Coronary artery lipid spot recognition method based on machine learning

Abstract

The invention provides a coronary artery lipid spot recognition method based on machine learning, belongs to the field of image recognition, and aims to recognize a lipid spot structure region from a coronary artery image. Anisotropic structure features are extracted through an asymmetric differential operator of the feature fractal dimension analysis module, and rotation invariant feature expression is established through a directional aggregation signature of the structure collaborative recognition module, so that the problem of insufficient recognition precision of non-uniform lipid spots in a complex environment by a traditional method is effectively solved. And reliable technical support is provided for coronary heart disease screening and accurate auxiliary diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, and specifically relates to a method for recognizing coronary artery lipid plaques based on machine learning. Background Technology

[0002] Traditional methods for identifying lipid plaques in coronary arteries mainly rely on manual image reading or basic image processing algorithms, which have significant limitations. Under complex imaging conditions, traditional methods struggle to effectively extract the spectral and texture features of lipid plaques within the vessel wall. Furthermore, traditional algorithms are poorly adaptable to non-ideal clinical scenarios such as changes in vessel curvature angles and local calcification obscuring the plaques. Moreover, the coronary artery wall and surrounding tissues have highly similar brightness and texture, making it difficult for traditional methods to accurately separate lipid plaque areas from normal vessel wall structures, severely impacting the efficiency of lesion screening and risk assessment.

[0003] While current deep learning-based solutions have improved the recognition rate of some clear images, they still have significant shortcomings in actual clinical imaging. Most algorithms directly process single-modal images without optimizing for the spectral attenuation characteristics within the blood vessel lumen, leading to signal differences caused by different illumination intensities and imaging depths, which reduces recognition stability. Secondly, existing feature extraction methods mix vascular structural features and lipid plaque texture features, resulting in a high false detection rate in scenarios with similar tissue types. Most importantly, existing methods lack a unified representation capability for lipid plaques under multi-angle imaging, making it difficult to infer the overall lesion distribution from local fragment features.

[0004] This invention proposes a machine learning-based method for identifying coronary artery lipid plaques. It achieves stable extraction of lipid plaque features under complex lighting and imaging conditions through a spectral spatial enhancement module, solving the light sensitivity problem of traditional methods. A feature fractal analysis module separates and comprehensively enhances vascular structure and lipid texture, significantly improving the geometric expression of key lesion areas. Finally, a structural collaborative recognition module integrates multi-angle imaging features to achieve robust recognition in multi-angle scenes. This method can efficiently and accurately identify coronary artery lipid plaques, significantly improving its ability to cope with complex tissue structures and imaging interference, providing a precise and reliable lesion detection method for clinical diagnosis of coronary heart disease. Summary of the Invention

[0005] This invention provides a machine learning-based method for identifying coronary artery lipid plaques, which aims to achieve the identification of coronary artery lipid plaques through the synergistic effect of a spectral spatial joint enhancement module, a feature fractal dimension analysis module, and a structural collaborative identification module.

[0006] This invention aims to propose a coronary artery lipid plaque recognition model and provide a machine learning-based method for coronary artery lipid plaque recognition, comprising the following steps:

[0007] S1. Acquire raw medical images of coronary artery lipid plaques, and preprocess the raw images by converting the format and uniformly cropping to generate a coronary artery lipid plaque image dataset.

[0008] S2. A high-precision color space transformation is proposed, and a spectral space joint enhancement module is constructed. The coronary artery lipid plaque image is transformed into a multi-phase feature fusion enhanced image through the spectral space joint enhancement module.

[0009] S3. Design an asymmetric differential operator and construct a feature fractal analysis module. The multi-phase feature fusion enhancement image is obtained by the feature fractal analysis module to obtain a decoupled enhanced feature map.

[0010] S4. A structural cooperative operator and a directional aggregation signature strategy are proposed to construct a structural cooperative recognition module. The coronary artery lipid plaque image and the decoupled enhanced feature map are used to obtain the lipid plaque recognition result through the structural cooperative recognition module.

[0011] S5 integrates a spectral spatial joint enhancement module, a feature fractal dimension analysis module, and a structural collaborative recognition module to construct a coronary artery lipid plaque recognition model;

[0012] S6. Train the coronary artery lipid plaque recognition model. Input the coronary artery lipid plaque image to be recognized into the trained coronary artery lipid plaque recognition model for recognition, and output the lipid plaque recognition region in the coronary artery lipid plaque image.

