Dsa guidewire segmentation method and apparatus based on decoupling experts and anatomical constraints

Abstract

The application discloses a DSA guide wire segmentation method and device based on role decoupling experts and anatomical constraints, comprising: acquiring a DSA image, and extracting image features through an image encoder; generating a prompt embedding through a prompt encoder; generating a guide wire gating coefficient and a blood vessel gating coefficient according to the prompt embedding through a prompt router; generating global prediction, guide wire prediction and blood vessel prediction based on the image features through a role decoupling expert decoder with a residual routing, and performing residual fusion according to the guide wire gating coefficient and the blood vessel gating coefficient to obtain an initial guide wire mask and an initial blood vessel mask; and performing constraint correction on the initial guide wire mask according to the initial blood vessel mask through an anatomical constraint refinement module to output a final guide wire segmentation mask. The application corrects the guide wire position by using the morphological constraint between the guide wire and the blood vessel, improves the continuity and anatomical rationality of the guide wire segmentation, and is suitable for interventional navigation and intraoperative auxiliary display.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology and relates to a DSA guidewire segmentation technology solution based on role decoupling experts and anatomical constraints. Background Technology

[0002] Interventional vascular surgery is a type of minimally invasive surgery that relies on medical image guidance. In this type of surgery, the guidewire is usually an important tool for instrument navigation and path establishment, and its position, shape, and movement directly affect the advancement and manipulation of interventional instruments. Digital subtraction angiography (DSA) images can provide intraoperative imaging information on vascular structures and interventional instruments. Therefore, accurate segmentation of the guidewire in DSA images is of great significance for interventional surgery-aided navigation, guidewire contour display, path planning, and intelligent image analysis.

[0003] However, guidewires in DSA images are typically long and thin, with low contrast, blurred boundaries, and are susceptible to background interference. Because the guidewire occupies a relatively small percentage of pixels in the image, its structure is easily affected by image noise, motion blur, overlapping blood vessels, and imaging artifacts, leading to breakage, missed detection, or false detection. Existing guidewire segmentation methods mostly treat the guidewire as an independent binary segmentation target, focusing only on the pixel-level prediction results of the guidewire itself, and rarely explicitly utilizing the inherent constraints between the guidewire and vascular anatomy.

[0004] In practical interventional scenarios, guidewires are typically located inside or near the vascular lumen, and their feasible space is limited by the vascular anatomy. Therefore, guidewire segmentation is not simply a problem of segmenting a thin wire structure, but rather a type of instrument segmentation problem with anatomical constraints. Simultaneously segmenting the guidewire and the vessel and using vessel prediction to correct guidewire prediction is an important approach to reduce extravascular misdiagnosis and repair local fractures. If the segmentation method ignores the structural dependency between the guidewire and the vessel, it is prone to problems such as extravascular misdiagnosis, guidewire breakage, and topological discontinuities.

[0005] In recent years, basic visual models and deep learning segmentation models have made progress in natural and medical image segmentation tasks. For example, segmentation models based on encoder-decoder structures can extract semantic features of images and output segmentation masks, while segmentation models based on large-scale pre-training also have strong general visual representation capabilities. However, there are significant domain differences between DSA images and natural images, and directly using general segmentation models is difficult to adequately adapt to the grayscale distribution, low signal-to-noise ratio, and thin structural target features of angiography images.

[0006] Furthermore, existing models typically lack stable multi-role structural modeling mechanisms. This invention recognizes that for DSA guidewire segmentation, the guidewire structure and the blood vessel structure have different visual characteristics and functional roles: the guidewire has a slender, linear structure, while the blood vessel has a tree-like or branching anatomical structure. Using a single branch to process both simultaneously can easily lead to structural feature confusion; processing them completely independently makes it difficult to establish the spatial constraint relationship between the guidewire and the blood vessel. Therefore, a DSA guidewire segmentation system and method that can simultaneously achieve image domain adaptation, role decoupling modeling, conservative fusion, and anatomical constraint correction is needed. Summary of the Invention

[0007] The purpose of this invention is to provide a DSA guidewire segmentation technology based on role decoupling experts and anatomical constraints, in order to solve the problems existing in the current guidewire segmentation method, such as insufficient DSA image domain adaptation, easy breakage of the thin and long guidewire structure, insufficient utilization of the anatomical relationship between the guidewire and blood vessel, a large number of false detections outside the blood vessel, and insufficient model training stability.

[0008] To achieve the above objectives, this invention provides a DSA guidewire segmentation method based on role decoupling experts and anatomical constraints, comprising: A DSA image is acquired, and image features are extracted from the DSA image using an image encoder. Generate cue embeddings using a cue encoder; The router generates guidewire gating coefficients and vascular gating coefficients based on the prompts. Using a role-decoupled expert decoder with residual routing, global prediction, guidewire prediction, and vessel prediction are generated based on the image features. The global prediction, guidewire prediction, and vessel prediction are then residually fused according to the guidewire gating coefficient and the vessel gating coefficient to obtain an initial guidewire mask and an initial vessel mask. The anatomical constraint refinement module performs constraint correction on the initial guidewire mask based on the initial vascular mask, and outputs the final guidewire segmentation mask.

