A low-dose coronary CTA denoising method based on segmentation guidance

Abstract

The application discloses a low-dose coronary CTA denoising method based on segmentation guidance, which comprises the following steps: firstly, obtaining the DICOM image data of the coronary CTA scanning reconstruction under low radiation dose, and pre-processing the data; then inputting the image data into a coarse denoising network for preliminary denoising; then sending the coarse denoising result into a segmentation network for segmentation; finally, feeding the coarse denoising result and the segmentation result into a fine denoising network together, and obtaining a more fine denoising result by using the guiding effect of the segmentation result. The application realizes the overall denoising of the low-dose coronary CTA, and the image after fine denoising can provide technical support for the diagnosis of clinicians; meanwhile, the application can also give the coronary segmentation result with great reference significance, and can meet the diagnosis requirements of clinicians to a certain extent.

Description

Technical Field

[0001] This invention relates to a segmentation-guided low-dose coronary CTA noise reduction method, belonging to the field of computer image processing. Background Technology

[0002] The rapid development of artificial intelligence technology in the field of medical imaging has become a new frontier in assisting the diagnosis of cardiovascular diseases. Given that medical diagnosis relies heavily on high-quality images, building advanced imaging algorithm frameworks can not only improve diagnostic accuracy but also promote the application of domestically produced medical equipment in high-end medical imaging, meeting the national strategic need to improve the diagnosis and treatment of major diseases.

[0003] In the diagnosis of cardiovascular diseases, especially coronary artery diseases, high-quality imaging and accurate image analysis are crucial for disease management and treatment planning. Computed Tomography Angiography (CTA) is a non-invasive medical imaging technique that uses CT scanners to perform tomographic scans of the patient's heart and injects contrast agents intravenously to enhance the visualization of the coronary arteries, thus clearly showing their anatomical structure and blood flow. Due to its high imaging resolution, non-invasiveness, and convenience, coronary CTA has become an important clinical tool for early screening of coronary artery diseases. Clinically, traditional coronary CTA usually requires a high radiation dose to obtain high-quality images, which exposes patients to significant radiation doses, posing a potential health risk. Therefore, low-dose coronary CTA has become a current research hotspot. However, reducing the radiation dose increases noise and artifacts in coronary CTA imaging, leading to a decline in image quality. This not only seriously affects the diagnostic accuracy of physicians but also presents significant challenges to subsequent image processing. How to reduce radiation dose while maintaining image quality has become a key challenge in technological development.

[0004] In low-dose CTA image processing, denoising techniques are crucial for improving image quality. Currently, common denoising techniques include filtering algorithms and iterative reconstruction algorithms. While these techniques can reduce image noise to some extent, they often sacrifice image detail. This is especially true in coronary CTA images, where accurate identification of fine vascular structures and lesion areas is critical for diagnosis. More importantly, current denoising techniques often focus solely on noise removal without specifically optimizing for subsequent image processing tasks, such as coronary artery segmentation. In coronary CTA images, the goal of denoising is not only to improve image readability but, more importantly, to support accurate vessel segmentation. Over- or inappropriate denoising can blur vessel boundaries, making subsequent segmentation difficult and impacting overall diagnostic accuracy. Therefore, developing a segmentation-guided denoising method in conjunction with coronary artery segmentation is of great significance for improving the quality and diagnostic value of low-dose coronary CTA images.

[0005] This invention proposes a novel segmentation-guided low-dose coronary CTA denoising method. This method combines advanced image denoising algorithms with segmentation techniques to effectively eliminate noise while preserving crucial coronary artery structural information, and can also generate clinically significant coronary artery segmentation results. By designing a segmentation network to coarsely segment the coronary artery structures in the image, this method allows the denoising network to perform more targeted image denoising processing, thereby reducing noise while maximizing the preservation of important coronary artery structural information. This approach not only improves the overall image quality, enhancing the quality and accuracy of medical diagnosis and providing more accurate and reliable information for subsequent clinical analysis and diagnostic decisions, but also promotes the development of domestic medical technology, further strengthening the national healthcare system, and paving the way for the development of domestically produced medical equipment in the field of high-end medical imaging. Summary of the Invention

