Coronary artery CTA enhancement method based on cross-scale feature interaction network

Abstract

The invention discloses a coronary artery CTA enhancement method based on a cross-scale feature interaction network. The method comprises the following steps: acquiring an original low-quality coronary artery CTA image; the method comprises the following steps: constructing an enhancement model comprising a residual channel attention (RCAB) local branch, a Transform global branch, a cross-scale attention interaction (CSA) module and a color attention (CAB) correction module; and performing branch feature extraction, cross-scale interactive fusion, contrast correction and residual reconstruction on the original image based on the enhancement model to obtain an enhanced high-quality coronary CTA image. According to the method, the coronary artery anatomy topological characteristics can be comprehensively sensed, and the long-range connectivity of the blood vessel and the fidelity of the CT value are kept while the radiation noise is effectively inhibited.

Description

Technical Field

[0001] This invention relates to medical image processing and deep learning technology, specifically a method for enhancing coronary CT angiography (CTA) based on a cross-scale feature interaction network. Background Technology

[0002] Coronary CT angiography (CTA) is a core non-invasive examination method used clinically to assess coronary artery stenosis, diagnose coronary heart disease, and detect plaque characteristics. High-quality coronary CTA images can clearly display the anatomical details of the vessel wall, lumen, and plaque, providing a foundation for subsequent computer-aided diagnostic (CAD) and hemodynamic analysis (FFR-CT). However, in actual imaging, limited by the spatial resolution of CT detectors, radiation noise from low-dose scanning, motion artifacts caused by patient heart rate fluctuations, and interference from complex anatomical backgrounds, raw coronary CTA images often suffer from insufficient contrast, severe noise, and blurred display of small vessel terminals. While traditional medical image enhancement methods (such as median filtering, nonlocal mean filtering, or CLAHE) can achieve some noise reduction and contrast improvement, they often result in blurred vessel edges or loss of small plaque features when processing coronary artery images with complex branching structures. In recent years, deep learning-based methods, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs), have been widely applied to image enhancement tasks. Despite the progress made by these methods, the following key technical shortcomings remain: ① Imbalance between global morphology and local detail modeling: Coronary arteries have extremely complex spatial topologies. Due to their limited receptive field, CNN models tend to capture local textures of the vessel wall, making it difficult to maintain long-distance vascular connectivity. While Transformer-based methods possess global perception capabilities, they are prone to artifacts or breaks in anatomical structures when processing coronary branches (small vessels) with extremely small pixel proportions due to the lack of local generalization bias. ② Lack of cross-scale feature interaction: The coronary artery system consists of robust main arteries (such as the left coronary trunk and left anterior descending artery) and delicate collateral circulations, spanning a huge scale. Existing enhancement networks mostly perform feature extraction at a single resolution, failing to effectively utilize morphological information at large scales to guide detail repair at small scales, resulting in poor connectivity of enhanced coronary images at fine branches. ③ Difficulty in maintaining contrast (CT value) consistency: Clinical diagnosis of coronary CTA is highly dependent on the accuracy of CT values ​​to distinguish between calcified plaques, soft plaques, and the enhanced lumen. Existing enhancement models often lack dedicated correction mechanisms (such as color or contrast attention constraints) when improving image brightness or contrast, which can easily lead to shifts in the CT grayscale distribution of lesion areas, increasing the risk of clinical misdiagnosis. Therefore, there is an urgent need for a coronary CTA enhancement technique that can effectively combine local detail extraction with cross-scale global feature interaction and has precise contrast correction capabilities to meet the needs of high-quality auxiliary diagnosis. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a coronary CTA enhancement method based on a cross-scale feature interaction network. This method constructs a dual-branch cross-scale sensing network, which can achieve an organic unity between local detail restoration and global morphological preservation, significantly improving image quality while ensuring the accuracy of CT values.

