Scale-aware progressive learning method for x-ray coronary angiography stenosis segmentation

Abstract

The application discloses a scale perception progressive learning method for X-ray coronary angiography stenosis segmentation. The technical problem to be solved is how to solve the problem of insufficient segmentation accuracy caused by multi-scale characteristics and small target characteristics in coronary angiography stenosis segmentation. The application combines a scale perception strategy and a progressive feature reservation mechanism to provide a scale perception progressive learning method for X-ray coronary angiography stenosis segmentation. The application significantly improves the segmentation capability of multi-scale stenosis. The scale perception strategy dynamically aggregates context features of different scales, effectively adapts to the scale difference of coronary arteries in XCA images, enhances the sensitivity of the model to subtle stenosis changes, realizes more accurate stenosis boundary delineation, effectively solves the problem of easy loss of small target information, improves the robustness of the model to complex background interference, has stronger generalization capability, can adapt to different devices and imaging conditions, reduces the work burden of doctors and improves efficiency.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation and provides a scale-aware progressive learning method for segmenting stenosis in X-ray coronary angiography. Background Technology

[0002] Coronary artery disease is a leading cause of death worldwide. Its main pathological feature is narrowing or blockage of the coronary arteries due to plaque buildup, which affects the blood supply to the heart muscle. This narrowing can cause not only angina but also myocardial infarction, and in severe cases, even death. In clinical treatment, X-ray coronary angiography (XCA) is a key tool for assessing the degree of coronary artery stenosis. Through the visualization information provided by XCA, doctors can accurately analyze the morphology of the coronary arteries, thereby developing personalized treatment plans. Percutaneous coronary intervention (PCI) is a common interventional cardiovascular procedure widely used to treat coronary artery disease. PCI restores blood flow by implanting a stent in the narrowed or blocked coronary artery, thereby relieving angina, preventing myocardial infarction, and improving the patient's quality of life. During PCI, doctors rely on the precise segmentation results provided by XCA to determine the severity, length, and location of the stenosis, and then select the appropriate stent size and optimize the surgical strategy. Therefore, accurate segmentation of the stenotic area of ​​the coronary artery in XCA images is crucial to ensuring the success rate of the procedure and the patient's prognosis.

[0003] Recent research on coronary angiography stenosis segmentation has primarily relied on supervised deep learning models, which have improved segmentation accuracy to some extent. However, these methods still have significant limitations when dealing with the multi-scale and small-target characteristics of coronary artery stenosis. The multi-scale nature makes it difficult for existing methods to accurately delineate stenosis boundaries in complex backgrounds, while the small-target nature makes these methods prone to losing crucial information when dealing with small lesions. These problems not only reduce the reliability of segmentation results but may also lead to misjudgments in surgical strategies during clinical practice, increasing the treatment risks for patients.

[0004] To address the aforementioned problems, this invention attempts to solve them through a scale-aware approach, leveraging its ability to dynamically perceive features at different scales to improve segmentation accuracy. However, different coronary angiography images exhibit significant scale variations, and existing methods cannot effectively preserve truly valuable features, especially in small target feature preservation and multi-scale detail capture. To address this issue, this invention proposes a coronary angiography stenosis segmentation method based on Scale-Aware Progressive Learning (SAPL). This method improves segmentation accuracy from both multi-scale and small target dimensions by introducing a scale-aware strategy and a progressive feature preservation mechanism. Specifically, the scale-aware strategy dynamically perceives features at different scales and suppresses background interference, while the progressive feature preservation mechanism gradually preserves key features of small targets, significantly reducing information loss and thus achieving accurate segmentation of coronary artery stenosis areas. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to solve the problem of insufficient segmentation accuracy caused by the multi-scale characteristics and small target characteristics in coronary angiography stenosis segmentation. The invention proposes a scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation that combines a scale-aware strategy and a progressive feature preservation mechanism.