[0006] To eliminate noise in low-dose coronary CTA and restore high-quality images, while mitigating the problem of excessive denoising leading to the loss of subtle coronary artery structures in existing denoising algorithms, this invention provides a segmentation-guided low-dose coronary CTA denoising method. This method uses segmentation to guide the denoising process, effectively restoring subtle coronary artery branch structures, improving image quality, reducing noise, and generating coronary artery segmentation results with significant reference value, providing technical support for clinicians in the early diagnosis and treatment of cardiovascular diseases.

[0007] To address the aforementioned technical problems, this invention provides a segmentation-guided low-dose coronary CTA noise reduction method, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0008] Step 1: Acquire DICOM image data after low-dose coronary CTA reconstruction from clinical scans;

[0009] Step 2: Preprocess the ICO M image data obtained in step (1), set the FOV range according to the ICO M header file, and perform normalization processing;

[0010] Step 3: Input the preprocessed data from step (2) into the coarse noise reduction network to obtain a coarse noise reduction result with a lower noise level;

[0011] Step 4: The coarse noise reduction result of step (3) is fed into the segmentation network, so that the segmentation network can segment the coronary artery branches while avoiding some noise.

[0012] Step 5: Using the segmentation results obtained in step (4) to guide the coarse denoising results in step (3), input both into the fine denoising network to obtain the final, more refined denoising results.

[0013] Specifically,

[0014] In step 2, threshold filtering is used to preprocess the DICOM image data, and the FOV range is recorded.

[0015] In step 2, the DICOM image data is normalized according to the slope and intercept of the data acquisition device.

[0016] The coarse denoising network in step 3, the segmentation network in step 4, and the fine denoising network in step 5 share the same network architecture, each consisting of an encoder and a decoder. The encoder has four encoder blocks to learn denoising features at different scales, with the feature map size decreasing and the number of channels increasing progressively with each level. Correspondingly, the decoder also has four decoder blocks, with the feature map size increasing and the number of channels decreasing progressively with each level. Adjacent encoder blocks are connected by subpixel downsampling convolutional layers, and adjacent decoder blocks are connected by subpixel upsampling convolutional layers. Subpixel convolutional layers achieve upsampling and downsampling by increasing or decreasing the number of channels and scrambling pixels, effectively restoring image details, reducing image blurring, and avoiding the checkerboard artifact effect caused by transposed convolution. Each encoder and decoder level consists of two residual channel attention modules and a cascaded convolutional layer. The residual channel attention module integrates the advantages of channel attention and residual mechanisms, enabling it to fully learn effective information. The channel attention mechanism allows the network to focus on important information along the channel dimension during feature extraction, as different channels typically contain different feature representations. Channel attention learns a weight vector to weight each channel, enabling the network to focus more on important feature channels. The residual mechanism allows the network to learn residual mappings, which helps address the vanishing gradient and representation bottleneck problems in deep networks. By introducing residual connections, the network can more easily learn the direct mapping relationship between input and output, thereby improving denoising performance. Furthermore, skip connections exist between sibling encoder and decoder blocks. A novel coordinate gating module is embedded in these skip connections to fuse denoising features at different scales from sibling encoder and decoder blocks. The coordinate gating module first compresses the features from the encoder and decoder blocks at the same level by performing 3×3 convolutions and then average pooling operations in both the width and height dimensions. Next, the encoder and decoder features, pooled in the width dimension, are concatenated along the channel dimension, and a 1×1 convolution is used to reduce the number of channels to match the input feature dimension; the length dimension is handled similarly. Then, the compressed features obtained in the width and height dimensions are multiplied pixel-by-pixel and fused using a 3×3 convolution. Finally, the fused features are activated and multiplied by both the encoder and decoder features, then summed to obtain the fused feature representation. The feature transfer process of the coordinate gating module is shown below:

