Blood vessel wall image automatic analysis and lesion recognition system and method
By using an automated vascular wall image analysis system combined with multimodal image fusion technology, the problems of assessment discrepancies and time consumption in traditional vascular lesion identification have been solved, achieving high-precision and rapid lesion identification and assessment, and supporting real-time adjustment of interventional surgical pathways.
Patent Information
- Application Number
- CN202510885508.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional vascular lesion identification relies on doctors' subjective judgment, which results in large differences in assessment, long time consumption, and failure to effectively integrate VUS and DSA imaging information, making it difficult to meet the needs of emergency care.
It employs image acquisition, preprocessing, feature extraction, fusion computation, semantic segmentation, and lesion classification units, combined with sparse core association algorithm, dilated convolutional pyramid network, U-Net network, and support vector machine to achieve automated blood vessel wall image analysis, supporting multimodal image fusion and real-time feedback.
It has achieved reduced vascular boundary segmentation error, improved plaque volume calculation accuracy, shortened full-process automation time, improved catheter positioning accuracy, and enhanced consistency in stenosis assessment, especially improving the diagnostic accuracy of bifurcation lesions.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure SMS_5
Abstract
Description
Technical Field
[0001] This invention relates to the field of image analysis technology, specifically to an automatic analysis system and method for vascular wall images and lesion recognition. Background Technology
[0002] Cardiovascular and cerebrovascular diseases are among the leading causes of death worldwide, and their diagnosis heavily relies on the accurate analysis of vascular images. Traditional methods of identifying vascular lesions primarily rely on doctors' subjective judgment of DSA or IVUS images, which have the following drawbacks:
[0003] Different doctors can have 20% to 30% differences in their assessment of the same lesion. This is especially true in complex bifurcation lesions. Manually measuring vessel diameter and plaque volume can lead to significant errors. Existing semi-automated systems require manual annotation of key points, and a single analysis takes more than 15 minutes, which is insufficient to meet emergency needs. The anatomical and functional information from VUS and DSA images has not been effectively integrated, resulting in an incomplete assessment of the degree of stenosis. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an automatic blood vessel wall image analysis and lesion identification system and method. It solves the problems of errors in manually measuring blood vessel diameter and plaque volume in complex bifurcation lesions, the need for manual annotation of key points in semi-automatic systems, the time taken for a single analysis exceeding 15 minutes, which is difficult to meet emergency needs, and the failure to effectively integrate data, resulting in incomplete assessment of stenosis degree.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic blood vessel wall image analysis and lesion identification system, comprising:
[0006] The image acquisition unit is used to acquire real-time images of the target blood vessel through an intravascular ultrasound (IVUS) catheter or a digital subtraction angiography (DSA) device.
[0007] The preprocessing unit is used to perform noise reduction and contrast enhancement processing on the real-time image to generate a preprocessed image;
[0008] The feature extraction unit is used to perform multi-scale feature extraction on the preprocessed image to obtain shallow feature maps and deep feature maps;
[0009] The fusion computing unit is used to fuse the shallow feature map and the deep feature map using a sparse core association algorithm to generate a fused feature map;
[0010] The semantic segmentation unit is used to segment the blood vessel boundary and plaque region based on the fused feature map, and extract medical feature parameters, including plaque volume, minimum lumen diameter and blood vessel area.
[0011] The lesion classification unit is used to output lesion level labels based on the medical feature parameters and a preset lesion grading model.
[0012] The fusion computing unit satisfies the following formula:
[0013]
[0014] Among them, F 浅层 and F 深层 These are the shallow and deep feature maps, respectively, and M is the correlation weight matrix.
[0015] Preferably, the preprocessing unit includes:
[0016] The noise reduction module uses a nonlocal mean filtering algorithm to suppress noise in the image, and the filtering parameters satisfy the following:
[0017]
[0018] Where h is the filter intensity, σ is the noise standard deviation, and N is the total number of image pixels.
[0019] Preferably, the feature extraction unit employs a dilated convolutional pyramid network, comprising three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively, for extracting multi-scale vascular structure features.
[0020] Preferably, the fusion computing unit further includes:
[0021] The weight allocation module is used to perform position-by-position weighting of shallow and deep features according to the associated weight moments M, generating an optimized fused feature map F. 融合 ,satisfy:
[0022]
[0023] Here, ⊙ represents element-wise multiplication. This represents matrix multiplication.