[0006] The present invention solves the above-mentioned key problems by adopting the following technical solutions:

[0007] This invention discloses a scale-aware progressive learning method for stenosis segmentation in X-ray coronary angiography, comprising:

[0008] 1) Obtain X-ray coronary angiography images and corresponding ground truth masks from publicly available datasets;

[0009] 2) The obtained X-ray coronary angiography images are fed into the scale-aware convolutional aggregation module SCA to extract local consistency feature maps, stenosis boundary feature maps, and multi-scale context feature maps;

[0010] 3) Dynamically fuse the three-branch feature maps using 1×1 convolution to obtain a unified multi-scale feature map;

[0011] 4) Input the unified multi-scale feature map into the progressive feature preserving block (PFP), and obtain the enhanced feature map with narrow information at small scale through saliency selection, position encoding and channel attention;

[0012] 5) The enhanced feature map of the obtained small-scale narrow information and the unified multi-scale feature map are fused using a residual connection method to obtain an enhanced small target feature map;

[0013] 6) The enhanced small target feature map and the X-ray coronary angiography image are fed into the segmentation network to obtain the prediction mask;

[0014] 7) Construct a scale-aware loss function (SAL) and calculate weight coefficients based on the area difference between the predicted mask and the real mask. ;

[0015] 8) Adopt The weighted crossover ratio loss optimization network makes the model pay more attention to small-scale narrow regions during training, ultimately resulting in a segmentation model that can accurately segment coronary artery stenosis of different scales.

[0016] As a further improvement, the scale-aware convolutional aggregation module SCA of the present invention includes:

[0017] Standard convolutional branches are used to maintain local spatial consistency;

[0018] The central difference convolution branch is used to highlight gradient details at narrow boundaries, and its operation is defined as follows: ; This represents central difference convolution; Represents the standard convolutional response; This represents the central difference intensity adjustment coefficient, with a value range of 0-1; Indicates the central difference term;

[0019] The dilated convolution branch contains multiple convolution operations with different dilation rates to expand the receptive field and integrate contextual information at different scales.

[0020] As a further improvement, the present invention describes inputting a unified multi-scale feature map into a progressive feature preserving block (PFP) and obtaining an enhanced feature map with narrow information at small scales through saliency selection. Specifically, this includes: selecting the most saliency feature map from each channel of the input feature tensor. There are 1 feature points; among which The value is dynamically adjusted based on the depth of the network layers, and the calculation formula is: in, The number of channels in the input feature map. This is the preset offset. This represents the current network layer number.

[0021] As a further improvement, the present invention inputs a unified multi-scale feature map into a progressive feature preserving block (PFP), and obtains an enhanced feature map with narrow information at a small scale through position encoding, so that it has spatial position awareness capability, and obtains a position-aware feature matrix, which is used to enhance the correlation between local and global information.

[0022] As a further improvement, the present invention inputs a unified multi-scale feature map into a progressive feature preserving block (PFP), obtains an enhanced feature map of small-scale narrow information through channel attention, highlights important features of small-scale narrow regions through weighting, and suppresses irrelevant background features, thereby obtaining an enhanced feature map.

[0023] As a further improvement, the scale-aware loss function (SAL) described in this invention calculates weight coefficients based on the area difference between the predicted mask and the real mask. The calculation method is as follows:

[0024] in, To predict the pixel set, For real pixel set, This is a function for calculating the scalar variance.

[0025] As a further improvement, the present invention employs... The weighted intersection-union loss optimization network has the following form:

[0026] Used to enhance the model's focus on small-scale targets during training.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. Significantly improves the ability to segment narrow structures at multiple scales.

[0029] By dynamically aggregating contextual features at different scales using a scale-aware strategy, the model can effectively adapt to the scale differences presented by coronary arteries in XCA images, enhance the model's sensitivity to subtle stenosis changes, and achieve more accurate characterization of stenosis boundaries.

[0030] 2. Effectively solves the problem of easy loss of information in small targets.

[0031] A progressive feature preservation mechanism is adopted to gradually screen, locate and enhance key features in small-scale narrow regions, which significantly reduces the degradation of small target features in deep networks and improves the ability to identify small lesion regions.

[0032] 3. Improve the model's robustness to complex background interference.

[0033] Coronary angiography images are often accompanied by highly interfering factors such as contrast agent flow, rib shadows, and noise, which easily affect traditional methods. The scale-aware aggregation module in this invention can dynamically suppress invalid responses in the background region, allowing the network to focus on the main blood vessel and stenotic areas, thereby significantly improving the segmentation stability and robustness in complex backgrounds.

[0034] 4. It has stronger generalization ability and can adapt to different devices and imaging conditions.