[0017] E H =GAP H (E),E W =GAP W (E),

[0018] D H =GAP H (D),D W=GAP W (D),

[0019] M = Concat(E) H D H W1×Concat(E W D W W2,

[0020] E A =σ(MW3),D A =σ(MW4),

[0021] F = E A ×E+D A ×D,

[0022] Where E and D represent encoder features and decoder features, respectively, GAP H and GAP W E represents average pooling in the height and width dimensions, respectively. H and E W D represents the encoder features after average pooling in the height and width dimensions, respectively. H and D W W1, W2, W3, and W4 represent the decoder features after average pooling in the height and width dimensions, respectively. W1, W2, W3, and W4 represent 1×1 convolutions, Concat represents the concatenation operation, M represents the feature map to be activated, and E represents the feature map to be activated. A and D A represents the feature maps after the encoder features and decoder features are activated, respectively, and F represents the fused features.

[0023] The segmentation network in step 4 includes a unique cross-task guided refinement module for the interaction between the coarse denoising network and the segmentation network. This module utilizes unique intra-task information and complementary inter-task information to refine, fuse, and transfer denoising and segmentation features, enhancing and normalizing the prediction process for denoising and segmentation. This module consists of two parts: intra-task refinement and inter-task refinement. Intra-task refinement calculates an affinity matrix to filter and capture task similarity information to enhance contextual information. Inter-task refinement utilizes cross-attention and self-attention to refine and fuse cross-task features to achieve feature refinement. Specifically, in the intra-task refinement part, the input features pass through an affinity layer to calculate an affinity matrix. Since each row of the affinity matrix represents the correlation between a location and all other locations, the affinity matrix can utilize pixel-level similarity to achieve structure preservation for denoising and target region partitioning for segmentation. It provides a prerequisite for the interaction between the two tasks. In calculating the affinity matrix, the input features are first flattened in space, then multiplied with their transposed matrix. Finally, each row of the product is normalized to obtain the affinity matrix, as this reduces the impact of scale. After obtaining the affinity matrix, the input features are weighted and aggregated into the enhanced response. In the inter-task refinement, cross-attention is first used to calculate the similarity between the query vector Q and key vector K, guiding the weighting of their respective value vectors V to establish their correlations and highlight similar features. Then, a combined affinity matrix is ​​calculated by multiplying the affinity matrices of the two tasks element-wise to integrate complementary information between tasks. Subsequently, since each task contains unique features, self-attention is applied to each task separately, guided by the combined affinity matrix, to establish relationships between global pixels, thus fully fusing the features of the two tasks. When calculating the query vector Q, key vector K, and value vector V, depthwise convolution can better replace linear projection for embedding features. Compared to linear projection, depthwise convolution can preserve the spatial location information of pixels and reduce information loss, which is highly beneficial for noise reduction and segmentation. Finally, the fused high-level features are further refined to enhance their representational capabilities and then passed to the segmentation network and the fine denoising network respectively to assist in segmentation and denoising. The feature transfer process of the cross-task guided refinement module can be represented as follows:

[0024] M D =Aff(D),M S =Aff(S),

[0025]

[0026] D′=MHSA(F D ),S′=MHSA(F S ),

[0027] Where D and S represent denoising features and segmentation features, Aff represents affinity layer, and M represents M. D and M S Let denoise the denoising affinity matrix and the segmentation affinity matrix, respectively, and σ denote the sigmoid activation function. and This represents the refined denoising and segmentation features within the task, MHCA represents multi-head cross-attention, Concat represents the connection operation, and F... D and F S The symbols represent intermediate features refined across tasks, MHSA represents multi-head self-attention, and D′ and S′ represent the denoising and segmentation features obtained after cross-task guided refinement. Similarly, for simplicity, all convolutions are omitted in the flowchart.