[0009] Furthermore, the image encoder is equipped with a low-rank adaptation module, which is inserted into the self-attention module of each Transformer block of the image encoder to achieve adaptation to the DSA image domain while retaining the main structure of the pre-trained encoder.

[0010] Furthermore, the prompt encoder is used to receive explicit or implicit prompt information and convert the prompt information into a prompt embedding that can be processed by the prompt router; when there is an explicit prompt, a prompt tag in the prompt embedding is taken as the prompt vector; when there is no explicit prompt, the prompt embedding is averaged in the spatial dimension to obtain the prompt vector.

[0011] Furthermore, the role decoupling expert decoder with residual routing includes a global expert, a guidewire expert, and a vascular expert; the global expert is used to generate the basic segmentation results, the guidewire expert is used to enhance the slender structural features of the guidewire, and the vascular expert is used to enhance the anatomical structural features of the blood vessels.

[0012] Furthermore, the global expert, the guidewire expert, and the vascular expert share a decoding backbone and output global prediction results, guidewire prediction results, and vascular prediction results respectively through corresponding prediction heads; the decoding backbone includes a bidirectional Transformer module and a mask upsampling module.

[0013] Furthermore, the prompt router is implemented using a multilayer perceptron, which is used to perform weighted control on different expert branches according to the prompt embedding, and to incorporate the prediction results of the expert branches into the global prediction results in a residual fusion manner; the guidewire gating coefficient and the blood vessel gating coefficient are two independent gating coefficients.

[0014] Furthermore, the anatomical constraint refinement module includes a connected component filtering and denoising unit, a tolerance zone construction unit, a projection constraint unit, and a topology repair unit. The connected component filtering and denoising unit is used to remove isolated noise regions in the guidewire prediction results or vessel prediction results. The tolerance zone construction unit performs an expansion operation based on the initial vessel mask to construct a vessel tolerance region. The projection constraint unit is used to restrict the guidewire prediction results to the vessel tolerance region. The topology repair unit is used to detect the guidewire skeleton endpoints and bridge and repair fracture regions that meet preset conditions.

[0015] Furthermore, a progressive training strategy is employed to train the image encoder, the cue encoder, the cue router, and the role decoupling expert decoder with residual routing. This progressive training strategy includes a global adaptation phase, an expert specialization phase, and a collaborative routing phase performed sequentially. The global adaptation phase trains the low-rank adaptation module and the global expert. The expert specialization phase trains the guidewire expert and the vascular expert, and freezes the cue router. The collaborative routing phase trains the cue router so that the guidewire prediction and the vascular prediction are weighted by the guidewire gating coefficient and the vascular gating coefficient, respectively, and then fused with the global prediction in a residual manner.

[0016] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described above.

[0017] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described above.

[0018] In the aforementioned DSA guidewire segmentation technique based on role decoupling experts and anatomical constraints, the main focus is to transform guidewire segmentation from a simple fine-line target segmentation problem into a joint segmentation problem constrained by vascular anatomy. Simultaneously, both the guidewire and the vessel are predicted. The vessel prediction results are used to construct a feasible space for the guidewire, followed by spatial projection and endpoint bridging repair of the predicted guidewire. Therefore, guidewire prediction depends not only on pixel appearance but also on the positional relationships within the vessel and topological continuity.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The DSA medical image domain adaptation is achieved through the low-rank adaptation module, which reduces the parameter overhead and overfitting risk caused by full fine-tuning; (2) By using global experts, guidewire experts and vascular experts to achieve decoupled modeling of different structural roles, the feature expression ability of guidewire and vascular structures can be improved; (3) By using a prompting router and residual fusion mechanism, expert branches are introduced to enhance the results while preserving the stability of global prediction; (4) Enhance the continuity, topological rationality and anatomical consistency of guidewire segmentation results through vascular tolerance zone, projection constraint and endpoint bridging repair; (5) By using a progressive training strategy, the system gradually transitions from global adaptation to expert specialization and collaborative routing, thereby improving the stability of the training process; (6) The system does not depend on a specific neural network backbone and can be deployed on SAM, UNet, Transformer encoder or other encoder-decoder models, and has strong versatility and scalability.

[0020] The present invention is simple and convenient to implement, highly practical, and solves the problems of low practicality and inconvenience in actual application of related technologies. It can improve user experience and has significant market value. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall structure and training strategy in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram comparing the qualitative segmentation results of the present invention and the prior art. Detailed Implementation

[0023] The following will further explain the concept, specific structure and technical effects of the present invention in conjunction with the accompanying drawings and embodiments, so as to fully understand the purpose, features and effects of the present invention.

[0024] Example 1: This invention provides a DSA guidewire segmentation method based on role decoupling expert and anatomical constraints. It sets up an image encoder, a cue encoder, a cue router, a role decoupling expert decoder with residual routing, and an anatomical constraint refinement module to achieve guidewire segmentation. The guidewire segmentation implementation process includes: A DSA image is acquired, and image features are extracted from the DSA image using an image encoder. Generate cue embeddings using a cue encoder; The router generates guidewire gating coefficients and vascular gating coefficients based on the prompts. Using a role-decoupled expert decoder with residual routing, global prediction, guidewire prediction, and vessel prediction are generated based on the image features. The global prediction, guidewire prediction, and vessel prediction are then residually fused according to the guidewire gating coefficient and the vessel gating coefficient to obtain an initial guidewire mask and an initial vessel mask. The anatomical constraint refinement module performs constraint correction on the initial guidewire mask based on the initial vascular mask, and outputs the final guidewire segmentation mask.