[0004] The technical solution to achieve the objective of this invention is: A coronary CTA enhancement method based on cross-scale feature interaction networks includes the following steps: 1) Acquire raw low-quality coronary CTA images: Read the DICOM format coronary CTA sequence, perform standardization processing, and segment it into three-dimensional slices or two-dimensional projection images; 2) Constructing a cross-scale feature interaction perception enhancement model: The cross-scale feature interaction perception enhancement model includes a cross-scale feature interaction perception network and a contrast consistency correction and reconstruction network. The output of the cross-scale feature interaction perception network is connected to the input of the contrast consistency correction and reconstruction network. The cross-scale feature interaction perception network includes a local feature extraction branch, a global feature perception branch, and a cross-scale attention interaction branch (CSA). The CSA is positioned between the corresponding scale levels of the local and global branches. The local feature extraction branch includes a shallow feature extractor and several stacked residual channel attention blocks (RCABs). The global feature perception branch includes a multi-scale feature block and a set of stacked Transformer blocks. The contrast consistency correction and reconstruction network includes a color attention branch (CAB) and a residual reconstruction branch. The color attention branch (CAB) uses global statistics to calculate the contrast correction coefficient at the channel level. The color attention branch (CAB) has a global spatial pooling layer, a two-layer linear transformation layer, and a sigmoid function. The multi-scale feature block... The block is equipped with downsampling convolutional layers with a stride of 2 and parallel multi-scale convolutional operators to obtain vascular topological features under different receptive fields; 3) Obtaining enhanced high-quality coronary CTA images: Based on a cross-scale feature interaction perception enhancement model, the original low-quality coronary CTA images are enhanced to obtain enhanced high-quality coronary CTA images, specifically as follows: 3-1) The original low-quality coronary CTA image is input into the cross-scale feature interaction perception enhancement model. Based on the local feature extraction branch, local high-frequency detail feature extraction is performed on the image to obtain local enhancement features. Specifically, this includes: based on the shallow feature extractor Convolutional layers extract basic features, and based on the channel attention sub-modules in the residual channel attention blocks, global average pooling is used to obtain channel descriptors. Generate channel weight vector The local enhanced features are obtained by multiplying the weight vector element-wise with the input features; 3-2) Based on the global feature perception branch, long-range topology modeling is performed on the image to obtain global feature information. The steps of long-range topology modeling of the image by the global feature perception branch include: Based on the multi-scale feature extractor in the multi-scale feature block, multi-scale receptive fields are obtained according to convolutional layers with different dilation rates. Based on the multi-head transpose attention mechanism in the Transformer block, feature attention mapping is calculated in the channel dimension to identify the complete path of the coronary artery trunk. 3-3) Based on the cross-scale attention interaction branch, the local enhanced features and global feature information are interactively fused to obtain the fused enhanced features; 3-4) Based on the Color Attention Branch (CAB), the contrast consistency of the fused enhanced features is corrected to obtain the corrected feature map. The Color Attention Branch (CAB) has a global pooling layer, a 1×1 convolutional layer, a corrected linear unit, and a weight mapping layer. The Color Attention Branch (CAB) calculates the global correlation of the image in the RGB space or gray-level dynamic range to generate a contrast enhancement factor, which is used to correct the gray-level contrast distribution between the inside of the blood vessel lumen and the surrounding tissue. Specifically, the overall gray-level statistical features of the enhanced feature map are extracted based on the global spatial pooling layer of the Color Attention Branch (CAB). The correlation of different CT value intervals is explored based on the two-layer linear transformation layer to generate a contrast correction weight map. The correction weight map is multiplied by the fused enhanced features to correct the gray-level distribution shift during the enhancement process. 3-5) Based on the residual reconstruction branch, the corrected feature map is superimposed on the original low-quality coronary CTA image to obtain the enhanced high-quality coronary CTA image, specifically: Local detail mining: The original image is input into the local branch, and the RCABs module is used to suppress noise and enhance the edge sharpness of the blood vessel wall and calcified plaques while maintaining spatial resolution; Global topology modeling: The image is downsampled and input into the global branch. Transformer blocks are used to capture the overall shape of the large-scale structure of the left coronary and anterior descending branches, ensuring the connectivity of the slender branches; Cross-scale interactive fusion: The CSA module is used to inject the macroscopic vascular distribution information obtained from the global branch into the local branch to correct the breakage or artifacts that may occur in the local repair. Contrast correction and consistency maintenance: The interactive features are input into the CAB module to calculate the global contrast distribution weights, ensuring that the enhanced CT values ​​conform to the physical meaning of Hounsfield (HU) units; Residual reconstruction: The corrected feature map is superimposed with the original image using residuals to generate the final enhanced image. The model's training loss function consists of pixel loss L1 and structural consistency loss L... SSIM Perceived damage and loss of color contrast consistency Weighted composition is used to ensure consistency in visual effects and quantitative measurements of enhanced images.

