Medical image optimization method and system based on blood vessel branch selective virtualization

Abstract

The application discloses a medical image optimization method and system based on blood vessel branch selective virtualization. The method comprises the following steps: pre-processing and blood vessel enhancement are performed on a three-dimensional angiography DICOM image; an encoder-decoder network containing a wide activation and residual hollow space pyramid module is used for accurate blood vessel segmentation; through blood vessel topology analysis, interference branches that shield target blood vessel structures under a specific working angle are identified; based on a Poisson equation and a bidirectional convolution LSTM, selective virtualization repair is performed on the interference branches to generate an optimized image that is visually unshielded and maintains topological continuity. A safety mechanism of virtual-real combined display and multi-angle preplan is introduced to ensure the reliability of surgical navigation. Finally, the optimized 3D blood vessel model is integrated into a contrast system supporting real-time synchronization and used as a roadmap. The application can effectively solve the clinical problem that the best working angle cannot be used due to blood vessel shielding, and significantly improves the accuracy and safety of intravascular intervention surgery.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more particularly to a medical image optimization method and system based on selective blurring of vascular branches. This invention is particularly applicable to the real-time optimization of three-dimensional angiographic pathway maps during endovascular interventional surgery. By blurring vascular branches that obscure the target structure, it provides clear visual navigation for the surgeon to find and ensure the optimal working angle, thereby improving the accuracy and safety of the surgical procedure. Background Technology

[0002] Angiographic images are crucial medical imaging evidence for diagnosing cardiovascular diseases and cerebrovascular lesions. Currently, clinically used angiography machines support the use of 3D angiographic images as roadmaps for endovascular interventional procedures. This system can synchronize the angle of the 3D angiographic roadmap displayed on the monitor with the real-time acquired angiographic angle, providing doctors with important spatial positioning references during the procedure.

[0003] However, the clinical application of this technology, especially in endovascular interventional treatments with extremely high visualization requirements (such as intracranial aneurysm embolization), still faces significant challenges. Taking the interventional embolization of middle cerebral artery aneurysms or anterior communicating artery aneurysms as an example, the key to surgical success lies in the operator obtaining an ideal two-dimensional fluoroscopic working angle. At this angle, the two-dimensional image must simultaneously and clearly display all key target vascular (abnormal) structures without obstruction, including: the parent artery, the aneurysm neck, the aneurysm body, and its related branches. Under this perspective, the operator needs to precisely insert the microcatheter into the aneurysm cavity and ensure that the inserted coils densely fill the aneurysm body and neck, without protruding into the parent artery or accidentally occluding normal branch vessels.

[0004] However, such aneurysms often occur at densely packed confluences of blood vessels, making it difficult to find an ideal working angle in clinical practice due to obstruction by unrelated arteries. This manifests in two ways: first, due to the overlapping and obstruction of multiple vessels, an angle that meets the above requirements cannot be found at all; second, while a slight rotation of the 3D vascular map may reveal an ideal angle, at this angle, an unrelated artery may be obscuring the crucial target structure. Both situations force the surgeon to abandon the optimal operating perspective, significantly increasing the difficulty and risk of the surgery.

[0005] At the image processing level, traditional vascular image optimization methods (such as the Frangi filter based on the Hessian matrix) can enhance tubular structures, but their enhancement effect on low-contrast small vessels is limited, and they are prone to amplifying noise. Conventional segmentation methods are insufficient in maintaining the continuity of complex vascular topology, making it difficult to accurately distinguish between real vessels and artifacts. Existing technologies often use simple interpolation or region filling to process interfering branches, lacking a comprehensive consideration of the continuity and spatial consistency of vascular physiological structures. This results in unnatural vascular morphology after processing and fails to address the core clinical need for selective removal of occlusions at specific working angles.

[0006] In recent years, deep learning technology has been applied to vascular segmentation tasks. However, conventional encoder-decoder networks still have limitations in recognizing complex vascular structures, multi-scale branches, and microvessels, especially when vessels overlap with lesion areas, which can easily lead to missegmentation or missed segmentation. In addition, existing methods rarely involve intelligent recognition of interfering vascular branches or selective blurring to ensure surgical safety, and they lack the ability to deeply integrate with real-time roadmap systems to provide surgeons with multi-angle pre-planning strategies.

[0007] Therefore, there is an urgent need in the field for a medical image optimization method and system that can be seamlessly integrated with a real-time roadmap system, intelligently identify and process obstructed blood vessels at specific working angles, thereby providing clear, safe and reliable image navigation support for endovascular interventional surgery. Summary of the Invention

[0008] The purpose of this invention is to provide a medical image optimization method and system based on selective blurring of vascular branches, which is particularly suitable for use in 3D angiography roadmap systems with real-time synchronization, to assist doctors in quickly locating the optimal working angle and clearly displaying the target vascular structure during endovascular interventional surgery.

[0009] In a first aspect, embodiments of this application provide a medical image optimization method based on selective vascular branch blurring, the method comprising:

[0010] S1. Acquire 3D angiography DICOM images and perform grayscale conversion, contrast normalization, and gamma correction;

[0011] S2. The image is enhanced with a multi-scale Hessian matrix Frangi filtering algorithm, and the visibility of low-contrast blood vessels is further improved by combining contrast-limited adaptive histogram equalization.

[0012] S3. Use a deep learning segmentation network based on an encoder-decoder structure to perform vascular tree structure segmentation on the image. The network includes a combination of extensive activation modules and residual void spatial pyramid modules, and is trained and optimized through a joint regularization strategy.

[0013] S4. Identify and mark vascular branches that overlap with or interfere with the target vascular structure through vascular topology analysis and spatial location relationship; the target vascular structure includes the parent artery, aneurysm neck, aneurysm body and related branches that need to be clearly displayed in interventional surgery.

[0014] S5. Selectively blur the marked interfering blood vessel branches, and use an image editing method based on the Poisson equation combined with bidirectional convolutional LSTM to perform spatiotemporal feature fusion on the processed area to generate an optimized image that does not visually obstruct the view and maintains the continuity of blood vessels.

[0015] S6. Preserve the spatial information and topological structure of the original image to generate an optimized three-dimensional angiography image;

[0016] S7. Surgical safety assurance procedures, which include:

[0017] S7.1, Combined Real and Virtual Display: The interfering vascular branches processed in step S5 are displayed in the optimized image in a semi-transparent or dashed form, so as to eliminate visual obstruction and indicate their anatomical location to the surgeon.

[0018] S7.2 Multi-angle contingency planning: Based on the same original three-dimensional angiography image, multiple different working angles and corresponding optimized images are pre-planned and generated for a single interventional surgical operation. Each optimized image at each angle independently executes steps S1 to S6 to blur the interfering branches that obscure the target vascular structure at that angle.

[0019] S8. Output the optimized 3D angiography image in DICOM format and integrate it into an angiography system that supports real-time synchronization. Use it as a roadmap for endovascular interventional surgery to help the operator find the best working angle and safely navigate the interventional device.

[0020] Secondly, embodiments of this application provide a medical image optimization system based on selective vascular branch blurring, applied to the medical image optimization method based on selective vascular branch blurring as described in the first aspect, the system comprising:

[0021] The image preprocessing module is used to acquire 3D angiography DICOM images and perform grayscale conversion, contrast normalization, and gamma correction.