[0024] Preferably, the semantic segmentation unit adopts a U-Net network structure and introduces an attention mechanism in the output layer, wherein the calculation of the attention weight α satisfies:
[0025] a = Sigmoid(W·F) 融合 +b)
[0026] Where W and b are trainable parameters.
[0027] Preferably, the lesion classification unit includes:
[0028] The feature encoding module maps the medical feature parameters into high-dimensional vectors;
[0029] The hierarchical module uses a Support Vector Machine (SVM) model, and its decision function is:
[0030]
[0031] Where K(x) i ,x) is the radial basis kernel function.
[0032] Preferred options also include:
[0033] The real-time feedback unit is used to dynamically adjust the catheter advancement path according to the lesion grade label during interventional surgery, and the path correction amount Δd satisfies:
[0034]
[0035] Where k is the proportionality coefficient, S 斑块 and S 血管 These are the cross-sectional areas of the plaque and the blood vessel, respectively.
[0036] The teaching review module is used to record user annotation operations and system analysis results, and generate interactive 3D blood vessel models with a resolution of no less than 0.1 mm / pixel.
[0037] Preferably, the medical feature parameters also include plaque type, which is classified by analyzing the gray-scale distribution entropy HH of the plaque region:
[0038]
[0039] Where, p i Let be the probability of the i-th gray level. The correspondence between the entropy range and the patch type is as follows:
[0040] H < 2.0: Calcified plaque;
[0041] 2.0≤H≤3.5: fibrous plaque;
[0042] H>3.5: Lipid plaque.
[0043] Preferably, the system supports multimodal image fusion, and when acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is adopted:
[0044] I 融合 =λ·I IVUS +(1-λ)·I DSA
[0045] Where λ∈[0.4,0.6] represents the dynamically adjusted weights.
[0046] An automated method for analyzing blood vessel wall images and identifying lesions includes the following steps:
[0047] Step 1: Acquire real-time images of the target blood vessel using an intravascular ultrasound (IVUS) catheter or digital subtraction angiography (DSA) equipment, and perform noise reduction and contrast enhancement processing on the images to generate pre-processed images;
[0048] Step 2: Perform multi-scale feature extraction on the preprocessed image. Extract shallow and deep feature maps using a dilated convolutional pyramid network. The network contains three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively. Then, fuse the shallow and deep feature maps using a sparse core association algorithm to generate a fused feature map.
[0049] Step 3: Based on the fused feature map, the U-Net network with an attention mechanism is used to segment the blood vessel boundary and plaque region, and medical feature parameters are extracted, including plaque volume, minimum lumen diameter, blood vessel area and plaque type.
[0050] Step 4: Map the medical feature parameters into high-dimensional vectors, output lesion level labels through a support vector machine (SVM) model, dynamically adjust the catheter advancement path according to the lesion level labels during the interventional procedure, and generate an interactive three-dimensional vascular model with a resolution of not less than 0.1 mm / pixel.
[0051] Step 5: When acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is used to generate a fused image.
[0052] This invention provides an automated system and method for analyzing blood vessel wall images and identifying lesions. It offers the following advantages:
[0053] 1. This invention integrates multi-scale features through a sparse core association algorithm, reducing the vascular boundary segmentation error from ±0.5mm in traditional methods to ±0.1mm, and the relative error of plaque volume calculation is <5%. At the same time, the three-dimensional reconstructive model can restore the entire surgical process, supports reanalysis of arbitrary sections, and shortens the training cycle by 50%.
[0054] 2. The fully automated processing time of this invention is ≤3 minutes, supports real-time path correction of interventional devices during the procedure, and the catheter positioning accuracy reaches 0.05mm. Furthermore, weighted fusion improves the consistency of stenosis rate assessment to 95%, especially improving the diagnostic accuracy of bifurcation lesions by 40%. Detailed Implementation
[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0056] As one aspect of the present invention, the present invention provides an automatic blood vessel wall image analysis and lesion identification system, comprising:
[0057] The image acquisition unit is used to acquire real-time images of the target blood vessel through an intravascular ultrasound (IVUS) catheter or a digital subtraction angiography (DSA) device.