[0005] The residual channel attention block (RCABs) described in step 2) includes a first convolutional layer, a channel attention submodule (Channel Attention), and a second convolutional layer. The output of the first convolutional layer is connected to the input of the channel attention submodule, and the output of the channel attention submodule (Channel Attention) is connected to the input of the second convolutional layer. The channel attention submodule includes a global average pooling layer, a channel reduction convolutional layer, a ReLU activation layer, a channel expansion convolutional layer, and a Sigmoid activation layer. The channel attention submodule uses a global average pooling layer to compress the spatial dimension to obtain a channel-level global feature vector. It adopts a two-layer convolutional structure including channel reduction, ReLU activation, and channel expansion, and combines the Sigmoid function to calculate the gain weights. Through the nesting of long skip connections and short skip connections, the RCABs enable the network to focus on high-frequency detail components such as blood vessel edges and plaque textures.

[0006] The Transformer block described in step 2) includes a layer normalization layer (LN), a multi-head transposed attention mechanism (MHTA), and a gated feed-forward network (GFFN). The output of the layer normalization layer is connected to the input of the multi-head transposed attention mechanism and the input of the gated feed-forward network, respectively. The multi-head transposed attention mechanism uses cross-covariance calculation in the channel dimension to model the long-range topological correlation of the vascular tree, which is used to model the long-range spatial dependence of the vascular tree.

[0007] Step 2) describes the Cross-Scale Attention Interaction Branch (CSA), which is set at the same resolution stage as the local feature extraction branch and the global feature perception branch. It uses cross-attention mapping to dynamically weight and correct the detailed features in the local branch, generates a Key and Value based on global features, and generates a Query based on local features for attention matrix calculation. Specifically: The cross-attention fusion layer uses the morphological feature maps extracted from the global branch as keys and values, and the detail feature maps extracted from the local branch as queries. Based on matrix multiplication, it dynamically guides the repair of local details with global information, as shown in formula (1): Among them, F L G represents local features, and G represents global features.

[0008] This technical solution acquires raw, low-quality coronary CTA images, then introduces a cross-scale feature interaction perception architecture and a color attention correction mechanism to construct a coronary CTA enhancement model. It mines high-frequency local detail features such as vessel edges, lumen details, and plaque textures through a local feature extraction branch centered on a convolutional neural network, and mines the long-range spatial topology and global anatomical information of the coronary tree through a global feature perception branch centered on a Transformer. Furthermore, it achieves deep integration of local refined repair and global morphological guidance through a cross-scale attention interaction module, enhancing the visualization and connectivity of small vessel terminals through a multi-scale interaction mechanism. While maintaining excellent computational efficiency and feature representation accuracy, it achieves dynamic synergy between global branch morphological information and local branch detail information through a cross-attention-based feature interaction module. This fully leverages the complementarity and structural correlation of features under different receptive fields. Finally, based on the coronary CTA enhancement model, the original image is enhanced and contrast-corrected to obtain high-resolution, topologically complete, and CT-value-preserving coronary CTA enhanced images. These images can comprehensively perceive and restore the complex anatomical features of the coronary arteries, effectively suppress radioactive noise, and correct grayscale shifts. This provides high-quality image support for subsequent vascular stenosis assessment and hemodynamic analysis, demonstrating good performance and strong generalization ability.

[0009] This method constructs a dual-branch, cross-scale sensing network that can achieve an organic unity between local detail restoration and global morphological preservation, significantly improving image quality while ensuring the accuracy of CT values. Attached Figure Description

[0010] Figure 1 The diagram shows the structural block diagram of the MSINet model, which includes the CNN branch, the base Transformer branch, and the color correction branch in the embodiment. Figure 2A schematic diagram of the Transformer module structure in the embodiment; Figure 3 A schematic diagram of the cross-scale attention mechanism (CSA) in the embodiment. Detailed Implementation

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0012] Example:

[0013] A coronary CTA enhancement method based on cross-scale feature interaction networks includes the following steps: 1) Acquire raw low-quality coronary CTA images: Read the coronary CTA sequence in clinical DICOM format, preprocess the three-dimensional volume data, including spatial resampling and grayscale normalization, and map the CT values ​​(Hounsfield Unit, HU) to the [0, 1] interval to adapt to the input requirements of the deep learning model; 2) Construct a cross-scale feature interaction perception enhancement model: such as Figure 1 , Figure 2 , Figure 3 As shown, the enhanced model (MSINet) includes a cross-scale feature interaction perception network and a contrast consistency correction and reconstruction network. The output of the cross-scale feature interaction perception network is connected to the input of the contrast consistency correction and reconstruction network. The cross-scale feature interaction perception network includes a local feature extraction branch, a global feature perception branch, and a cross-scale attention interaction module (CSA), specifically: 2-1) The local feature extraction branch has a shallow feature extractor and several stacked residual channel attention blocks (RCABs). The shallow extractor consists of a 3×3 convolutional layer. The RCABs have a first convolutional layer, a channel attention submodule (CA), and a second convolutional layer. The channel attention submodule compresses the spatial dimension through a global average pooling layer to obtain a channel-level global feature vector. Then, through a two-layer 1×1 convolutional structure containing channel reduction, ReLU activation, and channel expansion, the gain weight of each channel is calculated using the Sigmoid function. This is used to adaptively identify and enhance the features of blood vessel edges while filtering out high-frequency noise. 2-2) The global feature perception branch includes multi-scale feature blocks and several stacked Transformer blocks. The multi-scale feature blocks obtain multi-scale receptive fields through convolutional layers with different dilation rates. The Transformer blocks specifically include layer normalization (LN), multi-head transposed attention mechanism (MDTA), and gated feedforward network (GDFN). The multi-head transposed attention mechanism (MDTA) can model the long-range topological correlation of vascular trees at a lower computational cost by calculating cross-covariance in the channel dimension rather than the spatial dimension. 2-3) The Cross-Scale Attention Interaction Module (CSA) is set at the same resolution stage of the local and global branches. CSA uses the anatomical topological information provided by the global branch as a guide and performs dynamic weighted correction on the detailed features in the local branch through cross-attention mapping. Its formulaic expression is to use global features to generate Key and Value, and use local features to generate Query for attention matrix calculation. More specifically, the contrast consistency correction reconstruction network includes the Color Attention Branch (CAB) and the residual reconstruction module. The CAB branch extracts the overall gray-level statistical features of the enhancement feature map through the global spatial pooling layer, explores the correlation of different CT value intervals using a two-layer linear transformation layer (1×1 convolution), and generates a contrast correction weight map. This branch can effectively constrain the gray-level distribution shift that the depth model may cause during the enhancement process, and ensure that the CT value contrast relationship between calcified plaques, fat plaques and luminal contrast agents meets the clinical diagnostic criteria. 3) Enhancement processing of coronary CTA images based on enhancement models: including: 3-1) Parallel Feature Extraction and Preliminary Local Feature Mining: The original image is input into the local feature extraction branch of the enhancement model. The Residual Channel Attention (RCAB) mechanism is used to capture high-frequency local information such as patch texture and pipe wall edges. During this process, the channel attention submodule adaptively rescales the feature channels through global statistical information. First, channel descriptors are obtained through global average pooling. As shown in formula (2): Next, the weight vectors s of each channel are generated by double convolution and the Sigmoid function as shown in formula (3): The final local enhancement features are obtained. ,in For channel reduction layer, For channel expansion layer, and These are the ReLU and Sigmoid functions, respectively. 3-2) Multi-scale global feature perception and long-range modeling: After downsampling, the image is input into the global feature perception branch. The Transformer block perceives the global structure of the coronary artery image through the MDTA mechanism. MDTA obtains spatial semantic information by calculating the attention map between channels, as shown in formula (4): in, It is used for control With its distributed learnable scaling operator, MDTA can identify the complete path of the main coronary arteries (such as the left main coronary artery LM and the left anterior descending artery LAD) and maintain long-range connectivity of the vessels even under conditions of extremely low local signal-to-noise ratio. 3-3) Cross-scale feature interaction fusion and structural correction: based on CSA module and The CSA module performs feature interaction processing by injecting low-frequency structural information from the global branch into the local branch, thus correcting the local breaks caused by artifacts as shown in formula (5): This step ensures that the enhanced image has both clear details and anatomical structural integrity. 3-4) Color (Contrast) Attention Correction and Residual Reconstruction: Based on the Color Attention Branch (CAB), contrast consistency constraints are applied to the fused features. The CAB calculates the cross-covariance mapping of global features and generates a grayscale correction coefficient vector. As shown in formula (6): This process prevents the model from exhibiting "false negative" or "false positive" enhancements (e.g., mistakenly enhancing mild calcification as severe calcification). Finally, using a residual learning mechanism, the corrected feature map is reconstructed into a prediction residual, which is then superimposed on the input image to output the final high-resolution coronary CTA image as shown in Equation (7):