[0028] In step 5, the fine denoising network, on the one hand, uses both the coarse denoising result and the segmentation result as input to provide global guidance for the segmentation result; on the other hand, it integrates a novel aggregation layer within a higher-level network to aggregate denoising features from the fine denoising network, high-level denoising features output from the cross-task guided refinement module, and high-level segmentation features from the segmentation network. Specifically, the coarse denoising result and the segmentation result are input into the fine denoising network together, allowing the segmentation result to guide the denoising process globally, helping the denoising network identify locations that require significant denoising and detail restoration. Accordingly, the aggregation layer first concatenates the denoising features and high-level segmentation features along the channel dimension, and then sequentially feeds them into a 1×1 convolutional layer, a 3×3 depthwise convolutional layer, and a 1×1 convolutional layer for preliminary feature fusion. Depthwise convolution is used instead of standard convolution to perform feature fusion in both the channel and spatial dimensions, making the fused features more targeted while significantly reducing the number of parameters and lowering the model's spatial complexity. Next, the preliminarily fused features are added element-wise with the high-level denoising features and then concatenated with the segmentation features along the channel dimension. The aim is to achieve the full aggregation and refinement of denoising features by leveraging the guiding role of segmentation features in deep semantics. Finally, the concatenated features are further fused and the output channels are adjusted through 3×3 convolution to obtain higher-level denoising features with richer dimensionality matching and stronger representational capabilities.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] 1. This invention takes into account the image processing task after noise reduction, combining noise reduction and segmentation. It uses segmentation to guide the noise reduction process, which can effectively restore the fine coronary artery structure damaged by noise, improve image quality, and provide coronary artery segmentation results with high reference value. This provides strong technical support for doctors' clinical diagnosis and has important clinical application value.

[0031] 2. This invention integrates coordinate gating modules in all three networks. These modules use coordinate information and gating mechanisms to capture key noise reduction and segmentation features, significantly improving the recovery of coronary artery fine structures and the segmentation effect, thus minimizing the impact of noise on the segmentation network's generated segmentation results.

[0032] 3. This invention applies a cross-task guided refinement module in the segmentation network. This module uses unique information within a task and complementary information between tasks to refine, fuse, and transfer denoising and segmentation features. This not only reduces the loss of denoising and segmentation information during interaction but also implicitly leverages the guiding role of segmentation on denoising, enabling the fine denoising network to achieve more precise enhancement of the coronary artery's fine structure.

[0033] 4. This invention constructs an aggregation layer within the fine denoising network. This layer aggregates denoising features from the fine denoising network, high-level denoising features output by the cross-task guided refinement module, and high-level segmentation features from the segmentation network. The aim is to aggregate richer and more representative high-level denoising features, thereby achieving high-quality restoration of low-dose coronary CTA images.

[0034] Without losing the delicate structure of the coronary arteries.

[0035] In summary, this invention, through advanced network and algorithm design, has significant advantages over existing technologies in improving image quality, reducing radiation risks, and enhancing diagnostic efficiency, and is of great significance for clinical medical practice and patient health. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0037] Figure 2 This is a schematic diagram of the basic structure of the three networks in this invention, taking the coarse noise reduction network as an example;

[0038] Figure 3 A schematic diagram of the residual channel attention module that constitutes the three networks;

[0039] Figure 4 This is a structural diagram of the coordinate gating module;

[0040] Figure 5 A structural diagram of the detailed module for cross-task guidance;

[0041] Figure 6 This is a schematic diagram of the polymer layer structure;

[0042] Figure 7 The first row is a schematic diagram of the noise reduction results of the present invention; the second row is an enlarged view of the red rectangle in the first row.

[0043] Figure 8 This is a schematic diagram of the segmentation result of the present invention. Detailed Implementation

[0044] The technical solution provided by the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0045] Example: Figure 1 As shown, a segmentation-guided low-dose coronary CTA noise reduction method includes the following steps:

[0046] Step 1: Acquire DICOM image data after low-dose coronary CTA reconstruction from clinical scans;

[0047] Step 2: Preprocess the ICO M image data obtained in step (1), set the FOV range according to the ICO M header file, and perform normalization processing;

[0048] Step 3: Input the preprocessed data from step (2) into the coarse noise reduction network to obtain a coarse noise reduction result with a lower noise level;

[0049] Step 4: The coarse noise reduction result of step (3) is fed into the segmentation network, so that the segmentation network can segment the coronary artery branches while avoiding some noise.