[0035] Since different hospitals, imaging equipment, and shooting angles can cause significant differences in image scale and quality, this invention adopts a dynamic scale-aware strategy and a progressive learning mechanism to enable the model to automatically adapt to differences in image resolution, scale changes, and noise conditions, thereby improving the generalization performance across devices and data sources and making it more stable on multi-center data.

[0036] 5. Reduce doctors' workload and improve efficiency.

[0037] Traditional manual analysis of XCA images is time-consuming and relies on experience. This method can automatically generate high-quality segmentation results, significantly reducing the workload of doctors' manual annotation and repeated verification, improving the efficiency of clinical workflows, and enhancing overall segmentation accuracy and clinical reliability.

[0038] 6. Improve the safety and success rate of interventional PCI procedures.

[0039] Precise stenosis segmentation results provide accurate vascular structure and lesion contours for PCI surgery, enabling doctors to rationally select stent type and surgical path, reducing the risk of stent misimplantation, over-expansion, or undertreatment, thereby improving treatment outcomes and reducing intraoperative complications.

[0040] 7. Promote the standardized analysis of coronary angiography data.

[0041] Differences in angiography equipment, imaging angles, and contrast distribution among different hospitals lead to inconsistent analytical standards. This invention utilizes a scale-aware strategy to improve robustness across devices and scenarios, enabling the output of more consistent stenosis measurement results. This contributes to the construction of unified coronary artery disease assessment standards and provides a technical foundation for large-scale multi-center studies and database development.

[0042] 8. Reduce patients' overall medical costs and time consumption.

[0043] Automated intelligent segmentation can reduce the need for repeated angiography, shorten operation time, and lower the risk of secondary treatment due to misdiagnosis. Higher precision stenosis segmentation can improve treatment accuracy, reduce the burden on patients in terms of both economy and time, and has significant social benefits. Attached Figure Description

[0044] Figure 1 This is a diagram illustrating the overall architecture of a coronary angiography stenosis segmentation method based on progressive scale-aware SAPL. Detailed Implementation

[0045] The present invention will be further described in conjunction with the accompanying drawings and specific embodiments:

[0046] This invention addresses the aforementioned key problems using the following technical solution: a coronary angiography stenosis segmentation method based on progressive scale perception. This method dynamically perceives features at different scales and suppresses background interference through a scale perception strategy, and progressively preserves key features of small targets through a progressive feature preservation mechanism, thereby significantly reducing information loss and improving segmentation accuracy. The method specifically includes the following steps:

[0047] Step 1: Obtain X-ray coronary angiography images and corresponding ground truth masks from public datasets.

[0048] Step 2: The obtained X-ray coronary angiography images are fed into the Scale-Aware Convolutional Aggregation (SCA) module. The SCA module first maintains local spatial consistency through standard convolution operations, which effectively preserves key patterns at close range. Next, a central difference convolution operation is used to highlight details at narrow boundaries, crucial for distinguishing fine structures in XCA images. Furthermore, a dilated convolution operation expands the receptive field by using different dilation rates, thereby integrating contextual information at different scales.

[0049] Step 3: Finally, the outputs of the above three branches are dynamically fused through a 1×1 convolutional layer to generate a unified feature map containing multi-scale stenosis lesion information. The output of the standard convolutional branch is represented as:

[0050]

[0051] Where X is the input feature map, This is a standard convolution operation. The output of the center difference convolution branch is represented as:

[0052]

[0053] in, It is the central difference convolution operation, defined as:

[0054] ; This refers to center difference convolution. Center difference convolution guides the network to focus on local intensity changes rather than absolute intensity by subtracting the weighted response of the center pixel from the standard convolution response, thereby effectively enhancing narrow boundary and fine-grained structural information. Represents the standard convolutional response; This represents the central difference intensity adjustment coefficient, with a value range of 0-1; Indicates the central difference term;

[0055] The output of the dilated convolution branch is represented as:

[0056]

[0057] in, It is a dilated convolutional block, containing multiple convolutional operations with different dilation rates. The feature fusion process is represented as:

[0058]

[0059] in, This means concatenating the outputs of the three branches along the channel dimension. It is a 1×1 convolution operation.