[0025] The image encoder receives the input DSA image and extracts its features. It includes a low-rank adaptation module to efficiently adapt parameters to the DSA image domain while preserving the main structure of the pre-trained encoder.

[0026] The cue encoder is used to generate cue embeddings. Cue embeddings can be obtained from explicit or implicit cue information and serve as input to the cue router.

[0027] The router is used to generate guidewire gating coefficients and vessel gating coefficients based on the embedded prompts. These coefficients control the degree to which guidewire and vessel experts participate in the final prediction result.

[0028] The role-decoupled expert decoder with residual routing includes a global expert, a guidewire expert, and a vascular expert. The global expert is used to generate the basic segmentation results, the guidewire expert is used to enhance the slender structural features of the guidewire, and the vascular expert is used to enhance the anatomical features of the blood vessels. The role-decoupled expert decoder incorporates the prediction results of the guidewire expert and the vascular expert into the global prediction result through residual fusion, thereby introducing expert enhancement information while maintaining the stability of the basic prediction.

[0029] The anatomical constraint refinement module is used to constrain and correct the guidewire prediction results based on the vessel prediction results. This module includes a connected component filtering and denoising unit, a tolerance zone construction unit, a projection constraint unit, and a topology repair unit. Through these units, the system can remove isolated noise regions, construct vessel tolerance regions, restrict the guidewire prediction to a reasonable anatomical range, and perform endpoint bridging repair on broken guidewire areas, ultimately outputting a guidewire segmentation mask.

[0030] Preferably, the image encoder can employ a pre-trained Segment Anything Model image encoder, and achieves efficient parameter adaptation to the DSA image domain through a low-rank adaptation module; the role-decoupled expert decoder shares a decoding backbone and outputs basic predictions and role-specific predictions through global experts, guidewire experts, and vascular experts respectively; the cue router generates guidewire gating coefficients and vascular gating coefficients based on cue embedding, and fuses the expert outputs in a conservative residual manner; the anatomical constraint refinement module constructs a vascular tolerance zone using the vascular prediction results, and performs projection constraints and endpoint bridging repair on the guidewire prediction, thereby improving the continuity, stability, and anatomical rationality of guidewire segmentation. Since the guidewire is usually located inside or near the vascular cavity in real interventional surgery, this invention uses the morphological constraints between the guidewire and the blood vessel to correct the guidewire position, which is suitable for interventional navigation, intraoperative auxiliary display, and intelligent medical image analysis.

[0031] Example 2: This embodiment provides a DSA guidewire segmentation system based on role decoupling experts and anatomical constraints, including an image encoder, a cue encoder, a cue router, a role decoupling expert decoder with residual routing, and an anatomical constraint refinement module. The image encoder can employ a pre-trained Segment Anything Model image encoder and achieves efficient parameter adaptation to the DSA image domain through a low-rank adaptation module. The role decoupling expert decoder shares a decoding backbone and outputs basic predictions and role-specific predictions through global experts, guidewire experts, and vessel experts, respectively. The cue router generates guidewire gating coefficients and vessel gating coefficients based on cue embeddings and fuses expert outputs using a conservative residual approach. The anatomical constraint refinement module constructs a vessel tolerance zone using vessel prediction results and performs projection constraints and endpoint bridging repair on the guidewire prediction, thereby improving the continuity, stability, and anatomical rationality of guidewire segmentation. Since guidewires are usually located inside or near the vascular cavity in real interventional surgery, this invention utilizes morphological constraints between the guidewire and the vessel to correct the guidewire position, making it suitable for interventional navigation, intraoperative auxiliary display, and intelligent medical image analysis.

[0032] In one possible embodiment, a DSA guidewire segmentation system based on role decoupling experts and anatomical constraints is provided, comprising: (1) Image encoder, used to receive input DSA images and extract image features; The image encoder is equipped with a low-rank adaptation module to achieve adaptation to the DSA image domain while retaining the main structure of the pre-trained encoder.

[0033] (2) A cue encoder, used to generate cue embeddings; The prompt encoder is used to receive explicit or implicit prompt information and convert the prompt information into a prompt embedding that can be processed by the prompt router.

[0034] (3) A prompt router is used to embed and generate guidewire gating coefficients and vascular gating coefficients based on the prompts; The prompt router is used to perform weighted control on different expert branches according to the prompt embedding, and to incorporate the prediction results of the expert branches into the global prediction results in a residual fusion manner.

[0035] (4) A role decoupling expert decoder with residual routing, used to generate global prediction, guidewire prediction and blood vessel prediction based on the image features respectively, and to fuse them according to the guidewire gating coefficient and blood vessel gating coefficient; The role decoupling expert decoder with residual routing includes a global expert, a guidewire expert, and a vascular expert, wherein: Global experts are used to generate basic segmentation results; Guidewire experts use it to enhance the slender structural features of guidewires; Vascular specialists use it to enhance the anatomical features of blood vessels.