[0050] Step 5: Using the segmentation results obtained in step (4) to guide the coarse denoising results in step (3), input both into the fine denoising network to obtain the final, more refined denoising results.

[0051] In step 2, threshold filtering is used to preprocess the DICOM image data, and the FOV range is recorded.

[0052] In step 2, the DICOM image data is normalized according to the slope and intercept of the data acquisition device.

[0053] The coarse noise reduction network in step 3, the segmentation network in step 4, and the fine noise reduction network in step 5 have the same network architecture, each consisting of an encoder and a decoder, such as... Figure 2As shown, the encoder has four encoder blocks to learn noise reduction features at different scales. The feature map size output by each encoder block decreases progressively, while the number of channels increases progressively. Correspondingly, the decoder also has four decoder blocks, with the feature map size output by each decoder block increasing progressively and the number of channels decreasing progressively. Adjacent encoder blocks are connected by subpixel downsampling convolutional layers, and adjacent decoder blocks are connected by subpixel upsampling convolutional layers. Subpixel convolutional layers achieve upsampling and downsampling by increasing or decreasing the number of channels and scrambling pixels, which can effectively restore image details, reduce image blurring, and avoid the checkerboard artifact effect caused by transposed convolution. Each encoder and decoder consists of two residual channel attention modules and one convolutional layer cascaded together, as shown below. Figure 3 As shown, the residual channel attention module integrates the advantages of channel attention and residual mechanisms, enabling it to fully learn effective information. The channel attention mechanism allows the network to focus on important information along the channel dimension during feature extraction, as different channels typically contain different feature representations. Channel attention learns a weight vector to weight each channel, allowing the network to focus more on important feature channels. The residual mechanism allows the network to learn residual mappings, which helps solve the gradient vanishing and representation bottleneck problems in deep networks. By introducing residual connections, the network can more easily learn the direct mapping relationship between input and output, thereby improving denoising performance. Furthermore, skip connections exist between sibling encoder and decoder blocks. A novel coordinate gating module is embedded in these skip connections to fuse denoising features at different scales from sibling encoder and decoder blocks. A schematic diagram of the coordinate gating module is shown below. Figure 4 As shown in the diagram, the coordinate gating module first compresses the scale of features from the encoder and decoder blocks at the same level by performing 3×3 convolutions and then average pooling operations in both the width and height dimensions. Next, the encoder and decoder features, pooled in the width dimension, are concatenated along the channel dimension, and a 1×1 convolution is used to reduce the number of channels to match the input feature dimension; the length dimension is handled similarly. Then, the compressed features obtained in the width and height dimensions are multiplied pixel-by-pixel and fused using a 3×3 convolution. Finally, the fused features are activated and multiplied by both the encoder and decoder features, then summed to obtain the fused feature representation. The feature transmission flow of the coordinate gating module is shown below:

[0054] E H =GAP H (E),E W =GAP W (E),

[0055] D H =GAP H (D),D W =GAPW (D),

[0056] M = Concat(E) H D H W1×Concat(E W D W W2,

[0057] E A =σ(MW3),D A =σ(MW4),

[0058] F = E A ×E+D A ×D,

[0059] Where E and D represent encoder features and decoder features, respectively, GAP H and GAP W E represents average pooling in the height and width dimensions, respectively. H and E W D represents the encoder features after average pooling in the height and width dimensions, respectively. H and D W W1, W2, W3, and W4 represent the decoder features after average pooling in the height and width dimensions, respectively. W1, W2, W3, and W4 represent 1×1 convolutions, Concat represents the concatenation operation, M represents the feature map to be activated, and E represents the feature map to be activated. A and D A represents the feature maps after the encoder features and decoder features are activated, respectively, and F represents the fused features.