[0013] Preferably, in S1, constructing the coronary artery lipid plaque image dataset specifically includes the following steps: using clinical imaging equipment to acquire multiple sections and angles of the coronary artery lumen and vessel wall to obtain raw medical image data including coronary artery lipid plaques; converting the acquired raw medical image data into PNG format and uniformly cropping it to a size of 768×432 pixels.

[0014] Preferably, coronary artery lipid plaque images acquired under complex clinical conditions are susceptible to changes in tissue reflectance intensity, imaging depth attenuation, and scattering noise, leading to local distortion of lipid plaque structural features. In addition, the expression of frequency domain features and spatial domain features is fragmented, making it difficult to coordinate the optimization of fine plaque boundaries and the overall vessel wall contour. Furthermore, there are significant differences in tissue optical characteristics between different imaging devices and different patients, making it difficult for a single processing mode to meet the lipid plaque recognition needs of all scenarios. Therefore, a spectral-spatial joint enhancement module is proposed. A unified tissue representation space is established through a high-precision color space transformation strategy to reduce device differences. A spectral-spatial coupling filter is used to achieve a balance between structural detail enhancement and noise suppression. Finally, the optimal features of different tissue reflectance phases are adaptively integrated through a multi-phase feature fusion strategy, thereby significantly improving the robustness and discriminativeness of lipid plaque feature expression.

[0015] Preferably, in S2, constructing the spectral spatial joint enhancement module specifically includes the following steps:

[0016] A color space transformation strategy is constructed based on the difference between the RGB channels of the image and the global color mean. Different transformation rules are adopted according to whether the difference exceeds a set threshold. The threshold size is determined by the global color mean of the image and its proportional relationship, so as to generate a coronary artery fatty plaque image after color space transformation.

[0017] The horizontal and vertical components of each pixel in the coronary artery fatty plaque image after color space transformation are determined. An adaptive standard deviation is constructed based on these components, and a spatially sensitive filter kernel is generated. Using the spatially sensitive filter kernel, combined with a frequency domain set containing low, medium, and high frequencies, a coupled filter feature is obtained through tensor product operation.

[0018] Based on the main diagonal length and phase offset of the coronary artery lipid plaque recognition image after color space transformation, a phase feature basis for the corresponding phase is constructed. Based on the phase feature basis, corresponding phase weights are introduced, and the Hadamard product is performed on the coupled filter features to sequentially complete the feature response fusion under each phase. The fusion results of all phases are then weighted and integrated to output a multi-phase feature fusion enhanced image.

[0019] Preferably, the spectral spatial joint enhancement module constructed through S2 uses an adaptive color space transformation strategy based on the difference between the RGB channels and the global color mean. This allows the image to maintain a more stable spectral expression under changes in illumination and background, improving the color discrimination of lipid plaque regions. An adaptive standard deviation constructed using pixel coordinates forms a spatially sensitive filter kernel, which is then used to perform tensor product operations with a frequency domain set covering low, mid, and high frequencies. This simultaneously enhances the image's spatial structure and spectral details, improving the resolution of plaque edges, fine textures, and local structural differences. Furthermore, a phase feature base is constructed using the image's main diagonal length and phase offset, and multi-phase fusion is performed, further enhancing the directional and hierarchical structure. Overall, this module achieves synergistic optimization in spectral consistency, spatial detail expression, and phase structure enhancement, providing clearer and more discriminative feature inputs for the lipid plaque recognition model, improving recognition accuracy and stability.

[0020] Preferably, although the enhanced feature map output by the spectral spatial joint enhancement module constructed by S2 effectively overcomes the problems of insufficient response to weak contrast patch regions, detail breakage caused by frequency domain filtering, and boundary discontinuity caused by multi-directional texture interference in traditional image enhancement methods, the feature map may still have problems such as unbalanced local structural expression, inconsistent cross-scale detail response, and false enhancement of some high-frequency features in complex backgrounds. Therefore, this invention proposes a feature fractal analysis module, which extracts multi-directional geometric features through asymmetric differential operators, uses structure-texture manifold decomposition to achieve accurate separation of identity features and apparent texture, and finally enhances the discriminative region response through adaptive feature recombination, thereby constructing a lipid patch feature expression with pose robustness.