[0036] The global expert, guidewire expert, and vascular expert share the decoding backbone and output global prediction results, guidewire prediction results, and vascular prediction results respectively through their corresponding prediction heads.

[0037] The role decoupling expert decoder with residual routing outputs an initial guidewire mask and an initial vascular mask, which are then input into the anatomical constraint refinement module.

[0038] (5) Anatomical constraint refinement module, which is used to constrain and correct the guidewire prediction results based on the blood vessel prediction results and output the final guidewire segmentation mask.

[0039] The anatomical constraint refinement module includes: 1) Connectivity filtering and denoising unit; The connected component filtering and denoising unit is used to remove isolated noise regions from the guidewire prediction results or blood vessel prediction results.

[0040] 2) Tolerance band building blocks; The tolerance zone construction unit constructs a vascular tolerance region based on the vascular prediction mask, which serves as the spatial constraint range for the guidewire prediction results.

[0041] 3) Projection constraint unit; The projection constraint unit is used to limit the guidewire prediction results to the vascular tolerance area in order to remove unreasonable guidewire responses outside the vascular area.

[0042] 4) Topology repair unit.

[0043] The topology repair unit improves the connectivity of the guidewire segmentation results by detecting the endpoints of the guidewire skeleton and bridging and repairing fractured areas that meet preset conditions.

[0044] The system employs a progressive training strategy, including a global adaptation phase, an expert specialization phase, and a collaborative routing phase. The global adaptation phase is used to train the low-rank adaptation module and the global prediction branch.

[0045] The expert specialization phase is used to train guidewire specialists and vascular specialists, and to freeze the prompt router.

[0046] The cooperative routing phase is used to freeze the expert head and train the cue router, so that the expert branches gradually participate in the generation of the final prediction result.

[0047] Example 3: The preferred implementation scheme provided in this embodiment is as follows: I. System Overall Structure like Figure 1 As shown in (a), this embodiment provides a DSA guide wire segmentation system based on role decoupling expert and anatomical constraints, including an image encoder, a cue encoder, a cue router, a role decoupling expert decoder with residual routing, and an anatomical constraint refinement module.

[0048] The input DSA image is first fed into an image encoder. The image encoder is used to extract the structural features of blood vessels, guide wires, and background in the image. To adapt to the imaging characteristics of DSA images, the image encoder includes a low-rank adaptation module. This low-rank adaptation module is used to introduce a small number of trainable parameters while freezing or partially freezing the pre-trained encoder parameters, enabling the encoder to adapt to the grayscale distribution, low-contrast targets, and thin structural features of angiography images.

[0049] The cue encoder generates a cue embedding, which is then input to the cue router. The cue router outputs guidewire gating coefficients and vessel gating coefficients based on the cue embedding. The guidewire gating coefficients control the contribution of the guidewire expert branch to the final prediction result, while the vessel gating coefficients control the contribution of the vessel expert branch to the final prediction result.

[0050] The role-decoupled expert decoder with residual routing receives image features output from the image encoder and generates global, guidewire, and vessel predictions through global, guidewire, and vessel expert methods, respectively. Subsequently, the decoder generates initial guidewire and vessel masks via residual fusion. These initial guidewire and vessel masks are further input into the anatomical constraint refinement module to generate the final guidewire segmentation mask.

[0051] II. Image Encoder and Low-Rank Adaptor Module In one embodiment, the image encoder employs a pre-trained Segment Anything Model (SAM) image encoder, specifically a visual Transformer structure. The SAM image encoder itself is a component of existing pre-trained basic segmentation models, and its basic structure will not be described in detail. The improvement of this invention lies in its application to the joint segmentation of DSA guidewires and blood vessels, and the addition of a low-rank adaptation module at its self-attention projection to complete the medical image domain adaptation.

[0052] The low-rank adaptation module adopts the LoRA form. Specifically, for the original projection matrix in the image encoder, the adapted projection matrix can be obtained through a low-rank update method, with the following form:

[0053] in, Represents the original projection matrix. This indicates a low-rank update term. Represents a low-rank fitness matrix. Let be a real number, and d represent the feature dimension of the projection matrix in the image encoder, i.e., the hidden layer dimension of each Transformer block in the pre-trained image encoder. In one embodiment, the rank... = 4. Matrix A is randomly initialized, and matrix B is initialized to 0, so that ΔW is approximately 0 at the beginning of training, thus maintaining the initial output of the pre-trained image encoder. The model output basically maintains the original behavior of the pre-trained encoder, and then gradually learns DSA domain features during the training process.

[0054] The image features after low-rank adaptation can be represented as:

[0055] in, This represents the feature representation of the input image. This represents the image features after low-rank adaptation.

[0056] The low-rank adaptation module is specifically inserted at the query projection q and value projection v in the self-attention module of each Transformer block of the SAM image encoder, while the key projection K remains unchanged.

[0057] In this way, the system enables the encoder to adapt to DSA grayscale distribution, low signal-to-noise ratio and thin structure targets without changing the structure of the SAM main encoder.