[0022] The vessel enhancement module is used to enhance the blood vessels in the image using a multi-scale Hessian matrix Frangi filtering algorithm, and further improves the visibility of low-contrast blood vessels by combining contrast-limited adaptive histogram equalization.

[0023] The blood vessel segmentation module is used to segment the image into a blood vessel tree structure using a deep learning segmentation network based on an encoder-decoder structure. The network includes a combination of a broad activation module and a residual void spatial pyramid module, and is trained and optimized through a joint regularization strategy.

[0024] The interference branch identification module is used to identify and mark vascular branches that overlap with or interfere with the target vascular structure through vascular topology analysis and spatial location relationship;

[0025] The branch blurring module is used to selectively blur the marked interfering blood vessel branches using an image editing method based on the Poisson equation, and combines bidirectional convolutional LSTM to perform spatiotemporal feature fusion on the processed area to generate an optimized image that does not visually obstruct the view and maintains the continuity of the blood vessels.

[0026] The image output module is used to preserve the spatial information and topological structure of the original image, generate optimized three-dimensional angiography images, and output them in DICOM format.

[0027] The surgical safety control module controls the branch blurring processing module to display interfering branches as semi-transparent or dashed lines; it manages the generation and retrieval of multiple pre-planned working angles and their corresponding optimized images based on the same original image; and it integrates the optimized images into the angiography system for real-time navigation of endovascular interventional surgery.

[0028] Thirdly, embodiments of this application provide an electronic device, including:

[0029] processor;

[0030] Memory used to store processor-executable instructions;

[0031] The processor is configured to implement the medical image optimization method based on selective vascular branch blurring as described in the first aspect when executing the instructions.

[0032] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to perform the medical image optimization method based on selective blurring of vascular branches as described in the first aspect.

[0033] This invention utilizes a series of innovative technologies, including multi-scale vascular enhancement, deep learning segmentation, topology analysis, and intelligent repair, to achieve effective identification and precise blurring of interfering branches in angiographic images, while maintaining the anatomical rationality and spatial continuity of the vascular tree. Compared with traditional methods, this invention significantly improves the clarity and diagnostic value of vascular images, providing more reliable image support for clinical surgical navigation and preoperative planning.

[0034] Beneficial effects:

[0035] 1. Precise service for interventional surgery navigation: This invention directly addresses the core challenge of selecting the working angle in endovascular interventional surgery. Through intelligent image processing, it enables the surgeon to obtain an unobstructed view of the target vascular structure under the optimal fluoroscopic angle, greatly improving the success rate and safety of delicate operations such as aneurysm embolization.

[0036] 2. Excellent vessel enhancement and segmentation results: Through the synergistic enhancement of multi-scale Frangi filtering and CLAHE, as well as the segmentation network combining wide activation and hollow spatial pyramid, the segmentation accuracy of complex vascular topology is significantly improved. In particular, the visibility and segmentation integrity of low-contrast small vessels are guaranteed, laying a solid foundation for subsequent accurate identification of interfering branches.

[0037] 3. Intelligent and accurate identification of interference branches: Based on vascular topology analysis and deep learning spatiotemporal features, it can intelligently distinguish between real vascular structures and interference branches, ensuring that only specific blood vessels that affect the display of the target area are marked, effectively avoiding misprocessing caused by segmentation errors.

[0038] 4. Natural and coherent vascular blurring: The image editing based on the Poisson equation and the spatiotemporal feature fusion of bidirectional convolutional LSTM are used to blur the interfering branches while maintaining the physiological continuity and visual naturalness of the vascular tree. The processed area is seamlessly connected with the surrounding blood vessels, avoiding structural breaks or morphological distortions.

[0039] 5. Ensuring surgical safety while improving clarity: An innovative dual safety mechanism is introduced, combining virtual and real-view display with multi-angle contingency planning. The blurred display eliminates visual obstruction while retaining crucial anatomical references, preventing surgeon misjudgment; the multi-angle alternatives provide redundant safety assurance for complex situations, effectively reducing surgical risks.

[0040] 6. Seamless integration with clinical workflow: The optimized images retain the original DICOM information and can be directly imported into angiography systems with real-time synchronization capabilities for use as high-definition roadmaps. This perfectly integrates into existing interventional surgical procedures, enhancing the intuitiveness and convenience of surgical navigation.

[0041] 7. High degree of automation and strong practicality: The fully automated processing from image preprocessing to branch blurring repair significantly improves the efficiency and consistency of vascular image optimization, can meet the needs of real-time or near-real-time clinical processing, and has broad clinical application prospects. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of a medical image optimization method based on selective blurring of blood vessel branches, provided in an embodiment of this application.

[0043] Figure 2 The system architecture diagram of the medical image optimization system based on selective blurring of blood vessel branches provided in this application.

[0044] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] 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 a part of the embodiments of this application, and not all of them.

[0046] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0047] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] This invention provides a medical image optimization method based on selective blurring of vascular branches. Its core objective is to address the clinical challenge in endovascular interventional surgeries (such as intracranial aneurysm embolization) where obstruction by unrelated vascular branches prevents the acquisition of a clear two-dimensional perspective view of the target vascular structure (including the parent artery, aneurysm neck, aneurysm body, and related branches). This method intelligently processes three-dimensional angiography images, selectively blurring obstructed vessels at specific viewpoints, and integrates with the real-time path mapping system of modern angiography machines, providing surgeons with optimal visual guidance and safety assurance.

[0049] Example 1

[0050] Figure 1 This is a schematic flowchart of a medical image optimization method based on selective vascular branch blurring, provided as an embodiment of this application. Figure 1 As shown, a medical image optimization method based on selective blurring of blood vessel branches includes:

[0051] S1. Acquire 3D DICOM angiography images and perform grayscale conversion, contrast normalization, and gamma correction. Image preprocessing involves standardizing the original 3D DICOM angiography images to eliminate imaging differences and optimize image quality. Specifically, grayscale conversion converts the original image to grayscale, reducing data dimensionality; contrast normalization unifies the image contrast distribution, enhancing structural consistency; and gamma correction adaptively adjusts image brightness and detail, improving the distinction between blood vessels and the background. This step provides a high-quality image foundation for subsequent processing, which is crucial for accurately identifying millimeter-level microvessels in interventional procedures.

[0052] Specifically, in this embodiment, S1 includes:

[0053] (1.1) The original DICOM format angiography image is preprocessed. The image is a grayscale image. The signal-to-noise ratio of the vascular structure is maximized in the early stage of grayscale conversion to provide better input features for subsequent processing.

[0054] (1.2) Perform contrast normalization on the grayscale image to adjust its pixel value distribution to a standard normal distribution with a mean of 0 and a standard deviation of 1. Specifically, the conversion formula is:

[0055] ,

[0056] in, The mean of the image data. The standard deviation is used for contrast standardization to unify the distribution of image data. It adjusts the pixel value distribution to a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating brightness differences between different images and improving the stability and convergence efficiency of deep learning model training.

[0057] (1.3) A contrast-limited adaptive histogram equalization method is used to locally enhance the standardized image. This involves dividing the image into multiple local regions and performing histogram cropping and equalization within each region to enhance the vascular structure and suppress noise. Specifically, the image is divided into multiple non-overlapping local regions. A gray-level histogram is calculated within each region, and a preset cropping threshold is used to crop the histogram to limit excessive enhancement of local contrast. Subsequently, the cropped pixel gray levels are uniformly redistributed to the entire histogram, and adaptive histogram equalization is performed based on the processed histogram. This enhances the vascular structure in low-contrast regions while suppressing the amplification of background noise. Contrast-limited adaptive histogram equalization enhances local contrast by performing histogram cropping and equalization in different regions. This effectively suppresses excessive amplification of background noise while enhancing the vascular structure in low-contrast regions, balancing local detail enhancement with noise control.