[0060] Step 4: Existing methods have significant limitations in preserving features of small targets, especially in feature selection and location information preservation. Traditional methods often capture only limited feature information, making it difficult to accurately distinguish target regions in complex backgrounds. Therefore, this invention introduces a Progressive Feature Preservation Block (PFP) to progressively preserve key features of coronary artery stenosis. The PFP first selects the most salient feature from each channel of the input feature tensor. These points form a feature matrix. The value is dynamically determined based on the network layer depth, ensuring that deeper networks sample more feature points. The mathematical expression is:

[0061]

[0062] in, It is the input tensor. Represents the selected feature vector. This represents the corresponding position index. The smaller 's' indicates the number of points selected, which is a parameter variable. The larger 'S' represents the final number of feature points selected in the current layer, obtained through a later formula; it is a constant, set to s=S in this layer. Then, it is calculated using the following formula... :

[0063]

[0064] in, It is the number of channels in the input tensor. It is the preset offset. This refers to the actual number of layers in the network structure. Furthermore, spatial context is provided for selected feature points through positional encoding, generating a position-aware feature matrix. This is specifically achieved through the following steps:

[0065]

[0066]

[0067] in, It is a position embedding matrix and It contains the two-dimensional coordinate information of selected feature points. Position-aware feature matrix. The calculation is as follows:

[0068]

[0069] Step 5: Finally, the model's focus on narrow regions is further refined through a channel attention layer, while suppressing background noise. This is achieved through the following steps:

[0070]

[0071]

[0072] Enhanced feature map of small targets.

[0073] Step 6: Input the enhanced small target feature map and the X-ray coronary angiography image into the segmentation network to obtain the prediction mask.

[0074] Step 7: Existing loss functions (such as IoU loss) are not sensitive to the scale of narrow regions, making it difficult to effectively optimize segmentation results at different scales, resulting in insufficient model performance when handling small targets. To address this, a scale-aware loss function (SAL) is introduced, which dynamically adjusts weights to adapt to narrow regions at different scales, improving sensitivity to small narrow regions while maintaining stability for large narrow regions.

[0075] In practical applications, SAL first calculates the predicted pixel set. and real pixel set The size of the value. Then, the weight is calculated using the following formula. :

[0076]

[0077] in, It is a function that calculates the provided scalar variance. Weights This is used to adjust the sensitivity of the loss function to narrow regions at different scales. Specifically, when the size difference between the predicted and the actual pixel set is large, This will decrease, thereby increasing the weight of small target regions, making the model pay more attention to the segmentation accuracy of small targets.

[0078] Finally, the SAL loss function is defined as:

[0079]

[0080] The loss function dynamically adjusts the weights. This allows the model to achieve good segmentation performance for narrow regions of different scales during training. For small target regions, SAL assigns higher weights, thereby increasing the model's sensitivity to these regions; while for large target regions, SAL maintains stable optimization performance, ensuring overall segmentation quality.

[0081] In summary, addressing the limitations of existing coronary angiography stenosis segmentation methods in handling multi-scale and small-target characteristics, this invention proposes a coronary angiography stenosis segmentation method based on Progressive Scale-Aware Convolutional Aggregation (SAPL), effectively improving segmentation accuracy and reliability. Specifically, firstly, existing methods often only capture limited local information when processing multi-scale features, making it difficult to accurately distinguish target regions in complex backgrounds. To address this, this invention introduces a Scale-Aware Convolutional Aggregation (SCA) module. This module captures local details, boundary information, and multi-scale contextual information through three branches: standard convolution, central difference convolution, and dilated convolution. These features are then dynamically fused using 1×1 convolutions to generate a unified multi-scale feature map. However, while SCA effectively captures multi-scale features, it still falls short in preserving small-target features.

[0082] Secondly, to address the issue of feature preservation for small targets, this invention introduces a Progressive Feature Preservation Block (PFP). PFP progressively preserves key features of small targets through three stages: feature selection, location allocation, and feature preservation. First, the most salient feature points are selected from the input feature tensor to form a feature matrix; then, spatial context is provided to these feature points through location encoding, generating a location-aware feature matrix; finally, a channel attention mechanism further refines the model's focus on narrow regions and suppresses background noise. The introduction of PFP effectively solves the problem of feature loss for small targets and significantly improves segmentation accuracy.