[0058] III. Role Decoupling Expert Decoder like Figure 1 As shown in section (a), the role-decoupled expert decoder with residual routing includes a global expert, a guidewire expert, and a vascular expert. These three types of experts are not three completely independent U-Net decoders, but rather different prediction heads are set on a shared SAM mask decoding backbone. The shared components include a Two-Way Transformer, a mask upsampling module, and shared upsampling features.

[0059] The Two-Way Transformer and the mask upsampling module are cascaded together to form the decoding backbone. The Two-Way Transformer performs bidirectional cross-attention fusion on the image features output by the image encoder and the cue embedding output by the cue encoder, outputting the fused image features and cue token. The mask upsampling module progressively upsamples the low-resolution fused feature map output by the Two-Way Transformer to the original image resolution. The global expert, guidewire expert, and vascular expert share the Two-Way Transformer and mask upsampling module, separating them only at each target prediction head to achieve parameter-efficient role decoupling modeling.

[0060] In one specific embodiment, the Transformer channel dimension is set to C. The mask upsampling module includes a first transposed convolutional layer, a two-dimensional normalization layer, a GELU activation layer, a second transposed convolutional layer, and a GELU activation layer. The first transposed convolutional layer has a kernel size of 2 and a stride of 2, transforming the number of channels from C to C / 4; the second transposed convolutional layer has a kernel size of 2 and a stride of 2, transforming the number of channels from C / 4 to C / 8. The above upsampling layers are shared by global experts, guidewire experts, and vascular experts.

[0061] The global expert uses the original masked supernetwork prediction head of the basic SAM mask decoder. This branch retains the general segmentation capability of the base model, providing a stable foundation for subsequent expert enhancement. This prediction head outputs a supernetwork vector based on the mask token and multiplies it with shared upsampled features to obtain the global prediction L_global. The guidewire expert and vascular expert each add a lightweight supernetwork prediction head, both employing a three-layer multilayer perceptron: input dimension C, hidden dimension C, and output dimension C / 8. The first two layers use ReLU activation, and the last layer outputs a supernetwork vector. Both the guidewire expert and vascular expert prediction heads use the shared global token output as conditional input and multiply the output vector with the same shared upsampled features to obtain L_wire and L_vessel, respectively.

[0062] Because the guidewire expert and vascular expert only add two lightweight multilayer perceptron prediction heads without replicating the Two-Way Transformer and upsampling layer, the number of parameters and memory overhead can be reduced, while avoiding structural redundancy caused by completely independent decoders. The global expert focuses on the overall segmentable region, the guidewire expert focuses on slender, low-contrast, and locally continuous features, and the vascular expert focuses on vascular regions, orientation, and branching structures.

[0063] IV. Prompt for router and residual fusion After the cue encoder generates the cue embedding, the cue embedding is input to the cue router. The cue embedding output by the cue encoder can be a sparse cue embedding of shape [B,N,D] or a dense cue embedding of shape [B,D,H,W], where B is the batch size, D is the cue embedding dimension, and N represents the number of sparse cue tokens contained in each image. In the SAM ViT-B embodiment, D is 256. When there are explicit cues such as points or boxes, a cue token from the sparse cue embedding can be taken as the cue vector p, where p belongs to R^D; when there are no explicit cues, the dense embeddings without cues can be averaged over spatial dimensions H and W to obtain p belonging to R^D. Here, R^D represents a D-dimensional real vector space.

[0064] The router is implemented using a two-layer multilayer perceptron to output the guidewire gating coefficient and the blood vessel gating coefficient. z = ReLU(W1 p + b1), [ , = sigmoid(W2 z + b2). Where p∈R^D represents the cue vector output by the cue encoder; z∈R^128 represents the hidden layer vector of the router; W1∈R^(128×D) represents the weight matrix of the first linear mapping layer; b1∈R^128 represents the bias vector of the first linear mapping layer; W2∈R^(2×128) represents the weight matrix of the second linear mapping layer; b2∈R^2 represents the bias vector of the second linear mapping layer; and Let W1 represent the guidewire gating coefficient and the vessel gating coefficient, respectively. W1 maps the D-dimensional cue vector to a 128-dimensional hidden vector, and W2 maps the 128-dimensional hidden vector to a 2-dimensional output. The sigmoid function constrains the two outputs to the interval [0,1]. The two gating coefficients are independent and are not required to be... + = 1, therefore guidewire specialists and vascular specialists can be enhanced simultaneously, inhibited simultaneously, or participate in fusion at different intensities.

[0065] To ensure initial training stability, the weights and biases of the last layer of the prompt router can be initialized to zero, and the expert contribution is controlled by the residual injection coefficient alpha. When alpha is 0, the final prediction degenerates into the global expert output; as alpha gradually increases, the expert output participates in the fusion as residuals. The fusion expression is:

[0066] in, This represents the global expert output. This indicates the output results from the guidewire expert. This indicates the results output by the vascular specialist. Indicates the guidewire gating factor. Indicates the vascular gating coefficient. This represents the residual injection coefficient. This conservative residual fusion method (see...) Figure 1 The residual logit fusion always retains the global prediction path, avoiding catastrophic degradation caused by insufficient training of expert branches.