[0060] The segmentation network in step 4 includes a unique cross-task guided refinement module for the interaction between the coarse denoising network and the segmentation network. This module utilizes unique intra-task information and complementary inter-task information to refine, fuse, and transfer denoising and segmentation features, thereby enhancing and standardizing the prediction process for denoising and segmentation. This module consists of two parts: intra-task refinement and inter-task refinement, such as... Figure 5 As shown in (a), intra-task refinement computes an affinity matrix to filter and capture task similarity information to enhance contextual information. Inter-task refinement utilizes cross-attention and self-attention to refine and fuse cross-task features to achieve feature refinement. Specifically, in the intra-task refinement part, the input features pass through an affinity layer to compute an affinity matrix, as shown in (a). Figure 5As shown in (b), since each row of the affinity matrix represents the correlation between a location and all other locations, the affinity matrix can utilize pixel-level similarity to achieve structure preservation in noise reduction and target region segmentation. It provides a prerequisite for the interaction between the two tasks. In the process of calculating the affinity matrix, the input features are first flattened in the spatial dimension, then multiplied with their own transposed matrix, and finally each row of the product is normalized to obtain the affinity matrix, as this reduces the influence of scale. After obtaining the affinity matrix, the input features are weighted and summarized into the enhanced response. In the inter-task refinement part, cross-attention is first used to calculate the similarity between the query vector Q and the key vector K between the tasks, and guides the weighting of their respective value vectors V to establish their correlation, thereby highlighting similar features. Then, the combined affinity matrix is ​​calculated by multiplying the affinity matrices of the two tasks element by element to integrate the complementary information between the tasks. Afterwards, since each task contains some unique features, self-attention is applied to each task separately under the guidance of the combined affinity matrix to establish the relationship between global pixels, thereby fully fusing the features of the two tasks. When computing the query vector Q, key vector K, and value vector V, depthwise convolution can better replace linear projection for embedding features. Compared to linear projection, depthwise convolution can preserve the spatial location information of pixels and reduce information loss, which is very beneficial for denoising and segmentation. Finally, the fused high-level features are sufficiently refined to enhance representational power and are passed to the segmentation network and the fine denoising network respectively to assist in segmentation and denoising. The feature transfer process of the cross-task guided refinement module can be represented as follows:

[0061] M D =Aff(D),M S =Aff(S),

[0062]

[0063] D′=MHSA(F D ),S′=MHSA(F S ),

[0064] Where D and S represent denoising features and segmentation features, Aff represents affinity layer, and M represents M. D and M S Let denoise the denoising affinity matrix and the segmentation affinity matrix, respectively, and σ denote the sigmoid activation function. and This represents the refined denoising and segmentation features within the task, MHCa represents multi-head cross-attention, Concat represents the connection operation, and F... D and F SThe symbols represent intermediate features refined across tasks, MHSA represents multi-head self-attention, and D′ and S′ represent the denoising and segmentation features obtained after cross-task guided refinement. Similarly, for simplicity, all convolutions are omitted in the flowchart.

[0065] In step 5, the fine denoising network, on the one hand, uses both the coarse denoising result and the segmentation result as input to exert a global guiding role for the segmentation result; on the other hand, it integrates a novel aggregation layer within a high-level network to aggregate denoising features from the fine denoising network, high-level denoising features output from the cross-task guided refinement module, and high-level segmentation features from the segmentation network. Its structural diagram is shown below. Figure 6 As shown in the diagram. Specifically, the coarse denoising result and the segmentation result are input into the fine denoising network together. This allows the segmentation result to guide the denoising process globally, helping the denoising network identify locations that require significant denoising and detail restoration. Correspondingly, the aggregation layer first concatenates the denoising features and high-level segmentation features along the channel dimension, then sequentially feeds them into a 1×1 convolutional layer, a 3×3 depthwise convolutional layer, and a 1×1 convolutional layer for initial feature fusion. Depthwise convolutions are used instead of standard convolutions to perform feature fusion in both the channel and spatial dimensions, making the fused features more targeted while significantly reducing the number of parameters and lowering the model's spatial complexity. Next, the initially fused features are added element-wise with the high-level denoising features and then concatenated with the segmentation features along the channel dimension. The aim is to achieve sufficient aggregation and refinement of the denoising features through the deep semantic guidance of the segmentation features. Finally, the concatenated features are further fused and the output channels are adjusted using a 3×3 convolution to obtain richer, more dimensionally matched, and more representative high-level denoising features.