[0021] Preferably, in S3, constructing the feature fractal dimension parsing module specifically includes the following steps:

[0022] The L2 norm and infinite norm are calculated for the multi-phase feature fusion enhancement image to generate adaptive weights corresponding to each pixel position. Based on the adaptive weights, the positive asymmetric differential calculation component of the enhancement image is extracted in the main diagonal direction, and the negative asymmetric differential calculation component of the enhancement image is extracted in the sub-diagonal direction.

[0023] Based on the positive and negative asymmetric differential computation components, their Frobenius norms over the entire image are calculated respectively; a decomposition angle is constructed according to the ratio of the two norms; based on the decomposition angle, the positive and negative asymmetric differential computation components are rotated and combined to construct structural components and texture components.

[0024] Nonlinear mapping is performed on the texture components, and a structure-texture balance coefficient is introduced to jointly enhance the structure and texture in the image region, outputting a decoupled enhanced feature map.

[0025] Preferably, the feature fractal analysis module constructed by S3 uses adaptive weights based on the L2 norm and the infinite norm to achieve differentiated adjustment of different image regions, and uses asymmetric differential components in the main diagonal and sub-diagonal directions to obtain more direction-sensitive differential features; at the same time, it constructs decomposition angles based on the Frobenius norms of the positive and negative asymmetric differential components and performs rotational combination, so that structural trends and texture perturbations are effectively separated in terms of components, enhancing the clarity of patch contours, regional gradients and fine textures; furthermore, it performs nonlinear mapping on the texture components and introduces a structure-texture balance coefficient, so that the image can highlight key texture details while maintaining structural integrity.

[0026] Preferably, although the decoupled enhanced feature map output by the feature fractal analysis module constructed by S3 achieves effective separation of vascular structure and lipid tissue texture, there is still a problem that the geometric structure correlation is not fully modeled in lesion regions with different lipid content levels. This results in a lack of synergistic constraints between the global distribution pattern and local texture changes of the three types of lipid plaques. Furthermore, the anisotropic features caused by lipids in the decoupled enhanced feature map do not form a unified structural expression. Therefore, this invention introduces a structural collaborative recognition module, which encodes the global structural distribution features of lipid plaques through structural collaborative operators and directional aggregation signature strategies. Finally, it achieves highly robust classification and recognition of the three types of lipid plaques through standardized signatures. Overall, this module forms a synergistic enhancement in terms of weight adaptability, directional sensitive difference, and structure-texture decomposition, providing a clearer and more discriminative feature representation for coronary artery lipid plaque recognition, and improving the model's accuracy and stability in recognizing complex morphologies and weakly textured plaques.

[0027] Preferably, in S4, constructing the structural collaborative recognition module specifically includes the following steps:

[0028] Based on the horizontal and vertical gradients of coronary artery lipid plaque images, image gradient direction information is obtained; the gradient direction information and the difference between the decoupled enhancement feature map are weighted and combined to construct a structural cooperative operator; based on the structural cooperative operator, the residual information between the original image brightness and the decoupled enhancement feature map is structurally adjusted to extract a joint feature map;

[0029] Using the radius of the inscribed circle of the joint feature map as the integration range, the joint feature map is converted into a polar coordinate representation; within the integration range, the polar coordinate joint feature map is weighted and integrated based on the radial attenuation factor to construct a global structural signature;

[0030] Calculate the feature mean and feature variance of the global structural signature; based on the feature mean and feature variance, standardize the global structural signature, propose a saliency feature projection, project the standardized global structural signature onto the image space to generate a saliency heatmap; based on the saliency heatmap, construct an adaptive recognition strategy, determine the recognition mask for the target lipid spot region, and obtain the final lipid spot recognition result.

[0031] Preferably, the structural collaborative recognition module constructed by S4 obtains a joint feature map of the original coronary artery image and the decoupled enhanced feature map through the structural collaborative operator, laying the foundation for obtaining the directional aggregation signature in the next step. The directional aggregation signature strategy encodes the distribution pattern of coronary artery lipid plaques on the vessel wall into rotation-invariant feature vectors through polar coordinate integration, which efficiently models the anisotropic geometric features and global distribution morphology of lipid plaques in complex vascular structures. By standardizing the global structural signature and combining it with a fast similarity calculation strategy, the recognition efficiency is significantly improved while improving the accuracy of lipid plaque recognition.