[0058] The preprocessing unit is used to perform noise reduction and contrast enhancement processing on the real-time image to generate a preprocessed image, including:
[0059] The noise reduction module uses a nonlocal mean filtering algorithm to suppress noise in the image, and the filtering parameters satisfy the following: Where h is the filter intensity, σ is the noise standard deviation, and N is the total number of image pixels;
[0060] The feature extraction unit employs a dilated convolutional pyramid network, which contains three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively. This unit is used to extract multi-scale vascular structure features and to perform multi-scale feature extraction on the preprocessed image, resulting in shallow and deep feature maps.
[0061] The fusion computing unit is used to fuse the shallow feature map and the deep feature map using a sparse core association algorithm to generate a fused feature map, including:
[0062] The weight allocation module is used to perform position-by-position weighting of shallow and deep features according to the associated weight moments M, generating an optimized fused feature map F. 融合 ,satisfy: Here, ⊙ represents element-wise multiplication. Represents matrix multiplication;
[0063] A U-Net network structure semantic segmentation unit is used to segment blood vessel boundaries and plaque regions based on the fused feature map, and to extract medical feature parameters, including plaque volume, minimum lumen diameter, and blood vessel area. An attention mechanism is introduced in the output layer, and the attention weight α is calculated according to: a = Sigmoid(W·F) 融合 +b) Where W and b are trainable parameters, and the medical feature parameters are classified by analyzing the gray-level distribution entropy HH of the plaque region:
[0064]
[0065] Where, p i Let be the probability of the i-th gray level. The correspondence between the entropy range and the patch type is as follows:
[0066] H < 2.0: Calcified plaque;
[0067] 2.0≤H≤3.5: fibrous plaque;
[0068] H>3.5: Lipid plaque;
[0069] A lesion classification unit, used to output lesion level labels based on the medical feature parameters and a preset lesion grading model, including:
[0070] The feature encoding module maps the medical feature parameters into high-dimensional vectors;
[0071] The hierarchical module uses a Support Vector Machine (SVM) model, and its decision function is:
[0072]
[0073] Where K(x) i (x) is the radial basis kernel function;
[0074] The fusion computing unit satisfies the following formula:
[0075]
[0076] Among them, F 浅层 and F 深层 These are the shallow and deep feature maps, respectively, and M is the correlation weight matrix.
[0077] The real-time feedback unit is used to dynamically adjust the catheter advancement path according to the lesion grade label during interventional surgery, and the path correction amount Δd satisfies:
[0078]
[0079] Where k is the proportionality coefficient, S 斑块 and S 血管 These are the cross-sectional areas of the plaque and the blood vessel, respectively.
[0080] The teaching review module is used to record user annotation operations and system analysis results, and generate interactive 3D blood vessel models with a resolution of no less than 0.1 mm / pixel.
[0081] The system supports multimodal image fusion. When acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is used.
[0082] I 融合 =λ·I IVUS +(1-λ)·I DSA
[0083] Where λ∈[0.4,0.6] represents the dynamically adjusted weights.
[0084] In another aspect, the present invention provides a method for automatic analysis of blood vessel wall images and lesion identification, comprising the following steps:
[0085] Step 1: Acquire real-time images of the target blood vessel using an intravascular ultrasound (IVUS) catheter or digital subtraction angiography (DSA) equipment, and perform noise reduction and contrast enhancement processing on the images to generate pre-processed images;
[0086] Step 2: Perform multi-scale feature extraction on the preprocessed image. Extract shallow and deep feature maps using a dilated convolutional pyramid network. The network contains three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively. Then, fuse the shallow and deep feature maps using a sparse core association algorithm to generate a fused feature map.
[0087] Step 3: Based on the fused feature map, the U-Net network with an attention mechanism is used to segment the blood vessel boundary and plaque region, and medical feature parameters are extracted, including plaque volume, minimum lumen diameter, blood vessel area and plaque type.
[0088] Step 4: Map the medical feature parameters into high-dimensional vectors, output lesion level labels through a support vector machine (SVM) model, dynamically adjust the catheter advancement path according to the lesion level labels during the interventional procedure, and generate an interactive three-dimensional vascular model with a resolution of not less than 0.1 mm / pixel.
[0089] Step 5: When acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is used to generate a fused image.
[0090] The following description, in conjunction with specific embodiments, will be provided.
