model complexity. Experimental results are shown in Table 3.

[0109] Table 3 Comparative Experimental Results

[0110] LSWNet 0.786 0.889 0.816 0.651 0.857 489,788 NestedU-net 0.797 0.895 0.831 0.666 0.861 36,654,548 Unet 0.635 0.846 0.67 0.476 0.743 31,919,380 ENet 0.637 0.842 0.855 0.475 0.746 373,148 MobileNet 0.634 0.823 0.696 0.468 0.752 3,469,448 DFANet 0.326 0.731 0.561 0.201 0.588 1,809,588 DeepLabV3 0.727 0.869 0.809 0.576 0.810 41,068,676 ICNet 0.692 0.832 0.712 0.534 0.793 6,742,932 Segnet 0.743 0.871 0.801 0.595 0.822 11,744,980 BIOnet 0.566 0.697 0.540 0.405 0.697 14,994,612 Attentionunet 0.796 0.894 0.823 0.665 0.862 31,921,528 Swinunet 0.78 0.888 0.826 0.644 0.845 8,992,764 r2unet 0.803 0.899 0.837 0.675 0.864 68,094,612 This application method 0.804 0.898 0.830 0.676 0.888 6,271809

[0111] In this experiment, various deep learning methods, including classic and novel models, were selected to evaluate the performance of the object segmentation task. A comprehensive comparison was conducted using metrics such as Dice coefficient, accuracy (ACC), precision (Pre), mean intersection-over-union (mIoU), and AUC. Furthermore, considering the complexity of different models,

[0112] The number of parameters is an important reference for measuring model size. As shown in the table, DualSI-Net significantly reduces the number of parameters and computational complexity while maintaining high segmentation accuracy. Experimental results demonstrate that our model exhibits superior performance across multiple segmentation metrics, showing clear competitiveness compared to other mainstream models. Furthermore, compared to the lightweight model ENet, our model achieves a significant improvement in segmentation accuracy, fully demonstrating the effectiveness of its optimization. In addition, our model achieves ACC and AUC of 0.898 and 0.888 respectively, demonstrating excellent robustness and generalization ability. Regarding the Dice coefficient, our model reaches 0.804, and the mIoU metric reaches 0.676, performing close to the high-complexity r2unet, but with only 6.27M parameters, significantly lower than r2unet (68.09M) and NestedU-net (36.65M), fully showcasing its lightweight characteristics. In contrast, while some models (such as r2unet and NestedU-net) have slight advantages in certain metrics, their large number of parameters limits their practical applications. Our model, through optimization of its structure, effectively reduces computational complexity while achieving high segmentation accuracy, successfully striking a good balance between lightweight design and high efficiency. This advantage makes it more practical in resource-constrained real-world scenarios, particularly suitable for medical image segmentation tasks with high requirements for computational efficiency and storage space, demonstrating significant application potential.

[0113] Furthermore, to more intuitively compare the performance of different models in the sublingual vessel segmentation task, this application selects several representative test samples and visualizes the prediction results of each model with the corresponding ground truth annotations, such as... Figure 4 As shown in the figure, the first row is the original image of the sublingual blood vessels, and the remaining rows are the segmentation results of different models.

[0114] From an overall segmentation perspective, most models can identify major vascular structures relatively well, but significant differences remain in small vascular branches, vascular boundaries, and complex intersection regions. For example, classic models such as U-Net and SegNet achieve good segmentation results in main vascular regions, but are prone to breaks or missed detections in small vascular structures. Lightweight models such as ENet and MobileNet still suffer from background missegmentation in some regions, resulting in noise points in the prediction results. In contrast, improved structures such as Nested U-Net, Attention U-Net, and Swin-Unet show improved performance in small vascular recognition, but still exhibit some structural discontinuities in complex vascular branching regions. Furthermore, in some image edge regions, these models still suffer from blurred vascular boundaries or missing small branches.