[0032] Preferably, in S5, constructing the coronary artery lipid plaque image recognition model specifically includes the following steps:

[0033] Step S51: The original coronary artery lipid plaque image is processed by a spectral spatial joint enhancement module to obtain a multi-phase feature fusion enhanced image. The mathematical model is as follows:

[0034] ;

[0035] in, This represents an image enhanced by multi-phase feature fusion. This indicates the spectral spatial joint enhancement module. This represents an image of the original coronary artery lipid plaque.

[0036] Step S52: The multi-phase feature fusion enhancement image is processed by the feature fractal analysis module to obtain the decoupled enhanced feature map. The mathematical model is as follows:

[0037] ;

[0038] in, This represents the decoupling enhancement feature map. This represents the feature fractal dimension parsing module.

[0039] Step S53: The original coronary artery lipid plaque image and the decoupled enhanced feature map are processed by the structural collaborative recognition module to obtain the final recognition result. The mathematical model is as follows:

[0040] ;

[0041] in, This indicates the final results of coronary artery lipid plaque identification. This indicates the structure collaboration recognition module.

[0042] In summary, compared with existing technologies, the beneficial effects of this invention are as follows: The spectral spatial joint enhancement module, through high-precision color space transformation and frequency-space coupling filtering, enables the model to effectively address feature distortion caused by tissue reflection differences, imaging depth attenuation, and scattering noise, significantly improving the integrity and feature representation of lipid plaque features; the feature fractal dimension analysis module, through asymmetric differential operators and structure-texture manifold decomposition, significantly improves the geometric integrity of lipid plaque boundaries and lesion contours and the discriminative power of lipid plaque representation, reducing the model's false recognition rate; the structural collaborative recognition module obtains a joint feature map through structural collaborative operators, and encodes the joint feature map into directional aggregation signature features through a directional aggregation signature strategy, realizing efficient modeling of the global distribution pattern of anisotropic lipid plaque structures, and improving the model's recognition efficiency through rapid similarity calculation; the three modules work together to form a coronary artery lipid plaque recognition method based on a fractal structure fusion strategy, making the model adaptable to imaging conditions, feature discriminative, and robust to multiple angles, significantly improving the recognition accuracy, robustness, and efficiency of coronary artery lipid plaque images under complex clinical imaging interference. Attached Figure Description

[0043] Figure 1 This is a step-by-step diagram of a machine learning-based method for identifying lipid plaques in coronary arteries.

[0044] Figure 2 This is a structural diagram of the spectral spatial joint enhancement module.

[0045] Figure 3 This is a structural diagram of the feature fractal analysis module.

[0046] Figure 4 This is a structural diagram of the structural collaborative identification module.

[0047] Figure 5 This is a diagram showing the overall structure of the coronary artery lipid plaque image recognition model.

[0048] Figure 6 A schematic diagram for identifying lipid plaques in the anterior and posterior coronary arteries. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a method for identifying coronary artery lipid plaques based on machine learning.

[0051] S1. Acquire raw medical images of coronary artery lipid plaques, and preprocess the raw images by converting the format and uniformly cropping to generate a coronary artery lipid plaque image dataset.

[0052] Furthermore, as shown in the appendix Figure 1 The dataset of coronary artery lipid plaque images described in S1 uses clinical imaging equipment to acquire images of the coronary artery lumen and vessel wall from multiple sections and angles, obtaining raw medical image data including coronary artery lipid plaques. The acquired raw medical image data is converted into PNG format and then uniformly cropped to a size of 768×432 pixels.

[0053] S2. A high-precision color space transformation is proposed, and a spectral space joint enhancement module is constructed. The coronary artery lipid plaque image is transformed into a multi-phase feature fusion enhanced image through the spectral space joint enhancement module.

[0054] Furthermore, as shown in the appendix Figure 1 The construction steps of the spectral spatial joint enhancement module described in S2 are attached. Figure 2 As shown, the specific implementation of the module includes the following steps:

[0055] Further, in step S21, a high-precision color space transformation is proposed, and a high-precision color space transformation strategy is constructed. The mathematical model is as follows:

[0056] ;

[0057] in, Image showing lipid plaques in the coronary arteries. This represents an image of coronary artery lipid plaques after color space transformation. Represents the RGB three-channel components of a coronary artery lipid plaque image. Indicates the color space transformation threshold. , The mathematical model for representing the global color mean of an image is:

[0058] ;

[0059] in, Indicates the total number of pixels. Indicates pixel index, They represent the first The red, green, and blue pixel values ​​of each pixel. This indicates the matrix transpose.