[0083] Finally, existing loss functions (such as IoU loss) are insensitive to the scale of narrow regions, making it difficult to effectively optimize segmentation results at different scales. To address this, this invention proposes a scale-aware loss function (SAL), which dynamically adjusts weights to adapt to narrow regions at different scales, improving sensitivity to small targets. SAL dynamically adjusts weights based on the size difference between the predicted and actual pixel sets, allowing the model to assign higher weights to small target regions during training, thereby improving the segmentation performance for small targets.

[0084] Practice has proven that the SAPL-based method proposed in this invention captures multi-scale features through the SCA module, retains key features of small targets with the help of the PFP module, and optimizes the segmentation results using the SAL function. It demonstrates excellent segmentation performance in both multi-scale and small target characteristics, significantly improving the accuracy and reliability of coronary angiography stenosis segmentation, providing more reliable guidance for clinical surgery, and has significant application value.

[0085] The above description of the embodiments is provided to enable those skilled in the art to understand and apply the present invention. Those skilled in the art can readily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made to the present invention by those skilled in the art based on the disclosure thereof should be within the scope of protection of the present invention.

Claims

1. A scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation, characterized in that, The method comprises the following steps: 1) obtaining X-ray coronary angiography images and corresponding real masks from a public data set; 2) inputting the obtained X-ray coronary angiography images into a scale-aware convolution aggregation module SCA to extract local consistency feature maps, stenosis boundary feature maps and multi-scale context feature maps; 3) using 1x1 convolution to dynamically fuse the three-branch feature maps to obtain unified multi-scale feature maps; 4) inputting the unified multi-scale feature maps into a progressive feature preservation block PFP to obtain enhanced feature maps of small-scale stenosis information through saliency selection, position coding and channel attention; 5) fusing the obtained enhanced feature maps of small-scale stenosis information and the unified multi-scale feature maps in a residual connection manner to obtain enhanced small target feature maps; 6) inputting the enhanced small target feature maps and the X-ray coronary angiography images into a segmentation network to obtain a predicted mask; 7) Constructing a scale-aware loss function SAL, calculating a weight coefficient based on the area difference between the predicted mask and the real mask ; 8) using The weighted IoU loss optimizes the network, so that the model pays more attention to small and narrow areas during training, and finally obtains a segmentation model that can accurately segment different sizes of coronary stenosis. The scale-aware convolution aggregation module SCA comprises: a standard convolution branch for maintaining local spatial consistency; The center difference convolution branch is used to highlight the gradient details of narrow boundaries, and its operation is defined as: ; represents the center difference convolution; represents the standard convolution response; represents the center difference intensity adjustment coefficient, the value range is 0-1; represents the center difference term; a dilated convolution branch comprising a plurality of convolution operations with different dilation rates to expand the receptive field and integrate context information of different scales.

2. The scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation according to claim 1, characterized in that: The unified multi-scale feature map is input into a progressive feature preservation block (PFP), and a strengthened feature map of small-scale narrow information is obtained through saliency selection, specifically including: in each channel of the input feature tensor, the most salient feature point is selected ; wherein The value is dynamically adjusted according to the depth of the network layer, and the calculation formula is: Wherein, is the number of channels of the input feature map, is a preset offset, is the current network layer.

3. The scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation according to claim 2, characterized in that: Inputting the unified multi-scale feature maps into the progressive feature preservation block PFP and obtaining enhanced feature maps of small-scale stenosis information through position coding, so that the enhanced feature maps have spatial position perception ability, and a position-aware feature matrix is obtained, which is used to enhance the relevance of local and global information.

4. The scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation according to claim 2 or 3, characterized in that: Inputting the unified multi-scale feature maps into the progressive feature preservation block PFP and obtaining enhanced feature maps of small-scale stenosis information through channel attention, highlighting important features in small-scale stenosis regions through a weighting manner, and suppressing irrelevant background features, so as to obtain enhanced feature maps.

5. The scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation according to claim 1, characterized in that: The construction scale perception loss function SAL calculates a weight coefficient based on an area difference between a predicted mask and a real mask , and the calculation manner is: ; wherein, is a set of predicted pixels, is a set of real pixels, is a function to compute the scalar variance.

6. The scale-aware progressive learning method for X-ray coronary angiography stenosis segmentation according to claim 1 or 5, characterized in that: The described adoption The weighted IoU loss optimization network is in the form of: ; to enhance the model's focus on small scale targets in training, for predicting a set of pixels, for a set of real pixels.

Images