[0058] (1.4) Adaptive gamma correction is performed on the enhanced image. The gamma value is dynamically set according to the global contrast statistics of the image to optimize the overall brightness and detail of the image. Specifically, gamma correction is performed on the enhanced image to further optimize the overall brightness and detail of the image. The gamma value γ of the enhanced image is adaptively set according to the global contrast statistics of blood vessels and background in the original image: First, the mean and standard deviation of the pixel gray level of the entire image or region of interest are calculated, and the contrast index is calculated accordingly. If the index is lower than the preset threshold, a correction strategy of γ<1 is adopted to improve the overall brightness and highlight the small blood vessels hidden in the dark area. If the index is higher than the preset threshold, a correction strategy of γ>1 is adopted to suppress the bright area, enrich the intermediate gray level details, and avoid overexposure of blood vessels. Through this adaptive gamma correction, the brightness and detail of blood vessel images under different imaging conditions are optimized in a differentiated manner, laying the foundation for subsequent accurate segmentation.

[0059] (1.5) Normalize the image by linearly scaling its pixel values ​​to the [0,1] interval based on the effective grayscale range to adapt to the training and inference of the subsequent deep learning model. Cooperative normalization optimizes the data distribution to adapt to the deep learning model. By linearly transforming the effective grayscale range after removing outliers, the pixel values ​​are mapped to the [0,1] interval. This works in conjunction with the batch normalization layer and weight regularization method in the network to accelerate model convergence and improve the stable recognition ability of microvessels.

[0060] The image after the above processing is normalized. This operation is not a simple global linear scaling, but a collaborative preprocessing based on the requirements of the extensive activation module and the joint regularization training mechanism: First, based on the overall gray-level histogram distribution of the image after gamma correction, outliers at both ends of the distribution caused by lesion areas or high-brightness noise are identified and removed; then, a linear transformation is performed based on the effective gray-level range after removal, mapping the pixel values ​​to the [0,1] interval; this normalization strategy aims to work in synergy with the batch normalization layer and weight regularization method in the network to jointly constrain the deviation of the data distribution during training, thereby accelerating model convergence and improving its stable recognition ability of vascular structural features, especially microvessels.

[0061] Each step forms a progressive image preprocessing pipeline. Through a chain of processes—grayscale conversion, standardization, local enhancement, global correction, and normalization—image quality is gradually optimized, ultimately providing high-quality input data with regular distribution, prominent features, and controlled noise for deep learning segmentation networks.

[0062] S2. The image is enhanced with a Frangi filtering algorithm based on a multi-scale Hessian matrix, and further improved in low-contrast vessel visibility by combining it with contrast-limited adaptive histogram equalization. Vessel enhancement highlights vascular structural features and suppresses noise interference. Specifically, Frangi filtering enhances tubular structures based on a multi-scale Hessian matrix; CLAHE fusion supplements vascular details in low-contrast areas, preventing the loss of small vessels. This step ensures that even in low-contrast areas, small perforating vessels associated with the target aneurysm can be effectively enhanced, laying the foundation for accurate segmentation and subsequent blurring processing.

[0063] Specifically, in this embodiment, the vascular enhancement pretreatment in S2 includes:

[0064] The initial enhancement is performed using the Frangi filtering algorithm based on the multi-scale Hessian matrix. By calculating the Hessian matrix feature values ​​of each pixel in the image at different scales, a blood vessel similarity function is constructed to enhance tubular structures and suppress non-vascular noise. The scale selection range covers the statistical distribution range of blood vessel diameters in the target blood vessel database.

[0065] The output image of the Frangi filter is weighted and fused with the original grayscale image after contrast-limited adaptive histogram equalization. The Frangi filter result has a higher weight to highlight the main vascular structure, while the CLAHE processing result retains an appropriate weight to maintain the natural texture of the tissue background and supplement the small vascular details that may be lost during the filtering process.

[0066] The fused enhanced image is used as input to a subsequent deep learning segmentation network based on an encoder-decoder structure. This network utilizes the advanced semantic extraction capabilities of its encoder part to further optimize and confirm the vascular features initially highlighted by the aforementioned hybrid enhancement method.

[0067] Specifically, in this embodiment, step (2) includes:

[0068] (2.1) Vascular enhancement using the Frangi filtering algorithm with a multi-scale Hessian matrix: Enhancement is performed based on vascular morphological features. Within a predefined set of scales... The above is processed, including the scale parameter. Covers the range of vessel diameters in the target vascular database. For each pixel, at each scale... Next, calculate its Hessian matrix. And find the eigenvalues ​​of the matrix. and (satisfy ).

[0069] Measurement of vascular similarity at this scale Calculated using the following formula:

[0070] ,

[0071] in: This is the difference ratio, used to distinguish between linear and patchy structures. For structural strength, it is used to filter low-contrast areas such as noise. and The constant hyperparameter is used to control the sensitivity of each component.

[0072] Ultimately, the Frangi filter response for that pixel Take the maximum response value across all scales:

[0073] ,

[0074] The vascular similarity function is calculated using the eigenvalues ​​of the multi-scale Hessian matrix. By leveraging the tubular (linear) structural characteristics of blood vessels, the vascular signal is enhanced at multiple scales, while effectively suppressing noise and plaque interference from non-vascular structures.

[0075] (2.2) The Frangi filtering result Enhanced results with CLAHE Adaptive weighted fusion: Adaptive weighted fusion combines the advantages of different enhancement methods. The fusion formula is defined as follows:

[0076] ,

[0077] Weight The image is dynamically adjusted based on the contrast characteristics of local regions, and the calculation formula is as follows:

[0078] ,

[0079] in, This is the base weighting coefficient, set to a value greater than 0.5 to ensure that the Frangi filtering result dominates the fusion process. It is the standard deviation of pixel intensity within a local window centered on the current pixel, used to quantify local contrast. It is an adjustment coefficient that controls the degree to which local contrast affects the weight.

[0080] This adaptive mechanism ensures that in areas of high contrast and clear structure, the weights are adjusted accordingly. Approaching 1, mainly relying on strong noise immunity. In areas with low contrast and rich detail, the weighting Appropriately reduce and introduce more Detailed information is obtained. The Frangi filtering result and the CLAHE processing result are weighted and fused. The Frangi filtering result has a dominant weight to highlight the main vascular structure and ensure noise resistance; the CLAHE processing result retains an appropriate weight to maintain the natural texture of the tissue background and supplement the small vascular details that may have been lost during the filtering process.

[0081] (2.3) The final enhanced image The image is then fed into a subsequent deep learning segmentation network. This provides optimized input to the segmentation network, bridging the image enhancement and segmentation stages. The fused enhanced image is then used as input to the subsequent deep learning segmentation network. This network, leveraging its powerful feature learning capabilities, can further optimize and confirm the vascular features initially highlighted by the aforementioned hybrid enhancement method, achieving a smooth transition from low-level image processing to high-level semantic segmentation.