[0060] Further, in step S22, a spectral-spatial coupled filter is designed to transform the color space of the coronary artery lipid plaque image. Filtering is performed to obtain the coupled filter characteristics. The mathematical model is as follows:

[0061] ;

[0062] in, Indicates the characteristics of coupled filtering. Represents the frequency domain. This represents a set of frequency domains, including low-frequency, mid-frequency, and high-frequency domains. This represents the tensor product operation. The spatially sensitive filter kernel is represented by the following mathematical model:

[0063] ;

[0064] in, Indicates the center coordinates of the space-sensitive filter kernel. The mathematical model for adaptive standard deviation is:

[0065] ;

[0066] The squared adaptive standard deviation ensures that the response decays non-linearly with increasing distance, enhancing the non-linear sensitivity to distance changes. This allows the filter kernel to decay rapidly when it is far from the center, thus forming a location-dependent local response region.

[0067] Furthermore, step S23 is based on the coupling filter features A multi-phase feature fusion strategy is constructed to output a multi-phase feature fusion enhanced image. The mathematical model is as follows:

[0068] ;

[0069] in, This represents an image enhanced by multi-phase feature fusion. Indicates phase weight, , This indicates the Hadamard product. Indicates phase The phase characteristic basis, mathematical model is:

[0070] ;

[0071] in, Indicates the length of the main diagonal of the image. Indicates phase The phase offset.

[0072] S3. Design an asymmetric differential operator and construct a feature fractal analysis module. The multi-phase feature fusion enhancement image is obtained by the feature fractal analysis module to obtain a decoupled enhanced feature map.

[0073] Furthermore, as shown in the appendix Figure 1 The feature fractal analysis module described in S3 is constructed using the following steps: Figure 3 As shown, the specific implementation of the module includes the following steps:

[0074] Further, in step S31, an asymmetric differential operator is designed to perform asymmetric differential operations on the multi-phase feature fusion enhancement image obtained by the spectral spatial joint enhancement module described in S2, resulting in positive and negative asymmetric differential calculation components. The mathematical model is as follows:

[0075] ;

[0076] ;

[0077] in, This represents the positive asymmetric differential calculation component along the main diagonal direction of the multi-phase feature fusion enhanced image. This represents the negative asymmetric differential calculation component along the sub-diagonal direction of the multi-phase feature fusion enhanced image. This represents an image enhanced by multi-phase feature fusion. Indicates the differential step size. The adaptive weights are represented by the following mathematical model:

[0078] ;

[0079] in, Represents the L2 norm. It represents the infinite norm.

[0080] Furthermore, step S32 proposes a structure-texture manifold decomposition algorithm to calculate the components of the positive asymmetric differential obtained in S31. and negative asymmetric differential calculation components Perform structure-texture manifold decomposition to obtain structure components and texture components. The mathematical model is as follows:

[0081] ;

[0082] in, Represents structural components, Represents texture components, To represent the decomposition angle, the mathematical model is:

[0083] ;

[0084] in, express Norm, mathematical model is:

[0085] ;

[0086] ;

[0087] in, and These represent the positions in the positive and negative asymmetric differential component matrices, respectively. The element value at that location.

[0088] Furthermore, step S33 is based on the structural components obtained in S32. and texture components Adaptive feature recombination is performed to generate decoupled enhanced feature maps. The mathematical model is as follows:

[0089] ;

[0090] in, This represents the decoupling enhancement feature map. Represents the structure-texture balance coefficient. , This represents the hyperbolic tangent nonlinear activation function, used to apply a smooth saturation mapping to the texture components, so that while maintaining the trend of texture changes, it suppresses excessive response and avoids excessive amplification of local textures.

[0091] S4. A structural cooperative operator and a directional aggregation signature strategy are proposed to construct a structural cooperative recognition module. The coronary artery lipid plaque image and the decoupled enhanced feature map are used to obtain the lipid plaque recognition result through the structural cooperative recognition module.

[0092] Furthermore, as shown in the appendix Figure 1 The structural collaborative recognition module described in S4 is constructed using the following steps: Figure 4 As shown, the specific implementation of the module includes the following steps:

[0093] Further, in step S41, a structural cooperative operator is constructed, and a joint feature map is extracted. The mathematical model is as follows:

[0094] ;

[0095] in, Represents the joint feature map. Represents structural cooperative operators, and They represent Gradient in the horizontal direction and gradient in the vertical direction and They represent Gradient in the horizontal direction and gradient in the vertical direction.