[0091] Example:
[0092] In the cardiovascular interventional center of a tertiary-level hospital, vascular imaging analysis and lesion identification were performed on 50 patients with coronary atherosclerosis to verify the clinical efficacy of the system of this invention. The patients' ages ranged from 45 to 75 years, and all underwent dual-modal imaging examinations including IVUS and DSA.
[0093] System configuration and parameter settings:
[0094] Image acquisition unit: using Philips IVUS catheter (40MHz frequency) and Siemens Artis Q DSA device (1024×1024 resolution); Preprocessing unit:
[0095] Noise reduction module: Non-local mean filtering, noise standard deviation σ = 0.05, total number of pixels N = 1024 × 1024, the calculated filter strength.
[0096] Contrast enhancement: Histogram equalization combined with gamma correction (γ = 1.2). Feature extraction unit: Dilated convolutional pyramid network with dilation rates of 1, 2, and 4, kernel size of 3×3, and stride of 1.
[0097] Fusion computing unit: Sparse core association algorithm, the association weight matrix M is generated by the Softmax function.
[0098] Semantic segmentation unit: U-Net network, attention mechanism parameters W and b are trained using Adam optimizer (learning rate 0.001).
[0099] Lesion classification unit: SVM model, radial basis function kernel width γ = 0.1, classification threshold is set based on clinical consensus.
[0100] Multimodal fusion: dynamic weight λ = 0.5, automatically adjusted according to the curvature of blood vessels (λ increases to 0.6 when the curvature is >30°).
[0101] Operating procedures
[0102] Step 1: Image Acquisition and Preprocessing
[0103] Cross-sectional images of blood vessels were acquired via IVUS catheter (30fps), and angiography sequences were acquired using DSA equipment (contrast agent injection rate 3mL / s).
[0104] The image is denoised and contrast enhanced to generate a pre-processed image.
[0105] Step 2: Multi-scale feature extraction and fusion
[0106] The dilated convolutional pyramid network extracts shallow features (edges, textures) and deep features (plaque calcification areas, vascular branching structures).
[0107] The sparse core association algorithm generates a fused feature map, using the following formula: in
[0108] Step 3: Semantic Segmentation and Feature Extraction
[0109] Net network segmentation of blood vessel boundaries and plaque regions, with a segmentation error of ±0.1mm (compared to ±0.5mm for traditional methods).
[0110] Extract medical feature parameters: plaque volume (relative error <3%), minimum lumen diameter (error ±0.05mm), and gray entropy H.
[0111] Step 4: Lesion Classification and Real-time Feedback
[0112] The SVM model outputs lesion grade labels (stable plaques, high-risk plaques) with a classification accuracy of 95% (compared to 80% for traditional methods).
[0113] The real-time feedback unit dynamically adjusts the conduit path and corrects the amount of correction. The intraoperative catheter positioning accuracy reached 0.05mm.
[0114] Step 5: Multimodal fusion and 3D debriefing
[0115] IVUS-DSA fused images were generated, with a 95% consistency in stenosis assessment.
[0116] The teaching review module generates a 3D blood vessel model (resolution 0.1mm / pixel) and supports resampling of any cross section.
[0117] Experimental Results and Comparison:
[0118]
[0119]
[0120] This embodiment verifies the significant advantages of the system of the present invention:
[0121] Accuracy: Multi-scale feature fusion and attention mechanism significantly reduce errors in vessel segmentation and plaque volume calculation.
[0122] Real-time performance: The fully automated processing time is only 3 minutes, which is significantly better than traditional semi-automated systems.
[0123] Safety: Dynamic catheter path correction combined with a three-dimensional replay model reduces the risk of intraoperative vascular injury.
[0124] Multimodal collaboration: The fusion of IVUS and DSA images improves the consistency of stenosis assessment, especially increasing the diagnostic accuracy of bifurcation lesions by 40%.