[0082] The steps constitute a hybrid enhancement strategy. It is not a simple chaining of two enhancement methods, but rather a complementary enhancement result achieved through model-based enhancement in (2.1) and adaptive fusion in (2.2). The Frangi filter in (2.1) has the advantage of utilizing explicit prior knowledge of vascular morphology, but it lacks sensitivity to low-contrast, small vessels; the CLAHE filter in (2.2) effectively enhances local contrast, but lacks shape specificity. The fusion of the two utilizes morphological priors while preserving details, and finally, (2.3) provides the segmentation network with high-quality input that is more feature-rich and noise-controlled, laying a solid foundation for subsequent accurate segmentation.

[0083] S3. A deep learning segmentation network based on an encoder-decoder structure is used to segment the image into a vascular tree structure. This network includes a combination of a wide activation module and a residual spatial pyramid module, and is trained and optimized through a joint regularization strategy. Through vascular segmentation, the vascular tree topology is accurately extracted. Specifically, the encoder-decoder network achieves end-to-end pixel-level vascular segmentation; the wide activation module and the spatial pyramid module enhance multi-scale feature extraction capabilities; and the joint regularization strategy improves model generalization and segmentation accuracy. High-precision vascular segmentation is a prerequisite for accurately identifying which vessels need to be blurred, directly affecting the accuracy of the final roadmap guidance.

[0084] Specifically, in this embodiment, the deep learning segmentation network based on the encoder-decoder structure mentioned in S3 is a wide activation segmentation network, and its specific structure includes:

[0085] (3.1) Feature Extraction Section: Composed of at least three cascaded width-optimized residual units. Each unit increases the channel dimension of the feature map before nonlinear activation by a channel expansion factor r (usually set to r=4), while compressing the channel number of shortcut connection paths by factor r. This expands the network width without significantly increasing the total number of network parameters, thus strengthening the forward propagation of underlying vascular morphology features. Each width-optimized residual unit is followed by a 2×2 max-pooling layer for spatial downsampling. The feature extraction section (encoder) efficiently extracts multi-scale vascular features. The number of feature channels is expanded before activation by width-optimized residual units, and the number of channels is compressed in shortcut connections. This design aims to increase the network width and representational ability without significantly increasing the number of parameters, strengthening the capture and forward propagation of vascular morphology features. The pooling layer gradually compresses the spatial size, expanding the receptive field.

[0086] Multi-scale information aggregation: Between the feature extraction and feature reconstruction parts, a multi-branch dilated convolution fusion module is introduced. This module consists of four parallel paths. Each path uses a dilated convolutional layer with a kernel size of 3×3 but different dilation rates. The dilation rates are set to d=1, 2, 4, and 8, respectively, to avoid the loss of spatial details caused by downsampling and to simultaneously capture contextual information covering different receptive fields. Finally, the output features of each path are additively fused.

[0087] The multi-scale information aggregation section captures multi-scale contextual information, avoiding the loss of detail. Parallel pathways are constructed using dilated convolutional layers with different dilation rates, simultaneously acquiring contextual information from different receptive fields without downsampling. This effectively resolves the contradiction between the loss of spatial detail caused by downsampling and the expansion of the receptive field, which is crucial for identifying vascular branches with significantly different diameters.

[0088] Feature reconstruction part: Its hierarchical structure is symmetrical to that of the feature extraction part. It gradually restores the spatial size of the feature map through upsampling operation, and uses cross-layer connection to stitch the high-resolution feature map of the corresponding layer of the feature extraction part with the upsampled feature map of the feature reconstruction part in the channel dimension, so as to fuse high-level semantic information and low-level positional details. Each upsampled block in the feature reconstruction part is also followed by a width-optimized residual unit.

[0089] The feature reconstruction section (decoder) gradually restores spatial details to achieve precise localization. Through upsampling operations and cross-layer connections, high-resolution detail information from the encoder path is fused with high-level semantic information from the decoder path. This symmetrical structure ensures that the segmentation result possesses both accurate semantics and precise alignment with the edges of blood vessels.

[0090] The joint regularization strategy specifically involves: during network training, batch normalization and weight parameter renormalization are performed simultaneously after each convolutional layer; the weight parameter renormalization decomposes the weight vector w of the convolutional kernel into a direction vector v and an magnitude scalar g for independent learning, represented as... The gradients of g and v are calculated and updated during backpropagation to stabilize the training process and improve the generalization performance of the model.

[0091] Joint regularization strategies enhance model training stability and generalization ability. Batch standardization mitigates internal covariate bias and accelerates training; weight parameter restandardization decouples the direction and magnitude of weights and optimizes them independently, helping to constrain model complexity, prevent overfitting, and thus improve the model's generalization performance on unseen data, especially in medical images with varied vascular morphology.

[0092] The aforementioned network architecture is a meticulously designed system: the encoder is responsible for understanding the image and extracting essential features; the multi-scale aggregation module is the core innovation, acting as a contextual information fusion unit to ensure the network can simultaneously see the forest (large blood vessels) and the trees (small blood vessels); the decoder is responsible for depicting the image, reconstructing the understood features into an accurate segmentation mask. The consistent regularization strategy acts as a stabilizer, ensuring that this complex model can be trained effectively and stably. These four parts work together to achieve high-precision and robust segmentation of complex vascular tree structures.

[0093] S3.1. The forward propagation process of the width-optimized residual unit is as follows: Let the module input be... The number of channels in the compressed quick connection path is The number of channels in the feature map before nonlinear activation is expanded to And satisfy ,in This is the channel expansion factor.

[0094] Then the output of this unit Represented as:

[0095] ,

[0096] Where σ is the ReLU activation function, and These represent the operations of the two convolutional layers preceding and following each other on the expansion path. To achieve channel expansion while maintaining similar computational complexity, the number of channels in the compressed shortcut connection path is... Number of channels compared to standard residual blocks Satisfying Relationship: Efficient feature transformation and information transfer are achieved through a forward propagation mechanism that optimizes the width of residual units. This is accomplished using the formula... Define the residual learning structure, including the channel expansion factor. and relation This design ensures that feature representation is enhanced by increasing network width within limited computational complexity. It balances network width with parameter efficiency, strengthening the capture of subtle vascular features.

[0097] S3.2. Output of the multi-branch dilated convolution fusion module It is obtained by summing the outputs of each parallel branch:

[0098] ,

[0099] in, Input for the module, Indicates the expansion rate 3×3 dilated convolution operation, The values ​​are 1, 2, 4, and 8 respectively. This indicates a re-standardization operation on the weight parameters. The activation function is used. The integration method of the multi-branch dilated convolution fusion module achieves effective fusion of multi-scale contexts. This is achieved through the formula... An additive fusion method for parallel multi-branch structures is defined. Dilated convolutions with different dilation rates (d=1,2,4,8) synchronously capture information from local details to global context, and feature integration is achieved through summation operations to enhance the model's adaptability to changes in vascular scale.

[0100] S3.3. The weight parameter re-standardization operation The weight tensor of the convolutional layer Decomposed into direction vectors and amplitude scalar To learn:

[0101] ,

[0102] During training, stochastic gradient descent is used to calculate... and Calculate and update the gradient:

[0103] , ,

[0104] in, For loss function, For the projection matrix, They are respectively , , The gradient; the specific implementation of weight parameter re-standardization, improving training stability and generalization ability. Through decomposition The weight learning is decomposed into independent optimizations of direction (v) and magnitude (g). This mechanism calculates the gradient separately during backpropagation, making weight updates more stable, effectively suppressing overfitting, and improving the model's robustness in complex blood vessel segmentation tasks.