[0096] Furthermore, step S42 will combine the feature maps. Converting to polar coordinates, designing a direction aggregation signature strategy to generate a global structure signature, the mathematical model is as follows:

[0097] ;

[0098] in, Indicates the global structure signature. The radius of the inscribed circle of the joint feature map is represented. The polar coordinate representation of the joint feature map. This represents the polar radius from the center of the image outwards. Represents polar coordinate angles. , This represents the radial attenuation control parameter. .

[0099] Further, step S43, signing the global structure. After standardization, a standardized global structural signature is obtained. The mathematical model is as follows:

[0100] ;

[0101] in, Represents a standardized global structure signature. The feature mean of the global structural signature. The square of the characteristic standard deviation of the global structural signature;

[0102] We propose a salient feature projection method to project the global signature back into the image space. The mathematical model is as follows:

[0103] ;

[0104] in, A heatmap indicating significance;

[0105] Adaptive identification was used to identify lipid plaque regions in the coronary arteries. The mathematical model is as follows:

[0106] ;

[0107] in, This represents the mask for the final recognition result. These represent the height and width of the saliency heatmap, respectively. .

[0108] S5 integrates a spectral spatial joint enhancement module, a feature fractal dimension analysis module, and a structural collaborative recognition module to construct a coronary artery lipid plaque image recognition model.

[0109] Furthermore, as shown in the appendix Figure 1The specific steps for constructing the coronary artery lipid plaque image recognition model described in S5 are attached. Figure 5 As shown, the specific implementation of the model includes the following steps:

[0110] Further, in step S51, the original coronary artery lipid plaque image is processed by a spectral spatial joint enhancement module to obtain a multi-phase feature fusion enhanced image. The mathematical model is as follows:

[0111] ;

[0112] in, This represents an image enhanced by multi-phase feature fusion. This indicates the spectral spatial joint enhancement module. This represents an image of the original coronary artery lipid plaque.

[0113] Furthermore, the S52 multi-phase feature fusion enhanced image is processed by the feature fractal analysis module to obtain a decoupled enhanced feature map. The mathematical model is as follows:

[0114] ;

[0115] in, This represents the decoupling enhancement feature map. This represents the feature fractal dimension parsing module.

[0116] Furthermore, S53, the original coronary artery lipid plaque image, and the decoupled enhanced feature map are processed through a structural collaborative recognition module to obtain the final coronary artery lipid plaque recognition result. The mathematical model is as follows:

[0117] ;

[0118] in, This indicates the final results of coronary artery lipid plaque identification. This indicates the structure collaboration recognition module.

[0119] S6. Train the coronary artery lipid plaque image recognition model. Input the coronary artery lipid plaque image to be recognized into the trained coronary artery lipid plaque image recognition model for recognition, and output the recognition result of the coronary artery lipid plaque image.

[0120] Furthermore, as shown in the appendix Figure 1 The coronary artery lipid plaque image recognition model described in S6 takes a coronary artery lipid plaque image as input, as shown in the attached image. Figure 6 As shown in (a), the recognition results in the coronary artery lipid plaque image are output, as attached. Figure 6 As shown in (b), the specific implementation includes the following steps:

[0121] Furthermore, in step S6, the operating system platform used by the model is CentOS, the language is Python 3.10.14, the processor is Jetson Xavier, the image processing libraries used are OpenCV and PIL, the server physical memory is 64G, and the dataset includes raw medical image data with coronary artery lipid plaques, totaling 1200 images.

[0122] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for identifying coronary artery lipid plaques based on machine learning, characterized in that, include: Raw images of coronary arteries were acquired, and the raw images were preprocessed by uniform cropping to construct a coronary artery lipid plaque image dataset. A high-precision color space transformation is proposed, and a spectral space joint enhancement module is constructed. Based on the spectral space joint enhancement module, the coronary artery lipid plaque image is processed to output a multi-phase feature fusion enhanced image. Based on the asymmetric differential operator, a feature fractal analysis module is constructed so that the multi-phase feature fusion enhanced image is obtained as a decoupled enhanced feature map through the feature fractal analysis module. Design a structural cooperative operator and a directional aggregation signature strategy, and construct a structural cooperative recognition module so that the coronary artery lipid plaque image and the decoupled enhanced feature map can be processed by the structural cooperative recognition module to obtain the lipid plaque recognition result; The spectral spatial joint enhancement module, the feature fractal dimension analysis module, and the structural collaborative recognition module are integrated into a coronary artery lipid plaque recognition model; The coronary artery lipid plaque recognition model is trained by inputting the image of the coronary artery lipid plaque to be recognized into the trained recognition model for inference, and outputting the lipid plaque recognition region of the coronary artery lipid plaque image.