[0125] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A system for automatic analysis of blood vessel wall images and identification of lesions, characterized in that, include: The image acquisition unit is used to acquire real-time images of the target blood vessel through an intravascular ultrasound (IVUS) catheter or a digital subtraction angiography (DSA) device. The preprocessing unit is used to perform noise reduction and contrast enhancement processing on the real-time image to generate a preprocessed image; The feature extraction unit is used to perform multi-scale feature extraction on the preprocessed image to obtain shallow feature maps and deep feature maps; The fusion computing unit is used to fuse the shallow feature map and the deep feature map using a sparse core association algorithm to generate a fused feature map; The semantic segmentation unit is used to segment the blood vessel boundary and plaque region based on the fused feature map, and extract medical feature parameters, including plaque volume, minimum lumen diameter and blood vessel area. The lesion classification unit is used to output lesion level labels based on the medical feature parameters and a preset lesion grading model. The fusion computing unit satisfies the following formula: Among them, F 浅层 and F 深层 These are the shallow and deep feature maps, respectively, and M is the correlation weight matrix.
2. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The preprocessing unit includes: The noise reduction module uses a nonlocal mean filtering algorithm to suppress noise in the image, and the filtering parameters satisfy the following: Where h is the filter intensity, σ is the noise standard deviation, and N is the total number of image pixels.
3. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The feature extraction unit employs a dilated convolutional pyramid network, which contains three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively, to extract multi-scale vascular structure features.
4. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The fusion computing unit also includes: The weight allocation module is used to perform position-by-position weighting of shallow and deep features according to the associated weight moments M, generating an optimized fused feature map F. 融合 ,satisfy: Here, ⊙ represents element-wise multiplication. This represents matrix multiplication.
5. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The semantic segmentation unit adopts a U-Net network structure and introduces an attention mechanism in the output layer. The calculation of the attention weight α satisfies: a=Sigmoid(W·F 融合 +b) Where W and b are trainable parameters.
6. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The lesion classification unit includes: The feature encoding module maps the medical feature parameters into high-dimensional vectors; The hierarchical module uses a Support Vector Machine (SVM) model, and its decision function is: Where K(x) i ,x) is the radial basis kernel function.
7. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, Also includes: The real-time feedback unit is used to dynamically adjust the catheter advancement path according to the lesion grade label during interventional surgery, and the path correction amount Δd satisfies: Where k is the proportionality coefficient, S 斑块 and S 血管 These are the cross-sectional areas of the plaque and the blood vessel, respectively. The teaching review module is used to record user annotation operations and system analysis results, and generate interactive 3D blood vessel models with a resolution of no less than 0.1 mm / pixel.
8. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The medical feature parameters also include plaque type, which is classified by analyzing the gray-scale distribution entropy HH of the plaque region: Where, p i Let be the probability of the i-th gray level. The correspondence between the entropy range and the patch type is as follows: H < 2.0: Calcified plaque; 2.0≤H≤3.5: fibrous plaque; H>3.5: Lipid plaque.
9. The automatic blood vessel wall image analysis and lesion identification system according to claim 1, characterized in that, The system supports multimodal image fusion. When acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is used. I 融合 =λ·I IVUS +(1-λ)·I DSA Where λ∈[0.4,0.6] represents the dynamically adjusted weights.
10. A method for automatic analysis of blood vessel wall images and identification of lesions, using a system for automatic analysis of blood vessel wall images and identification of lesions as described in any one of claims 1-9, characterized in that, The following steps are involved: Step 1: Acquire real-time images of the target blood vessel using an intravascular ultrasound (IVUS) catheter or digital subtraction angiography (DSA) equipment, and perform noise reduction and contrast enhancement processing on the images to generate pre-processed images; Step 2: Perform multi-scale feature extraction on the preprocessed image. Extract shallow and deep feature maps through a dilated convolutional pyramid network. The network contains three parallel convolutional layers with dilation rates of 1, 2, and 4, respectively. Subsequently, the shallow and deep feature maps are fused using the sparse core association algorithm to generate a fused feature map; Step 3: Based on the fused feature map, the U-Net network with an attention mechanism is used to segment the blood vessel boundary and plaque region, and medical feature parameters are extracted, including plaque volume, minimum lumen diameter, blood vessel area and plaque type. Step 4: Map the medical feature parameters into high-dimensional vectors, output lesion level labels through a support vector machine (SVM) model, dynamically adjust the catheter advancement path according to the lesion level labels during the interventional procedure, and generate an interactive three-dimensional vascular model with a resolution of not less than 0.1 mm / pixel. Step 5: When acquiring IVUS and DSA images simultaneously, a weighted fusion strategy is used to generate a fused image.