[0105] S3.4. Composite loss function used in network training By binary cross-entropy loss Compared to loss based on similarity coefficient Linear combination constitutes:

[0106] ,

[0107] Where λ is the balancing hyperparameter and the similarity loss term. Defined as:

[0108] ,

[0109] here, For real labels, To predict probabilities, K is a smoothing constant. The construction of the composite loss function fully supervises the model training process. Through... It combines binary cross-entropy loss and similarity coefficient loss. Ensure pixel-level classification accuracy. Enhancing regional consistency and complementing each other allows the model to simultaneously focus on local boundary accuracy and overall structural integrity.

[0110] S4. Through vascular topology analysis and spatial location relationships, identify and mark vascular branches that overlap with or interfere with the target vascular (abnormal) structure. The target vascular structure includes the parent artery, aneurysm neck, aneurysm body, and related branches that need to be clearly displayed during interventional surgery. By identifying interfering branches, locate vascular branches that overlap with or interfere with the target area. Specifically, construct a vascular connectivity map through topology analysis; set spatial location criteria and filter candidate branches through 3D bounding box intersection analysis; confirm the authenticity of branches using spatiotemporal features through continuity verification. This step is crucial for achieving selective processing. The system does not process all non-target vessels, but intelligently identifies specific branches that obscure key structures from a specific viewpoint.

[0111] Specifically, a vascular connectivity graph is constructed through topological analysis; spatial location determination is set, and candidate branches are screened through three-dimensional bounding box intersection analysis; and the authenticity of branches is confirmed by utilizing spatiotemporal features through continuity verification.

[0112] Specifically, in this embodiment, the process of identifying and marking interfering vascular branches in S4 specifically includes:

[0113] S4.1. Extract the vessel centerline based on the vessel segmentation results, construct the topological connectivity graph of the vessel tree, and record the morphological and spatial information of the vessel segments. Vessel topology reconstruction: Based on the initial vessel segmentation results obtained in step 3, extract the vessel centerline using a skeletonization algorithm and construct the topological connectivity graph of the vessel tree. ; in the picture Nodes represent the bifurcation points or endpoints of blood vessels, while edges... Representing the vessel segment, and simultaneously recording the average diameter of each vessel segment. ,length And spatial coordinate information (three-dimensional coordinate sequence of the centerline of the blood vessel segment).

[0114] S4.2. Calculate the three-dimensional spatial bounding box of the target region and identify all vascular branches that intersect with the bounding box as candidate interference branches. Target region spatial relationship determination: Calculate the three-dimensional spatial bounding box of the target region (e.g., aneurysm, stenosis) and correlate its spatial location with the vascular topology map; identify all vascular branches that pass through or intersect with the bounding box and preliminarily determine them as candidate interference branches.

[0115] Specifically, let the three-dimensional spatial bounding box of the target region (such as an aneurysm) be... For each vascular segment Calculate its centerline and bounding box Spatial relationships. If an intersection exists, it is marked as a candidate interference branch: ,in For vascular segment The set of centerline coordinates. Among them, For the first Each vascular segment (Edge) For the set of candidate interference branches, For the target region (such as an aneurysm), a three-dimensional bounding box. These are the minimum and maximum coordinates of the bounding box on each axis. To determine whether the centerline of the vessel segment is aligned with the bounding box There is a spatial intersection.

[0116] S4.3. Utilize the spatiotemporal features learned by the bidirectional convolutional LSTM module in the segmentation network to analyze the continuity of candidate interference branches and verify whether they represent real vascular structures. Continuity-assisted verification: Utilize the spatiotemporal features learned by the bidirectional convolutional LSTM module embedded in the segmentation network to enhance vascular continuity, analyze the continuity of candidate interference branches. Perform continuity analysis; if the vascular segments upstream and downstream of the target region are segmented into continuous structures with high confidence by the network, then the branch is confirmed to be a real blood vessel overlapping the target region, rather than noise introduced by segmentation errors.

[0117] Specifically, let the upstream and downstream feature sequences of the branch be... The continuity confidence score is calculated using LSTM units: where, where It is the first The feature vectors at each location (derived from the multi-scale features of the segmentation network decoder). ,like ( For example, a preset threshold. If the result is positive, then the branch is confirmed to be a real vascular structure rather than noise. It is a bidirectional convolutional LSTM module, simultaneously from the forward ( ) and backward ( Process the feature sequence and output the fused spatiotemporal features. The continuity confidence level represents the probability that the branch is a real blood vessel.

[0118] S4.4. Mark the vascular branches identified as overlapping or interfering with the target area, and attach vascular level, diameter, and positional relationship attribute information. Precise marking: Based on the above topology analysis, spatial relationship determination, and continuity verification results, precisely mark the vascular branches identified as overlapping or interfering with the target area, and attach vascular level, diameter, and positional relationship attribute information relative to the target area to each marked branch.

[0119] Specifically, for the interference branches that pass the verification, their attribute information is added: vascular level. (Based on topological distance, grading from trunk to tip); diameter Positional relationship relative to the target area (e.g., traversing adjacent areas). Generate a set of markers: .in, For the set of labels of the interference branches, For the marked interference branches, This is the vascular hierarchy, graded according to topological distance from the trunk (lower level) to the terminal (higher level). The average diameter of the vessel segment. For example, regarding the spatial relationship with the target area: Crossing: Blood vessels pass through the interior of the target area. Adjacent: The blood vessel is in contact with the boundary of the target area but does not penetrate it.

[0120] S5. Selectively blur the marked interfering vascular branches. A Poisson equation-based image editing method combined with bidirectional convolutional LSTM is used to fuse spatiotemporal features in the processed area to maintain vascular continuity and structural integrity. Branch blurring selectively blurs interfering branches and restores vascular continuity. Specifically, Poisson equation editing is used to achieve seamless pixel filling based on gradient fields; bidirectional convolutional LSTM is used to fuse spatiotemporal context information to ensure the structural naturalness of the processed area. In this step, "restoration" refers to generating a visually continuous and natural vascular path at the original location of the blurred vessel, but its display intensity (e.g., grayscale value) is reduced, or rendered in a semi-transparent, dashed mode, thus achieving a visually unobstructed but anatomically present blurring effect.

[0121] Specifically, in this embodiment, the process of selectively blurring the marked branch in step S5 includes:

[0122] Based on the vascular topology connectivity map, the upstream and downstream connection points of the marked interfering branches are determined, and a smooth virtual vascular path conforming to the vascular physiological morphology is generated as a repair guide. Topology-based repair guidance: Based on the established vascular topology connectivity map, the upstream and downstream connection points of the marked interfering branches are determined; using these upstream and downstream connection points as constraints, a smooth virtual vascular path conforming to the vascular physiological morphology that traverses the target region is generated as a structural guide for image repair.