2. The method for identifying coronary artery lipid plaques based on machine learning according to claim 1, characterized in that, A color space transformation strategy is constructed based on the difference between the RGB channels of the image and the global color mean. Different transformation rules are adopted according to whether the difference exceeds a set threshold. The threshold size is determined by the global color mean of the image and its proportional relationship, so as to generate a coronary artery fatty plaque image after color space transformation.

3. The method for identifying coronary artery lipid plaques based on machine learning according to claim 2, characterized in that, The horizontal and vertical components of the coronary artery fatty plaque image are determined based on the coordinate values ​​of each pixel position after color space transformation. An adaptive standard deviation is constructed based on the components, and a spatially sensitive filter kernel is generated. Using the spatially sensitive filter kernel, combined with a frequency domain set containing low, medium, and high frequencies, a coupled filter feature is obtained through tensor product operation.

4. The method for identifying coronary artery lipid plaques based on machine learning according to claim 3, characterized in that, Based on the main diagonal length and phase offset of the coronary artery lipid plaque recognition image after color space transformation, a phase feature basis for the corresponding phase is constructed. Based on the phase feature basis, corresponding phase weights are introduced, and the Hadamard product is performed on the coupled filter features to sequentially complete the feature response fusion under each phase. The fusion results of all phases are then weighted and integrated to output a multi-phase feature fusion enhanced image.

5. The method for identifying coronary artery lipid plaques based on machine learning according to claim 4, characterized in that, The L2 norm and infinite norm are calculated for the multi-phase feature fusion enhancement image to generate adaptive weights corresponding to each pixel position. Based on the adaptive weights, the positive asymmetric differential calculation component of the enhancement image is extracted in the main diagonal direction, and the negative asymmetric differential calculation component of the enhancement image is extracted in the sub-diagonal direction.

6. The method for identifying coronary artery lipid plaques based on machine learning according to claim 5, characterized in that, Based on the positive and negative asymmetric differential calculation components, their Frobenius norms over the entire graph are calculated respectively; the decomposition angle is constructed according to the ratio of the two norms. Based on the decomposition angle, the positive and negative asymmetric differential calculation components are rotated and combined to construct structural components and texture components. Nonlinear mapping is performed on the texture components, and a structure-texture balance coefficient is introduced to jointly enhance the structure and texture in the image region, outputting a decoupled enhanced feature map.

7. The method for identifying coronary artery lipid plaques based on machine learning according to claim 6, characterized in that, Based on the horizontal and vertical gradients of coronary artery lipid plaque images, image gradient direction information is obtained; the gradient direction information is weighted and combined with the differences between the decoupled enhancement feature map to construct a structural cooperative operator; Based on the aforementioned structural cooperative operator, structural adjustment is performed on the residual information between the original image brightness and the decoupled enhanced feature map to extract the joint feature map.

8. The method for identifying coronary artery lipid plaques based on machine learning according to claim 7, characterized in that, Using the radius of the inscribed circle of the joint feature map as the integration range, the joint feature map is converted into a polar coordinate representation. Within the integration range, the polar coordinate joint feature map is weighted and integrated based on the radial attenuation factor to construct a global structural signature.

9. The method for identifying coronary artery lipid plaques based on machine learning according to claim 8, characterized in that, Calculate the feature mean and feature variance of the global structural signature; based on the feature mean and feature variance, standardize the global structural signature, propose a saliency feature projection, project the standardized global structural signature onto the image space, and generate a saliency heatmap; based on the saliency heatmap, construct an adaptive recognition strategy and determine the recognition mask for the target lipid spot region.

10. The method for identifying coronary artery lipid plaques based on machine learning according to claim 9, characterized in that, The coronary artery lipid plaque image is processed by the spectral spatial joint enhancement module to obtain a multi-phase feature fusion enhanced image; the multi-phase feature fusion enhanced image is processed by the feature fractal dimension analysis module to obtain a decoupled enhanced feature map; the coronary artery lipid plaque image and the decoupled enhanced feature map are processed by the structural collaborative recognition module to obtain the final lipid plaque recognition result.

Images