[0014] The method described in this example has the following advantages compared to existing technologies: 1) To address the challenge of fragile small branches in coronary CTA, this method combines a dual-branch MSINet architecture that integrates local CNN perception with global Transformer modeling. Through the branches of the Transformer, this method can model the anatomical topology of the vascular tree from a global perspective, effectively compensating for the technical defects of traditional CNN models, such as limited receptive field and susceptibility to motion artifacts leading to loss of vascular continuity. 2) This method introduces a cross-scale attention interaction module (CSA), which breaks the isolation of information flow between different resolution levels in traditional multi-scale networks. Through the CSA module, the large-scale morphological information of the global branch can dynamically guide the local branch to perform fine repair, which significantly improves the accuracy of the reconstruction of the terminal coronary vessels in complex anatomical backgrounds (such as pericardial fat or lung tissue interference). 3) In response to the stringent requirements for the accuracy of CT values ​​(HU values) in clinical medical imaging, the color (contrast) attention branch (CAB) and the corresponding contrast consistency correction logic in this method can accurately lock the key grayscale range of medical images and correct the color shift and grayscale drift generated by the deep learning model during the image contrast enhancement process. This ensures that the enhanced images are not only visually clear, but can also be reliably used for subsequent precision medical applications such as stenosis rate quantification, plaque component analysis and FFR-CT.

Claims

1. A method for enhancing coronary CTA based on cross-scale feature interaction networks, characterized in that, Includes the following steps: 1) Acquire raw low-quality coronary CTA images: Read the DICOM format coronary CTA sequence, perform standardization processing, and segment it into three-dimensional slices or two-dimensional projection images; 2) Constructing a cross-scale feature interaction perception enhancement model: The cross-scale feature interaction perception enhancement model includes a cross-scale feature interaction perception network and a contrast consistency correction and reconstruction network. The output of the cross-scale feature interaction perception network is connected to the input of the contrast consistency correction and reconstruction network. The cross-scale feature interaction perception network includes a local feature extraction branch, a global feature perception branch, and a cross-scale attention interaction branch (CSA). The cross-scale attention interaction branch (CSA) is set between the corresponding scale levels of the local and global branches. The local feature extraction branch includes a shallow feature extractor and several stacked residual channel attention blocks (RCABs). The global feature perception branch includes a multi-scale feature block and a set of stacked Transformer blocks. The contrast consistency correction and reconstruction network includes a color attention branch (CAB) and a residual reconstruction branch. The color attention branch (CAB) uses global statistics to calculate the contrast correction coefficient at the channel level. The color attention branch (CAB) has a global spatial pooling layer, a two-layer linear transformation layer, and a sigmoid function. The multi-scale feature block has a downsampling convolutional layer with a stride of 2 and parallel multi-scale convolution operators to obtain vascular topological features under different receptive fields. 3) Obtaining enhanced high-quality coronary CTA images: Based on a cross-scale feature interaction perception enhancement model, the original low-quality coronary CTA images are enhanced to obtain enhanced high-quality coronary CTA images, specifically as follows: 3-1) The original low-quality coronary CTA image is input into the cross-scale feature interaction perception enhancement model. Based on the local feature extraction branch, local high-frequency detail feature extraction is performed on the image to obtain local enhancement features. Specifically, this includes: based on the shallow feature extractor Convolutional layers extract basic features, and based on the channel attention sub-modules in the residual channel attention blocks, global average pooling is used to obtain channel descriptors. Generate channel weight vector The local enhanced features are obtained by multiplying the weight vector element-wise with the input features; 3-2) Based on the global feature perception branch, long-range topology modeling is performed on the image to obtain global feature information. The steps of long-range topology modeling of the image by the global feature perception branch include: Based on the multi-scale feature extractor in the multi-scale feature block, multi-scale receptive fields are obtained according to convolutional layers with different dilation rates. Based on the multi-head transpose attention mechanism in the Transformer block, feature attention mapping is calculated in the channel dimension to identify the complete path of the coronary artery trunk. 3-3) Based on the cross-scale attention interaction branch, the local enhanced features and global feature information are interactively fused to obtain the fused enhanced features; 3-4) Based on the Color Attention Branch (CAB), the contrast consistency of the fused enhanced features is corrected to obtain the corrected feature map. The Color Attention Branch (CAB) has a global pooling layer, a 1×1 convolutional layer, a corrected linear unit, and a weight mapping layer. The Color Attention Branch (CAB) calculates the global correlation of the image in the RGB space or gray-level dynamic range to generate a contrast enhancement factor, which is used to correct the gray-level contrast distribution between the inside of the blood vessel lumen and the surrounding tissue. Specifically, the overall gray-level statistical features of the enhanced feature map are extracted based on the global spatial pooling layer of the Color Attention Branch (CAB). The correlation of different CT value intervals is explored based on the two-layer linear transformation layer to generate a contrast correction weight map. The correction weight map is multiplied by the fused enhanced features to correct the gray-level distribution shift during the enhancement process. 3-5) Based on the residual reconstruction branch, the corrected feature map is superimposed on the original low-quality coronary CTA image to obtain the enhanced high-quality coronary CTA image, specifically: Local detail mining: The original image is input into the local branch, and the RCABs module is used to suppress noise and enhance the edge sharpness of the blood vessel wall and calcified plaques while maintaining spatial resolution; Global topology modeling: The image is downsampled and input into the global branch. Transformer blocks are used to capture the overall shape of the large-scale structure of the left coronary and anterior descending branches, ensuring the connectivity of the slender branches; Cross-scale interactive fusion: The CSA module is used to inject the macroscopic vascular distribution information obtained from the global branch into the local branch to correct the breakage or artifacts that may occur in the local repair. Contrast correction and consistency maintenance: The interactive features are input into the CAB module to calculate the global contrast distribution weights, ensuring that the enhanced CT values ​​conform to the physical meaning of Hounsfield (HU) units; Residual reconstruction: The corrected feature map is superimposed with the original image using residuals to generate the final enhanced image. The model's training loss function consists of pixel loss L1 and structural consistency loss L... SSIM Perceived damage and loss of color contrast consistency Weighted composition is used to ensure consistency in visual effects and quantitative measurements of enhanced images.