[0067] Through the aforementioned residual fusion method, the global expert branch is always retained in the final prediction path, which avoids significant degradation of prediction results due to insufficient training of expert branches. Simultaneously, guidewire experts and vascular experts can gradually participate in the prediction through gating, thereby improving the continuity of the guidewire's thin structure and the rationality of vascular anatomy.

[0068] V. Anatomical Constraint Refinement Module like Figure 1As shown on the right side of section (a), the anatomical constraint refinement module is used for post-processing correction of the initial guidewire mask. This module utilizes the real surgical anatomical constraint that "the guidewire is located inside the vessel or in a feasible region adjacent to the vessel" to correct the guidewire prediction after network output. This module is not a simple smoothing post-processing, but rather transforms the vessel prediction into a guidewire feasible space, and performs denoising, projection, and topology repair within this space.

[0069] In one embodiment, the anatomical constraint refinement module includes the following steps.

[0070] First, connected component filtering and noise reduction are performed. The system filters isolated connected regions in the initial guidewire mask and initial vascular mask, removing noise responses that are clearly not part of the guidewire or vascular structure.

[0071] Second, construct the vascular tolerance zone. The system performs a dilate operation based on the initial vascular mask to obtain the vascular tolerance region. This process can be represented as:

[0072] in, This refers to the vascular prediction mask (Vessel). Indicates the radius of expansion. This represents the constructed vascular tolerance band.

[0073] Third, projection constraints are applied. The system confines the guidewire prediction mask within the vessel tolerance zone to eliminate unreasonable guidewire responses outside the vessel area. This process can be represented as:

[0074] in, This indicates the guide wire prediction mask (Wire).

[0075] Fourth, topology repair is performed. The system performs skeletonization processing on the guidewire prediction mask and detects the guidewire skeleton endpoints. For endpoint pairs whose spatial distance meets preset conditions, the system performs bridging repair within the vascular tolerance zone. This process can be represented as:

[0076] Through the above steps, the anatomical constraint refinement module can reduce extravascular misdiagnosis, repair local guidewire breakage, and improve the continuity, topological integrity, and anatomical rationality of the final guidewire segmentation mask.

[0077] VI. Progressive Training Strategy like Figure 1As shown in section (b), this invention employs a progressive training strategy of conservative expert integration to alleviate expert collapse and gating saturation problems caused by class imbalance in thin structures. This training strategy includes a global adaptation phase, an expert specialization phase, and a cooperative routing phase.

[0078] During the global adaptation phase, the low-rank adaptation module and global prediction path are trained, the cue router, guidewire, and vascular expert prediction head are frozen, and the residual injection coefficient alpha is set to 0. Figure 1 middle a This stage enables the model to first acquire basic segmentation capabilities for DSA images.

[0079] During the expert specialization phase, the cue router remains frozen while training the guidewire expert prediction head and the vascular expert prediction head, enabling them to learn the slender structure of the guidewire and the dendritic structure of the blood vessels, respectively. In this phase, auxiliary supervision loss can continue to constrain the guidewire and vascular expert outputs, and the residual injection coefficient alpha remains 0.

[0080] During the collaborative routing phase, the training prompts the router and can jointly fine-tune the guidewire expert prediction head and the vascular expert prediction head, allowing expert predictions to participate in the final fusion through a gated residual method. During training, alpha can be gradually increased from 0 to 1, or it can be set to 1 after the experts have stabilized.

[0081] Through the aforementioned progressive training strategy, the system can gradually introduce expert enhancement capabilities while maintaining global prediction stability, thereby improving model training stability and guide wire segmentation performance.

[0082] In one embodiment, the training process can employ a multi-branch joint loss function:

[0083] in, Indicates the global branch loss. Indicates guidewire branch loss. This indicates loss of vascular branches. and These represent the weighting coefficients for guidewire branches and vascular branches, respectively. The loss for each branch can be composed of Dice loss, cross-entropy loss, or a combination of both.

[0084] An example training configuration includes: AdamW optimizer, learning rate 0.005, LoRA rank 4, a combination of Dice loss and cross-entropy loss, and Dice weights of 0.8.

[0085] VII. Experimental Results To verify the effectiveness of this invention, experiments were conducted on a clinically acquired DSA image dataset. This dataset comprises 3069 DSA images from approximately 20 patients, covering various clinical scenarios including aneurysm repair, internal carotid artery stenosis treatment, and interventional procedures for vascular malformations. All images have pixel-level double annotations for both guidewires and blood vessels, and a patient-level segmentation method was used to avoid images of the same patient appearing simultaneously in both the training and test sets. The training set includes 2584 images, and the test set includes 485 images.

[0086] Statistical data shows that guidewires occupy an average of only 0.64% ± 0.30% of image pixels, indicating that they are typical small-scale, slender targets; blood vessels occupy an average of 13.27% ± 4.66% of image pixels, exhibiting complex tree-like and branching anatomical structures. Furthermore, the overlap between guidewires and blood vessels in the actual annotations reaches 99.7%, indicating a stable anatomical coupling relationship between the guidewire position and the vascular structure. Therefore, this invention models guidewire segmentation as a joint segmentation problem constrained by the vascular anatomy, rather than a standalone binary guidewire segmentation problem.