[0115] The differences between the different models can be more clearly observed from the red-marked areas in the figure. It can be seen that the DualSI-Net proposed in this application performs more stably in terms of preserving fine vascular structures, vascular connectivity, and boundary integrity. Compared with other models, DualSI-Net can more accurately recover the microvascular branch structure and effectively reduce vascular breakage and missegmentation, resulting in a more complete overall vascular network structure. DualSI-Net has stronger feature representation capabilities in complex vascular structures, accurately segmenting not only the main vessels but also preserving the fine vascular branch structure well.

[0116] To further evaluate the reliability and stability of the DualSI-Net segmentation results, this application systematically analyzed the consistency between model predictions and real labels, as well as the correlation between prediction results among models, using correlation analysis and visualization methods.

[0117] (1) Consistency analysis between DualSI-Net and labels

[0118] The Pearson correlation coefficient and Spearman rank correlation coefficient between the DualSI-Net prediction results and the ground truth labels were calculated. Experimental results show that both correlation coefficients reach 0.7137, indicating a significant positive correlation between the model prediction results and the manual annotations. This demonstrates that DualSI-Net can learn the distribution characteristics of vascular structures well at the pixel level, and its prediction results maintain a high degree of consistency with the ground truth annotations overall. For microvascular segmentation tasks, high correlation means that the model can accurately recover the spatial distribution of vascular regions, thus providing a reliable foundation for subsequent analysis of vascular morphology and microcirculation structure.

[0119] (2) Analysis of the predictive consistency between DualSI-Net and other models

[0120] To further analyze the prediction relationship between DualSI-Net and other mainstream segmentation models, this application conducted a visualization analysis from two perspectives: pixel-level scatter distribution and inter-model correlation matrix.

[0121] First, scatter plots were drawn comparing the pixel-level prediction results of DualSI-Net with those of R2U-Net, SwinUNet, NestedUNet, LSWNet, Attention U-Net, and SegNet. Figure 5 As shown in the figure, each scatter point represents the predicted value of the same pixel in the two models. When the scatter points are mainly distributed near the diagonal, it indicates that the two models have high consistency in pixel prediction. It can be observed from the figure that most of the scatter points are concentrated in the area near the diagonal, indicating that DualSI-Net has a good correlation with other models in terms of overall prediction trends. It also shows that while maintaining stable prediction characteristics, the model does not show obvious prediction shifts or abnormal distributions.

[0122] Based on this, the pairwise correlation coefficients between the prediction results of each segmentation model are further calculated and displayed in the form of a heatmap, such as... Figure 6 As shown, the results indicate that the correlation coefficient between DualSI-Net and other mainstream models is generally higher than 0.85, indicating a high degree of consistency in the overall prediction patterns among different models. This also reflects that DualSI-Net, while maintaining consistency with the prediction trends of mainstream segmentation models, can further improve vessel segmentation performance through structural optimization. This result further verifies the stability and reliability of the model's prediction results, consistent with the experimental conclusions of the quantitative evaluation indicators presented earlier.

[0123] Furthermore, to thoroughly evaluate the contributions of each module in DualSI-Net, this application designed a systematic ablation experiment, the results of which are summarized in Table 4. Using the complete model (Dice: 0.804, mIoU: 0.676) as a benchmark, the impact of each module on segmentation performance was quantitatively analyzed by removing or replacing key components one by one.

[0124] As shown in Table 4, after replacing the composite loss function with a single Dice loss, the model's Dice coefficient decreased from 0.804 to 0.791, and the mIoU decreased from 0.676 to 0.658. This result indicates that the pixel-wise gradient signal provided by the cross-entropy term in the composite loss function helps stabilize the training process and complements the region optimization of the Dice loss, thus synergistically improving segmentation accuracy and convergence stability. After removing the multi-scale feature fusion module, the model's Dice decreased to 0.788, and the mIoU decreased to 0.655. Given the data characteristics of large vessel diameter spans and the proportion of microvessels with small diameter pixels reaching 63%, this performance decrease demonstrates that multi-scale fusion plays an irreplaceable role in capturing vessel morphology at different scales. After removing the residual connections in the network, the model's Dice further decreased to 0.772, and the mIoU decreased to 0.633, indicating that the residual structure plays a crucial role in maintaining effective gradient propagation and mitigating the degradation problem of deep networks, especially in detailed segmentation tasks such as sublingual vessels, ensuring the effective transfer of shallow detailed features to deeper layers. After replacing the dual-branch network with a single-branch structure, the model's Dice dropped to 0.774, and the mIoU decreased sharply from 0.676 to 0.562. This significant difference verifies the effectiveness of the functional asymmetric design proposed in this application. The main branch provides global context to ensure the continuity of the backbone, while the auxiliary branch focuses on local details to improve the recall of small branches; both are indispensable. Finally, after significantly reducing the number of convolutional filters to compress the model size, Dice dropped to 0.736, and mIoU dropped to 0.588. This indicates that under the complexity of the current task, there is a certain balance between the model's representational power and the number of parameters; excessive compression will impair the ability to extract features from small blood vessels.