[0123] Specifically, the vascular topology map established based on the above steps Locating the marked interference branch Upstream and downstream connection points and Generate a smooth virtual vascular path that conforms to the physiological morphology of blood vessels. Its generation formula is based on cubic spline interpolation:

[0124] ,

[0125] Where the coefficient Determined by boundary conditions: Position constraints: , (T represents the tangent direction of the original blood vessel at the connection point). The topological information and spatial relationship attributes provided in the above steps are directly utilized. Generate an anatomically sound repair guide path. The three-dimensional spatial coordinates of the virtual blood vessel path are parameters. vector functions, For path parameters, starting from the origin To the finish line Continuous change. It is a cubic spline coefficient matrix (each dimension of the coordinate corresponds to a set of coefficients). The coordinates of the upstream connection point, These are the coordinates of the downstream connection point. The direction of the tangent at the upstream point. , The tangent direction at the downstream point. .

[0126] The processing region is defined centered on a virtual path. The Poisson equation is used to apply the texture and gradient information of healthy blood vessels as boundary conditions, achieving seamless filling of the processing region. Poisson image editing and content filling: The processing region is defined centered on the generated virtual path. An image editing method based on the Poisson equation is used, applying the texture and gradient information of healthy blood vessels at the boundary of the processing region as boundary conditions. By solving the Poisson equation, seamless filling of pixel values ​​within the processing region is achieved, ensuring visual continuity.

[0127] Specifically, using virtual paths Define the region Ω to be processed centered at the center. The core formula for editing the Poisson equation is:

[0128] , ,

[0129] in, The pixel values ​​within the processing area to be solved are... The guiding vector field is derived from gradient information from healthy blood vessel regions. , An enhanced image of the healthy blood vessel region. These are the pixel values ​​at the boundaries of the original image. To handle region boundaries, For the Laplace operator, This is the divergence operator. It utilizes the enhanced vascular texture information from the above steps. As a source of healthy vascular gradients, it ensures visual consistency in the treated area.

[0130] The preliminary repair results and the context feature maps from the segmentation network are input into a bidirectional convolutional LSTM network. Spatiotemporal sequence processing is used to optimize the repair results, ensuring vascular continuity and structural naturalness. Spatiotemporal feature fusion optimization: The preliminary repair results after Poisson editing are combined with features extracted at the corresponding levels by the segmentation network decoder, which contain complete vascular context information. Figure 1 The input is fed into a bidirectional convolutional LSTM network; this network processes the spatial sequence information around the processing area in both forward and backward directions, learns the extension pattern of blood vessels, optimizes the preliminary repair results, and outputs a final repaired image with higher spatiotemporal consistency, ensuring that the repaired blood vessels remain natural and continuous both visually and structurally.

[0131] Specifically, the preliminary results of Poisson editing Feature maps extracted by the segmentation network decoder at the corresponding levels Input bidirectional convolutional LSTM:

[0132] ,

[0133] ,

[0134] The final optimized output is:

[0135] ,

[0136] in, For forward convolution LSTM at time The hidden state, For backward convolution LSTM at time The hidden state, Edited for Poisson in position Preliminary repair results This refers to the context feature maps extracted by the segmentation network decoder at the corresponding layers. Forward convolution LSTM unit, For backward convolution LSTM units, This is a convolution fusion operation. This is the final optimized repair result.

[0137] The bidirectional convolutional LSTM module is directly invoked to optimize the repair results using the spatiotemporal extension pattern of blood vessels it learns.

[0138] Calculate the change in centerline curvature and diameter uniformity of the repaired blood vessel. If any indicators are abnormal, return to the step of adjusting the virtual path parameters for iterative optimization. Integrity verification: Calculate the change in centerline curvature and diameter uniformity of the repaired blood vessel in this region. If any indicators are abnormal, return to the step of adjusting the virtual path parameters for iterative optimization until the repair result meets biomechanical characteristics.

[0139] Specifically, integrity verification involves calculating morphological parameters of the repaired blood vessels:

[0140] Changes in centerline curvature: ,

[0141] Pipe diameter uniformity: ( This represents the average diameter of blood vessels in the processing area. (The standard deviation of the vessel diameter in the treatment area).

[0142] in, For path in arc length parameter The curvature value at that point, It is the first derivative of the path (tangent vector). The second derivative of the path (curvature vector). Let the vector magnitude be , For vector cross product,

[0143] like or Then return to the steps above to adjust the virtual path parameters for iterative optimization. Ensure the repair result meets the core objective of maintaining vascular continuity and structural integrity. The curvature threshold is typically set to 0.5-1.0 mm. -1 . The pipe diameter uniformity index is between 0 and 1, with the closer to 1 being more uniform. The uniformity threshold is typically set to 0.7-0.8. These parameters together form a complete mathematical framework from path generation and image restoration to result verification, ensuring the accuracy and physiological rationality of the vascular repair process.

[0144] S6. Preserving the spatial information and topological structure of the original image, generate optimized 3D angiography images and output them in DICOM format for integration with angiography systems as real-time navigation maps for endovascular interventional procedures. Image generation and output: Generate optimized 3D vascular images while maintaining clinical usability. Topological consistency reconstruction ensures the anatomical rationality of the vascular tree; DICOM format output is compatible with clinical diagnostic systems and supports surgical navigation and planning.

[0145] Specifically, in this embodiment, the process of generating the optimized three-dimensional angiography image in step S6 includes:

[0146] The 2D sequence segmentation results after branch blurring are reconstructed in 3D. Using the initial vascular topology connectivity map as constraints, consistency checks and smoothing are performed on potential topological breakpoints. Topological Consistency Reconstruction: The 2D sequence segmentation results after branch blurring are reconstructed in 3D. Using the established initial vascular topology connectivity map as constraints, consistency connectivity checks and smoothing are performed on potential topological breakpoints caused by branch blurring, ensuring the topological connectivity and anatomical rationality of the final 3D vascular tree.

[0147] The optimized vessel segmentation result is used as a foreground mask and fused with the background voxels of the original 3D DICOM image, fully preserving the spatial coordinates, grayscale values, and all positioning information in the DICOM file header of the original image. Lossless spatial information integration: The optimized vessel segmentation result is used as a foreground mask and fused with the background voxels of the original 3D DICOM image; during this process, the spatial coordinates, grayscale values, and all spatial positioning information (such as pixel spacing, slice thickness, image position, and orientation) of the unmodified voxels in the original image are fully preserved.

[0148] Automatically generate a structured report containing the number, location, and original diameter information of the blurred vascular branches, as well as key morphological parameters of the optimized vascular tree. Generating the structured report: Automatically generate a structured report associated with the optimized image. This report includes at least the number, location, and original diameter information of the blurred vascular branches, as well as key morphological parameters of the optimized vascular tree, including total length, number of bifurcation points, and average diameter of a specific region. The report is embedded in DICOM SR (Structured Report) format or XML format or output along with the image data.

[0149] The fused 3D image data and its structured report are packaged into a standard DICOM format file for use in clinical surgical navigation systems or preoperative planning software. Output and Integration: The fused 3D image data and its structured report are packaged into a standard DICOM format file, ensuring direct import into clinical surgical navigation systems or preoperative planning software for augmented reality visualization, blood flow simulation analysis, or precise surgical path planning.

[0150] As a key innovation of this invention, following S6, it also includes an application process to ensure surgical safety:

[0151] S7, Surgical Safety Navigation Application:

[0152] S7.1, Combined Real-Virtual Display: The optimized 3D vascular model is loaded into an angiography system that supports real-time synchronization. In the system, the blurred vascular branches are displayed in a semi-transparent or dashed line mode, so as not to obscure the target vascular structure, while providing the operator with a reference for its spatial position, preventing accidental entry into non-target vessels when operating microcatheters or microwires due to an overly clean path map.