[0087] Figure 2 The comparison of segmentation performance of different methods on representative DSA test images is presented. As can be seen from the figures, the comparative methods are prone to local breakage, missed detection, or endpoint misalignment in areas with slender guidewire structures. Some methods also produce unreasonable responses in the extravascular region. In contrast, the RDE-SAM + Refinement of this invention more closely approximates the actual spatial correspondence between the guidewire and the blood vessel, more completely preserving the continuous morphology of the guidewire and reducing spurious guidewire responses outside the blood vessel.

[0088] The orange circles highlight typical defects of different methods, such as guidewire breakage, missing wires, or structural discontinuities. This invention, through vascular tolerance zone projection constraints and endpoint bridging repair, allows guidewire prediction to fall more within the feasible vascular area and repairs short-distance local breaks, thus maintaining good structural continuity even in complex vascular courses and low-contrast regions.

[0089] The qualitative results further illustrate that the improvement of the present invention is not only reflected in the pixel overlap index, but also in the ability to improve the topological integrity and anatomical rationality of the guidewire as a continuous interventional device, which is more in line with the requirements of intraoperative auxiliary navigation for a stable and continuous guidewire profile.

[0090] Table 1 shows the quantitative comparison results of the present invention on the DSA test set. The method of the present invention achieved the highest Dice and the lowest HD95 on the DSA test set. Specifically, the RDE-SAM + Refinement of the present invention achieved a Dice of 84.2% and an HD95 of 2.36 pixels, which is superior to the comparison methods such as DARL, Medical-SAM-adapter, H-SAM, SAMed, and FGA-Net.

[0091] Compared to DARL, which is specifically designed for slender or vascular structures, this invention not only improves Dice but also significantly reduces HD95, indicating that it better maintains guidewire boundaries and structural continuity. Compared to SAM-type medical adaptation methods, this invention, while maintaining the expressive power of the pre-trained model, further reduces structural errors by introducing role-specific modeling of guidewires and blood vessels through role decoupling experts and residual routing. Compared to H-SAM, which performs relatively well, this invention reduces HD95 from 6.89 pixels to 2.36 pixels, demonstrating its significant advantages in maintaining the continuity of slender guidewire structures, endpoint position stability, and suppressing extravascular false detections.

[0092] The above results show that relying solely on general medical segmentation adaptation methods is insufficient to fully address the issues of low contrast, thinness, and easy breakage of DSA guidewires. In contrast, this invention, by jointly modeling guidewires and blood vessels and using vascular anatomy to constrain and correct guidewire predictions, can more effectively improve guidewire segmentation quality.

[0093] Table 1. Quantitative comparison results of the present invention on the DSA test set.

[0094] To analyze the contribution of each component to the final performance, further ablation experiments were conducted, and the results are shown in Table 2. When SAM + LoRA was used as the basic model, the model achieved efficient parameter adaptation to the DSA image domain through low-rank adaptation, achieving 80.2% Dice and 9.25 pixels of HD95.

[0095] Building upon this, the addition of a role decoupling expert improved Dice to 82.8% and reduced HD95 to 8.32 pixels, indicating that the guidewire expert and vascular expert can respectively learn the slender structure of the guidewire and the tree-like anatomical structure of the blood vessels, thereby enhancing the expressive ability of different target roles.

[0096] With the addition of a cue router, Dice improved to 83.9%, while HD95 significantly decreased to 2.98 pixels. These results demonstrate that the cue router can adaptively adjust the contributions of guidewire and vascular experts based on cue embedding, and incorporate expert predictions into the global prediction through conservative residual fusion, thereby improving thin structure continuity and boundary accuracy while maintaining the stability of the basic predictions.

[0097] Finally, after adding the anatomical constraint refinement module, Dice was further improved to 84.2%, and HD95 was reduced to 2.36 pixels. These results demonstrate that vascular tolerance zone projection and endpoint bridging repair can further utilize the anatomical constraint relationship between the guidewire and vessel after network prediction, reducing extravascular misdiagnosis and repairing short local fractures.

[0098] Ablation experiments show that the low-rank adaptation, role decoupling expert, cue router, and anatomical constraint refinement module in this invention all contribute to the final performance. Among them, the cue router and anatomical constraint refinement module have a particularly significant effect on improving HD95, indicating that this invention plays a key role in improving the continuity and boundary stability of the guidewire structure.

[0099] Table 2 Ablation Experiment Results of the Components of the Present Invention

[0100] To verify the architecture independence of the framework of this invention, it was also validated on different backbone models, and the results are shown in Table 3, demonstrating its generalization ability. When the role decoupling expert and residual routing concepts of this invention are applied to the UNet backbone, Dice increases from 80.7% to 82.0%, and HD95 decreases from 41.2 pixels to 40.0 pixels, indicating that the framework can enhance the modeling ability of the convolutional backbone for the relationship between guidewires and vascular structures.

[0101] When applied to the SAM backbone, Dice increased from 80.2% to 83.9%, and HD95 decreased from 9.25 pixels to 2.98 pixels, demonstrating a more significant performance improvement. This indicates that the high-capacity pre-trained image encoder can provide a strong basic representation for DSA images, while the role decoupling expert and conservative residual routing of this invention further introduce structural constraints between the guidewire and the blood vessel, thereby significantly improving the continuity and spatial rationality of segmentation of slender instruments.