[0125] Table 4. Segmentation results of the ablation experiment

[0126] DualSI-Net complete model 0.804 0.898 0.830 0.676 0.888 Reduce the number of filters 0.736 0.865 0.782 0.588 0.824 Replace with Dice loss 0.791 0.892 0.821 0.658 0.859 Remove multi-scale fusion 0.788 0.889 0.817 0.655 0.859 Single-branch structure 0.774 0.872 0.786 0.562 0.873 Remove residual connections 0.772 0.865 0.782 0.588 0.824

[0127] In summary, the multi-scale fusion mechanism, residual connections, and bi-branch structure are the core components for the effective operation of DualSI-Net. Their synergistic effect not only enhances the model's multi-level perception of vascular morphology but also ensures stable information transfer during training, jointly supporting the model's high-performance in the sublingual microcirculation vessel segmentation task.

[0128] This embodiment addresses the challenge of segmenting sublingual microcirculatory vessels, which are characterized by delicate morphology, low contrast, and the need to balance local details with overall topology. Inspired by the collaborative mechanism of the left and right hemispheres of the brain, it constructs a functionally asymmetric dual-branch network: one branch focuses on extracting local details and fine structures, while the other focuses on modeling global context and spatial relationships. These two branches collaborate efficiently through a biomimetic interaction module, thus simulating the specialized and integrated parallel processing mode in biological intelligence. Topological analysis confirms that this structure possesses high clustering and short-path small-world characteristics. Experiments on a self-built dataset show that this model achieves segmentation accuracy comparable to complex models such as R2U-Net (Dice: 0.804, mIoU: 0.676) with only 6.27M parameters, and demonstrates significant advantages in inference speed and memory usage. Ablation studies further clarify the crucial role of the multi-scale fusion module and residual connections in maintaining model performance.

[0129] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction, characterized in that, Includes the following steps:

101. Obtain the original image of the sublingual blood vessels, and perform image standardization preprocessing on the original image of the sublingual blood vessels to obtain the image to be segmented; 102. Construct an asymmetric dual-branch lateralized lightweight segmentation network, which includes a parallel left branch network, a right branch network, and multiple sets of cross-branch feature fusion units. The left branch network and the right branch network each have an independently set encoder, bottleneck layer, and decoder.

103. Input the image to be segmented into an asymmetric dual-branch lateralized lightweight segmentation network. Feature extraction is completed through the left branch network and the right branch network respectively. Relying on the cross-branch feature fusion unit, multi-level feature interaction between the two branch networks is realized in the encoder, bottleneck layer and decoder.

104. The features that have completed interactive fusion are sequentially decoded and classified to finally output the sublingual blood vessel segmentation result image.

2. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 1, characterized in that, The left and right branch networks are functionally differentiated asymmetric branch structures. Both branches adopt a lightweight convolutional architecture and set a small-world topology connection form. The left branch network internally constructs a global feature extraction link to extract global contextual features of the image to be segmented; The right branch network internally constructs a local feature extraction link to extract local detailed features of the sublingual blood vessels in the image to be segmented.

3. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 2, characterized in that, The encoder, bottleneck layer, and decoder of the left and right branch networks are all embedded with several multi-scale cyclic residual structure modules, which are the core lightweight feature extraction units of the network. The multi-scale cyclic residual structure module integrates three types of depthwise separable convolutions, specifically including 3×3 depthwise separable convolution, 1×1 depthwise separable convolution, and dilated depthwise separable convolution.

4. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 3, characterized in that, The multi-scale cyclic residual structure module performs channel-dimensional concatenation and fusion operations on the feature maps output by the three types of depthwise separable convolutions within it. The calculation formula for the fused features is as follows: in, To input the feature map of the multi-scale cyclic residual structure module, The output features are from a 3×3 depthwise separable convolution. The output features are those of a 1×1 depthwise separable convolution. The output features of the depth-separable convolution are the dilated features. This represents the concatenation operation of the channel dimension.

5. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 4, characterized in that, In addition to the convolution operation unit, the multi-scale cyclic residual structure module is also equipped with a cyclic residual connection structure to realize cross-layer transmission of features within the module. The cyclic residual connection structure establishes cross-layer direct connections between different convolutional layers and feature operation layers within the module, directly transmitting the feature maps generated in the shallow layers of the module to the deep operation nodes of the module.

6. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 5, characterized in that, The right branch network adopts an encoder-decoder architecture, with the encoder and decoder respectively paired with multi-scale cyclic residual structure modules and skip connection paths at corresponding levels; The encoder of the right branch network is composed of multiple downsampling layers connected in series, and each downsampling layer is connected to a multi-scale cyclic residual structure module at the same level. The decoder of the right branch network consists of multiple fine-grained upsampling layers. Each upsampling layer has an independent skip connection path, which connects the corresponding layers of the right branch network encoder and decoder.

7. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 6, characterized in that, The left branch network also adopts an encoder-decoder architecture, with an overall hierarchical structure corresponding to the right branch network, and independently configured with downsampling layers, upsampling layers, and skip connection structures. The encoder of the left branch network consists of a global modeling path composed of continuous multi-level downsampling layers, with downsampling layers and multi-scale cyclic residual structure modules arranged alternately. The decoder of the left branch network includes a multi-level upsampling module, and the skip connection structure in the decoder connects the shallow feature map of the encoder of the left branch network to the corresponding operation layer of the decoder.

8. The lightweight segmentation method for hypolingual vessels based on bibranch lateralization interaction according to claim 7, characterized in that, The cross-branch feature fusion unit realizes information interaction between branches. It is divided into three categories according to the network hierarchy, which correspond to different functional areas of the network to complete feature interoperability. The cross-branch feature fusion unit includes an encoding area fusion unit, a bottleneck area fusion unit, and a decoding area fusion unit; The encoding area fusion unit connects the same level nodes of the left branch network encoder and the right branch network encoder; The bottleneck fusion unit connects the bottleneck layer of the left branch network and the bottleneck layer of the right branch network. The decoding area fusion unit connects the same level nodes of the left branch network decoder and the right branch network decoder.

9. The lightweight segmentation method for hypolingual vessels based on bi-branch lateralization interaction according to claim 8, characterized in that, When training the dual-branch lateralized lightweight segmentation network, a composite loss function is used as a constraint for parameter iteration. This composite loss function is composed of the Dice loss function and the cross-entropy loss function, specifically including: The Dice loss function formula is: ; The formula for the cross-entropy loss function is: ; The formula for the composite loss function is: ; in, These are the network pixel prediction values. The actual pixel label value. These are the weight coefficients; during training, the gradient is backpropagated based on the composite loss function to correct the weight parameters of each convolutional layer and connection structure within the network.

10. The lightweight segmentation method for hypolingual vessels based on bi-branch lateralization interaction according to claim 9, characterized in that, The dataset completes network training and segmentation operations, and is divided into different subsets according to the usage scenario and then put into use. Construct a dedicated dataset for sublingual blood vessels containing multiple images and corresponding labeled images; The dataset is divided into a training subset and a test subset; Input the training subset into the network to perform iterative parameter training, and input the test subset into the trained network to perform sublingual vessel segmentation.

Images