[0153] S7.2 Multi-angle contingency planning: Before surgery, doctors can pre-process and save multiple alternative working angles and their corresponding optimized roadmaps based on the same set of raw data. When difficulties or uncertainties arise during operation at a certain angle, the surgeon can switch to another alternative angle with one click to verify the instrument position from different perspectives, ensuring surgical safety.

[0154] Example 2

[0155] like Figure 2 As shown, this application provides a medical image optimization system architecture diagram based on selective vascular branch blurring, which is applied to the medical image optimization system based on selective vascular branch blurring as described in Embodiment 1. It includes an image preprocessing module 11, a vascular enhancement module 12, a vascular segmentation module 13, an interference branch recognition module 14, a branch blurring processing module 15, an image output module 16, and a surgical safety control module 17.

[0156] Image preprocessing module 11 is used to acquire three-dimensional angiography DICOM images and perform grayscale conversion, contrast normalization and gamma correction.

[0157] The vessel enhancement module 12 is used to enhance the blood vessels in the image using a multi-scale Hessian matrix Frangi filtering algorithm, and further improve the visibility of low-contrast blood vessels by combining contrast-limited adaptive histogram equalization.

[0158] The blood vessel segmentation module 13 is used to segment the image into a blood vessel tree structure using a deep learning segmentation network based on an encoder-decoder structure. The network includes a combination of a broad activation module and a residual void spatial pyramid module, and is trained and optimized through a joint regularization strategy.

[0159] Interference branch identification module 14 is used to identify and mark vascular branches that overlap with or interfere with the target area through vascular topology analysis and spatial location relationship.

[0160] The branch blurring processing module 15 is used to selectively blur the marked interfering vascular branches using an image editing method based on the Poisson equation, and combines bidirectional convolutional LSTM to perform spatiotemporal feature fusion on the processed area to maintain the continuity and structural integrity of the blood vessels.

[0161] Image output module 16 is used to retain the spatial information and topological structure of the original image, generate an optimized three-dimensional angiography image and output it in DICOM format for integration into the surgical navigation system.

[0162] Furthermore, the system also includes a surgical safety control module 17 connected to the image output module. This module is used to: control the branch blurring processing module to achieve virtual-real combination display; manage the storage and retrieval of multiple pre-planned working angles and their optimized images; and send the final image to the angiography system.

[0163] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0164] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0165] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0166] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0167] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0168] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0169] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0170] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0171] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A medical image optimization method based on selective blurring of blood vessel branches, characterized in that, Includes the following steps: S1. Acquire 3D angiography DICOM images and perform grayscale conversion, contrast normalization, and gamma correction; S2. The image is enhanced with a multi-scale Hessian matrix Frangi filtering algorithm, and the visibility of low-contrast blood vessels is further improved by combining contrast-limited adaptive histogram equalization. S3. Use a deep learning segmentation network based on an encoder-decoder structure to perform vascular tree structure segmentation on the image. The network includes a combination of extensive activation modules and residual void spatial pyramid modules, and is trained and optimized through a joint regularization strategy. S4. Identify and mark vascular branches that overlap with or interfere with the target vascular structure through vascular topology analysis and spatial location relationships; the target vascular structure includes the parent artery, aneurysm neck, aneurysm body, and related branches that need to be clearly displayed during interventional surgery; the process of identifying and marking interfering vascular branches in S4 includes: S4.

1. Extract the vessel centerline based on the vessel segmentation results, construct the topological connection graph of the vessel tree, and record the morphological and spatial information of the vessel segments; S4.

2. Calculate the three-dimensional spatial bounding box of the target region and identify all vascular branches that intersect with the bounding box as candidate interference branches; S4.

3. Utilize the spatiotemporal features learned by the bidirectional convolutional LSTM module in the segmentation network to analyze the continuity of candidate interference branches and verify whether they are real vascular structures. S4.

4. Mark the vascular branches that are identified as overlapping or interfering with the target area, and add vascular hierarchy, diameter and positional relationship attribute information; S5. Selectively blur the marked interfering vascular branches, and use an image editing method based on the Poisson equation combined with bidirectional convolutional LSTM to perform spatiotemporal feature fusion on the processed area to generate an optimized image that does not visually obstruct the view and maintains the continuity of the blood vessels; the process of blurring the marked branches in S5 includes: Based on the vascular topology connection map, the upstream and downstream connection points of the marked interfering branches are determined, and a smooth virtual vascular path conforming to the vascular physiological morphology is generated as a repair guide. The processing area is defined with the virtual path as the center. The texture and gradient information of healthy blood vessels are used as boundary conditions by the Poisson equation to achieve seamless visual fusion between the processing area and the surrounding blood vessels. The blurring effect is achieved by adjusting the transparency. The preliminary repair results and the context feature map of the segmentation network are input into a bidirectional convolutional LSTM network. The repair results are optimized through spatiotemporal sequence processing to ensure the continuity of blood vessels and the naturalness of the structure. Calculate the change in the centerline curvature and the uniformity of the vessel diameter after processing. If the indicators are abnormal, return to adjust the virtual path parameters for iterative optimization. S6. Preserve the spatial information and topological structure of the original image to generate an optimized three-dimensional angiography image; S7. Surgical safety assurance procedures, which include: S7.1, Combined Real and Virtual Display: The interfering vascular branches processed in step S5 are displayed in the optimized image in a semi-transparent or dashed form, so as to eliminate visual obstruction and indicate their anatomical location to the surgeon. S7.2 Multi-angle contingency planning: Based on the same original three-dimensional angiography image, multiple different working angles and corresponding optimized images are pre-planned and generated for a single interventional surgical operation. Each optimized image at each angle independently executes steps S1 to S6 to blur the interfering branches that obscure the target vascular structure at that angle. S8. Output the optimized 3D angiography image in DICOM format and integrate it into an angiography system that supports real-time synchronization. Use it as a roadmap for endovascular interventional surgery to help the operator find the best working angle and safely navigate the interventional device.

2. The medical image optimization method based on selective vascular branch blurring according to claim 1, characterized in that, S1 includes: The original DICOM format angiography image is preprocessed; the image is a grayscale image. The grayscale image is contrast-normalized to adjust its pixel value distribution to a standard normal distribution with a mean of 0 and a standard deviation of 1. An adaptive histogram equalization method with limited contrast is used to locally enhance the standardized image. This involves dividing the image into multiple local regions and performing histogram cropping and equalization in each region to enhance the vascular structure and suppress noise. Adaptive gamma correction is applied to the enhanced image, and the gamma value is dynamically set according to the global contrast statistics of the image to optimize the overall brightness and detail of the image. The image is normalized by linearly scaling its pixel values ​​to the [0,1] range based on the effective grayscale range to adapt to the training and inference of the subsequent deep learning model.

3. The medical image optimization method based on selective vascular branch blurring according to claim 1, characterized in that, S2 includes: The Frangi filtering algorithm based on multi-scale Hessian matrix is ​​used to enhance blood vessels in images. By calculating the eigenvalues ​​of the Hessian matrix of pixels at different scales, a blood vessel similarity function is constructed to enhance tubular structures and suppress noise. The Frangi filtering result is weighted and fused with the original image after contrast-limited adaptive histogram equalization. The Frangi filtering result has a dominant weight and is used to highlight the main vascular structure, while the CLAHE processing result is used to supplement the details of small blood vessels and maintain the background texture. The fused enhanced image is used as input to a subsequent deep learning segmentation network for further optimization and confirmation of vascular features.