[0066] To verify the effectiveness of the present invention, a clinically acquired low-dose coronary CTA data sample was used to demonstrate the advantages of the method disclosed in this invention over existing algorithms. Figure 7 The denoising results obtained using the method provided in this invention and existing algorithms on low-dose coronary CTA data are presented. The comparison shows that the method provided in this invention can effectively suppress noise, obtain higher-quality coronary CTA images, and is significantly superior to existing algorithms in terms of coronary artery edge preservation and structural detail restoration. Figure 8 The results of segmentation using the method provided by this invention and existing algorithms on low-dose coronary CTA data are presented. Clearly, the method provided by this invention achieves superior segmentation results on low-dose coronary CTA, identifying and segmenting more coronary artery branches with a very high similarity to real coronary artery branches. The above comparison results all demonstrate that the method provided by this invention has profound significance for scientific research in clinical medicine and related fields.

[0067] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A segmentation-guided low-dose coronary CTA noise reduction method, characterized in that, Includes the following steps: Step 1: Acquire DICOM image data after low-dose coronary CTA reconstruction from clinical scans; Step 2: Preprocess the DICOM image data obtained in Step 1 by setting the FOV range according to the DICOM header file and performing normalization. Step 3: Input the preprocessed data from Step 2 into the coarse noise reduction network to obtain a coarse noise reduction result with a lower noise level. Step 4: Feed the coarse noise reduction result from Step 3 into the segmentation network, so that the segmentation network can segment the coronary artery branches while avoiding some noise. Step 5: Using the segmentation results obtained in Step 4 as a guide for the coarse denoising results in Step 3, input both into the fine denoising network to obtain the final, more refined denoising results. The coarse denoising network in step 3, the segmentation network in step 4, and the fine denoising network in step 5 share the same network architecture, each consisting of an encoder and a decoder. The encoder has four encoder blocks to learn denoising features at different scales, with the feature map size decreasing and the number of channels increasing progressively with each encoder block. Correspondingly, the decoder also has four decoder blocks, with the feature map size increasing and the number of channels decreasing progressively with each decoder block. Adjacent encoder blocks are connected by sub-pixel downsampling convolutional layers, and adjacent decoder blocks are connected by sub-pixel upsampling convolutional layers. Sub-pixel convolutional layers achieve upsampling and downsampling by increasing or decreasing the number of channels and shuffling pixels. Each encoder and decoder consists of two residual channel attention modules and a convolutional layer cascaded together. Channel attention learns a weight vector to weight each channel, enabling the network to focus more on important feature channels. The residual mechanism allows the network to learn residual mappings, which helps solve gradient vanishing and surface appearance problems in deep networks. To address the bottleneck issue, residual connections are introduced, allowing the network to more easily learn the direct mapping relationship between input and output, thereby improving denoising performance. Skip connections also exist between the encoder and decoder blocks at the same level. A novel coordinate gating module is embedded in these skip connections to fuse denoising features from different scales of the encoder and decoder blocks at the same level. The coordinate gating module first performs 3×3 convolutions on the features from the encoder and decoder blocks, followed by average pooling in both the width and height dimensions to compress the scale. Then, the encoder and decoder features, pooled in the width dimension, are concatenated in the channel dimension, and a 1×1 convolution is used to reduce the number of channels to match the input feature dimension; the same applies to the height dimension. Next, the compressed features obtained in the width and height dimensions are multiplied pixel-by-pixel and then fused using a 3×3 convolution. Finally, the fused features are activated, multiplied by the encoder and decoder features respectively, and then summed to obtain the fused feature representation. The feature transfer process of the coordinate gating module is shown below: in, and These represent encoder features and decoder features, respectively. and These represent average pooling in the height and width dimensions, respectively. and These represent the encoder features after average pooling in the height and width dimensions, respectively. and These represent the decoder features after average pooling in the height and width dimensions, respectively. , , and These represent 1×1 convolutions. Indicates a connection operation. This represents the feature map to be activated. and These represent the feature maps after activation of encoder and decoder features, respectively. Indicates the characteristics after fusion; The segmentation network in step 4 includes a unique cross-task guided refinement module for interaction between the coarse denoising network and the segmentation network. This module consists of two parts: intra-task refinement and inter-task refinement. Intra-task refinement calculates an affinity matrix to filter and capture task similarity information to enhance contextual information. Inter-task refinement utilizes cross-attention and self-attention to refine and fuse cross-task features to achieve feature refinement. In step 5, the aggregation layer first concatenates the denoising features and high-level segmentation features along the channel dimension, and then feeds them sequentially into a 1×1 convolutional layer, a 3×3 deep convolutional layer, and a 1×1 convolutional layer for initial feature fusion. Next, the pre-fused features are added element-wise to the high-level denoising features and then concatenated with the segmentation features along the channel dimension. The purpose is to achieve full aggregation and refinement of denoising features through the guiding role of segmentation features in deep semantics. Finally, the concatenated features are further fused and the output channels are adjusted through a 3×3 convolution to obtain higher-level denoising features with richer dimension matching and stronger representational capabilities.