2. The coronary CTA enhancement method based on cross-scale feature interaction networks according to claim 1, characterized in that, The residual channel attention block (RCABs) described in step 2) includes a first convolutional layer, a channel attention submodule (ChannelAttention), and a second convolutional layer. The output of the first convolutional layer is connected to the input of the channel attention submodule, and the output of the channel attention submodule (ChannelAttention) is connected to the input of the second convolutional layer. The channel attention submodule includes a global average pooling layer, a channel reduction convolutional layer, a ReLU activation layer, a channel expansion convolutional layer, and a Sigmoid activation layer. The channel attention submodule uses a global average pooling layer to compress the spatial dimension to obtain a channel-level global feature vector. It adopts a two-layer convolutional structure including channel reduction, ReLU activation, and channel expansion, and combines the Sigmoid function to calculate the gain weights. Through the nesting of long skip connections and short skip connections, the RCABs enable the network to focus on high-frequency detail components such as blood vessel edges and plaque textures.

3. The coronary CTA enhancement method based on cross-scale feature interaction networks according to claim 1, characterized in that, The Transformer block described in step 2) includes a layer normalization layer (LN), a multi-head transposed attention mechanism (MHTA), and a gated feedforward network (GFFN). The output of the layer normalization layer is connected to the input of the multi-head transposed attention mechanism and the input of the gated feedforward network, respectively. The multi-head transposed attention mechanism uses cross-covariance calculation in the channel dimension to model the long-range topological correlation of the vascular tree.

4. The coronary CTA enhancement method based on cross-scale feature interaction networks according to claim 1, characterized in that, Step 2) describes the Cross-Scale Attention Interaction Branch (CSA), which is set at the same resolution stage as the local feature extraction branch and the global feature perception branch. It uses cross-attention mapping to dynamically weight and correct the detailed features in the local branch, generates a Key and Value based on global features, and generates a Query based on local features for attention matrix calculation. Specifically: The cross-attention fusion layer uses the morphological feature maps extracted from the global branch as keys and values, and the detail feature maps extracted from the local branch as queries. Based on matrix multiplication, it dynamically guides the repair of local details with global information, as shown in formula (1): Among them, F L G represents local features, and G represents global features.

Images