4. The medical image optimization method based on selective vascular branch blurring according to claim 3, characterized in that, S2 further includes: S2.

1. Perform Frangi filtering on a predefined set of scales, where the scale parameters cover the range of vessel diameters in the target vessel database; for each pixel, calculate the eigenvalues ​​of the Hessian matrix at each scale, and calculate the vessel similarity measure at that scale based on the eigenvalue relationships; finally, take the maximum response value at all scales as the Frangi filter response for that pixel. S2.

2. Adaptively weightedly fuse the Frangi filtering result and the CLAHE enhancement result. The fusion weight is dynamically adjusted according to the contrast characteristics of the local region of the image: in the high contrast region, the weight approaches 1, mainly relying on the Frangi filtering result; in the low contrast region, the weight is appropriately reduced to introduce more detailed information from the CLAHE processing. S2.

3. Input the final enhanced image into the subsequent deep learning segmentation network.

5. A medical image optimization method based on selective vascular branch blurring according to claim 4, characterized in that, The deep learning segmentation network based on the encoder-decoder structure in S3 includes the following components: Feature extraction part: It consists of multiple cascaded width-optimized residual units. Each unit increases the number of feature map channels before nonlinear activation by a channel expansion factor, while compressing the number of channels in the shortcut connection path. Each unit is followed by a max pooling layer for spatial downsampling. Multi-scale information aggregation section: includes a multi-branch dilated convolution fusion module, which uses dilated convolutional layers with different dilation rates to process features in parallel and additively fuses the output features of each pathway; Feature reconstruction part: The spatial size of the feature map is gradually restored through upsampling operation, and high-level semantic information and low-level positional details are fused by cross-layer connections; The network training employs a joint regularization strategy, simultaneously implementing batch normalization and weight parameter renormalization after each convolutional layer. The weight parameter renormalization decomposes the convolutional kernel weights into direction vectors and magnitude scalars for independent learning.

6. A medical image optimization method based on selective vascular branch blurring according to claim 5, characterized in that, The specific implementation of the deep learning segmentation network further includes: S3.

1. The forward propagation process of the width-optimized residual unit is expressed as follows: ， in The input to the module is σ, where σ is the ReLU activation function. and These represent the operations of the two convolutional layers before and after the expansion path, and the number of channels in the compressed shortcut connection path. Number of channels compared to standard residual blocks satisfy , For channel expansion factor, This is the output of the module; S3.

2. Output of the multi-branch dilated convolution fusion module It is obtained by summing the outputs of each parallel branch: ， Among them, Indicates the expansion rate 3×3 dilated convolution operation, The values ​​are 1, 2, 4, and 8. This indicates a re-standardization operation on the weight parameters. For activation functions; S3.

3. The weight parameter re-standardization operation The weight tensor of the convolutional layer Decomposed into direction vectors and amplitude scalar To learn: ； During training, stochastic gradient descent is used to calculate... and Calculate and update the gradient: ， ， in, For loss function, For the projection matrix, They are respectively , , The gradient; S3.

4. Composite loss function used in network training By binary cross-entropy loss Compared to loss based on similarity coefficient Linear combination constitutes: , Where λ is the balancing hyperparameter and the similarity loss term. Defined as: ， here, For real labels, K is a smoothing constant used to predict probabilities.

7. A medical image optimization method based on selective vascular branch blurring according to claim 1, characterized in that, The process of generating the optimized three-dimensional angiography image in S6 includes: The two-dimensional sequence segmentation results after branch blurring are reconstructed in three dimensions. The initial vascular topology connection map is used as a constraint to perform consistency checks and smoothing on possible topological breakpoints. The optimized blood vessel segmentation result is used as a foreground mask and fused with the background voxels of the original 3D DICOM image to fully preserve the spatial coordinates, gray values, and all positioning information in the DICOM file header of the original image. Automatically generate a structured report containing the number, location, and original diameter information of the blurred vascular branches, as well as the key morphological parameters of the optimized vascular tree; The fused 3D image data and its structured report are packaged into a standard DICOM format file.

8. A medical image optimization system based on selective vascular branch blurring, applied to the medical image optimization method based on selective vascular branch blurring as described in any one of claims 1 to 7, characterized in that, The system includes: The image preprocessing module is used to acquire 3D angiography DICOM images and perform grayscale conversion, contrast normalization, and gamma correction. The vessel enhancement module is used to enhance the blood vessels in the image using a multi-scale Hessian matrix Frangi filtering algorithm, and further improves the visibility of low-contrast blood vessels by combining contrast-limited adaptive histogram equalization. The blood vessel segmentation module is used to segment the image into a blood vessel tree structure using a deep learning segmentation network based on an encoder-decoder structure. The network includes a combination of a broad activation module and a residual void spatial pyramid module, and is trained and optimized through a joint regularization strategy. The interference branch identification module is used to identify and mark vascular branches that overlap with or interfere with the target vascular structure through vascular topology analysis and spatial location relationships; the process of identifying and marking interfering vascular branches includes: S4.

1. Extract the vessel centerline based on the vessel segmentation results, construct the topological connection graph of the vessel tree, and record the morphological and spatial information of the vessel segments; S4.

2. Calculate the three-dimensional spatial bounding box of the target region and identify all vascular branches that intersect with the bounding box as candidate interference branches; S4.

3. Utilize the spatiotemporal features learned by the bidirectional convolutional LSTM module in the segmentation network to analyze the continuity of candidate interference branches and verify whether they are real vascular structures. S4.

4. Mark the vascular branches that are identified as overlapping or interfering with the target area, and add vascular hierarchy, diameter and positional relationship attribute information; The branch blurring module is used to selectively blur marked interfering vascular branches using an image editing method based on the Poisson equation, and combines bidirectional convolutional LSTM to perform spatiotemporal feature fusion on the processed area to generate an optimized image that does not visually obstruct the view and maintains the continuity of the blood vessels. The process of blurring the marked branches includes: Based on the vascular topology connection map, the upstream and downstream connection points of the marked interfering branches are determined, and a smooth virtual vascular path conforming to the vascular physiological morphology is generated as a repair guide. The processing area is defined with the virtual path as the center. The texture and gradient information of healthy blood vessels are used as boundary conditions by the Poisson equation to achieve seamless visual fusion between the processing area and the surrounding blood vessels. The blurring effect is achieved by adjusting the transparency. The preliminary repair results and the context feature map of the segmentation network are input into a bidirectional convolutional LSTM network. The repair results are optimized through spatiotemporal sequence processing to ensure the continuity of blood vessels and the naturalness of the structure. Calculate the change in the centerline curvature and the uniformity of the vessel diameter after processing. If the indicators are abnormal, return to adjust the virtual path parameters for iterative optimization. The image output module is used to preserve the spatial information and topological structure of the original image, generate optimized three-dimensional angiography images, and output them in DICOM format. The surgical safety control module controls the branch blurring processing module to display interfering branches as semi-transparent or dashed lines; it manages the generation and retrieval of multiple pre-planned working angles and their corresponding optimized images based on the same original image; and it integrates the optimized images into the angiography system for real-time navigation of endovascular interventional surgery.

Images