2. The segmentation-guided low-dose coronary CTA noise reduction method according to claim 1, characterized in that, In step 2, threshold filtering is used to preprocess the DICOM image data, and the FOV range is recorded.

3. The segmentation-guided low-dose coronary CTA noise reduction method according to claim 1, characterized in that, In step 2, the DICOM image data is normalized according to the slope and intercept of the data acquisition device.

4. The segmentation-guided low-dose coronary CTA noise reduction method according to claim 1, characterized in that, In step 4, within the intra-task refinement, the input features are processed by an affinity layer to calculate an affinity matrix. Since each row of the affinity matrix represents the correlation between a location and all other locations, the affinity matrix utilizes pixel-level similarity to achieve structure preservation for noise reduction and target region segmentation. It provides a prerequisite for the interaction between the two tasks. During the calculation of the affinity matrix, the input features are first flattened in the spatial dimension, then multiplied with their transposed matrix, and finally, each row of the product is normalized to obtain the affinity matrix. After obtaining the affinity matrix, the input features are weighted and aggregated into the enhanced response. In the inter-task refinement, cross-attention is first used to calculate the similarity between the query vector Q and the key vector K, guiding the weighting of their respective value vectors V to establish their correlation, thereby highlighting similarities. The features are then calculated by multiplying the affinity matrices of the two tasks element-wise to integrate complementary information between the tasks. Since each task contains unique features, self-attention is applied to each task separately under the guidance of the combined affinity matrix to establish relationships between global pixels, thus fully fusing the features of the two tasks. When calculating the query vector Q, key vector K, and value vector V, depthwise convolution can better replace linear projection for embedding features. Compared to linear projection, depthwise convolution preserves the spatial location information of pixels and reduces information loss, which is very beneficial for noise reduction and segmentation. Finally, the fused high-level features are sufficiently refined to enhance representational power and are passed to the segmentation network and the fine noise reduction network respectively to assist in segmentation and noise reduction. The feature transfer process of the cross-task guided refinement module is represented as follows: in, and Representing noise reduction features and segmentation features, Indicates the affinity layer. and Let these represent the denoising affinity matrix and the segmentation affinity matrix, respectively. This represents the Sigmoid activation function. and This represents the refined noise reduction and segmentation features within the task. This indicates multiple points of cross-attention. Indicates a connection operation. and Represents intermediate features that refine the relationships between tasks. This indicates the bulls' self-attention. and This represents the noise reduction and segmentation features obtained after cross-task-guided refinement.

Images