[0102] The above results show that the present invention does not depend on a specific neural network backbone, but can be used as a general anatomically constrained segmentation framework that can be transferred to different encoder-decoder structures for medical image segmentation tasks involving guide wires, blood vessels, and other medical images with anatomical spatial dependencies.

[0103] Table 3 Generalization Proof

[0104] In summary, the experimental results validated the effectiveness of this invention in four aspects: quantitative indicators, qualitative effects, module ablation, and cross-bone generalization. Compared with existing guidewire segmentation or medical SAM adaptation methods, this invention not only improves Dice overlap accuracy but also significantly reduces HD95 boundary distance error, indicating that it can improve the structural continuity and spatial positioning stability of the guidewire. This performance improvement mainly stems from the synergistic effect between guidewire-vessel joint modeling, role decoupling expert, cueing residual routing, and anatomical constraint refinement.

[0105] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0106] The following describes the DSA guidewire segmentation electronic device based on role decoupling expert and anatomical constraints provided by the present invention. The DSA guidewire segmentation electronic device based on role decoupling expert and anatomical constraints described below can be referred to in correspondence with the DSA guidewire segmentation method based on role decoupling expert and anatomical constraints described above.

[0107] The electronic device may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute a DSA wire segmentation method based on role decoupling experts and anatomical constraints, primarily including the software processing components described above.

[0108] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the software processing part of the DSA guide wire segmentation method based on role decoupling experts and anatomical constraints provided by the above methods.

[0110] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the software processing portion of the DSA guidewire segmentation method based on role decoupling experts and anatomical constraints provided by the above methods.

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A DSA guidewire segmentation method based on role decoupling experts and anatomical constraints, characterized in that: include: A DSA image is acquired, and image features are extracted from the DSA image using an image encoder. Cue embeddings are generated using a cue encoder; The router generates guidewire gating coefficients and vascular gating coefficients based on the prompts. Using a role-decoupled expert decoder with residual routing, global prediction, guidewire prediction, and vessel prediction are generated based on the image features. The global prediction, guidewire prediction, and vessel prediction are then residually fused according to the guidewire gating coefficient and the vessel gating coefficient to obtain an initial guidewire mask and an initial vessel mask. The anatomical constraint refinement module performs constraint correction on the initial guidewire mask based on the initial vascular mask, and outputs the final guidewire segmentation mask.

2. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints according to claim 1, characterized in that: The image encoder is equipped with a low-rank adaptation module, which is inserted into the self-attention module of each Transformer block of the image encoder. It is used to adapt to the DSA image domain while retaining the main structure of the pre-trained encoder.

3. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 1, characterized in that: The prompt encoder is used to receive explicit or implicit prompt information and convert the prompt information into a prompt embedding that can be processed by the prompt router. When an explicit prompt exists, a prompt tag in the prompt embedding is taken as the prompt vector. When no explicit prompt exists, the prompt embedding is averaged in the spatial dimension to obtain the prompt vector.

4. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 1, characterized in that: The role decoupling expert decoder with residual routing includes a global expert, a guidewire expert, and a vascular expert; the global expert is used to generate the basic segmentation results, the guidewire expert is used to enhance the slender structural features of the guidewire, and the vascular expert is used to enhance the anatomical structural features of the blood vessels.

5. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 4, characterized in that: The global expert, the guidewire expert, and the vascular expert share a decoding backbone and output global prediction results, guidewire prediction results, and vascular prediction results respectively through their corresponding prediction heads; the decoding backbone includes a bidirectional Transformer module and a mask upsampling module.

6. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 1, characterized in that: The prompt router is implemented using a multilayer perceptron, which is used to perform weighted control on different expert branches according to the prompt embedding, and to incorporate the prediction results of the expert branches into the global prediction results in a residual fusion manner; the guidewire gating coefficient and the blood vessel gating coefficient are two independent gating coefficients.

7. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 1, characterized in that: The anatomical constraint refinement module includes a connected component filtering and denoising unit, a tolerance zone construction unit, a projection constraint unit, and a topology repair unit. The connected component filtering and denoising unit is used to remove isolated noise regions in the guidewire prediction results or vessel prediction results. The tolerance zone construction unit performs an expansion operation based on the initial vessel mask to construct a vessel tolerance region. The projection constraint unit is used to restrict the guidewire prediction results to the vessel tolerance region. The topology repair unit is used to detect the guidewire skeleton endpoints and bridge and repair fracture regions that meet preset conditions.

8. The DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in claim 1, characterized in that: A progressive training strategy is employed to train the image encoder, the cue encoder, the cue router, and the role-decoupled expert decoder with residual routing. This progressive training strategy includes a global adaptation phase, an expert specialization phase, and a collaborative routing phase performed sequentially. The global adaptation phase trains the low-rank adaptation module and the global expert. The expert specialization phase trains the guidewire expert and the vascular expert, and freezes the cue router. The collaborative routing phase trains the cue router such that the guidewire prediction and the vascular prediction are weighted by the guidewire gating coefficient and the vascular gating coefficient, respectively, and then fused with the global prediction in a residual manner.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the DSA guidewire segmentation method based on role decoupling experts and anatomical constraints as described in any one of claims 1 to 8.

Images