A carotid artery plaque infrared polarization image enhancement processing method and storage medium

Abstract

The application provides a carotid artery plaque infrared polarization image enhancement processing method and a storage medium, and relates to the technical field of image enhancement; the method comprises the following steps: acquiring an infrared polarization single-channel gray image, calculating an edge image and a gamma enhancement image to construct three-channel input, calculating the weight of each channel image and performing normalized fusion to obtain a fused image, extracting an initial feature map, constructing a structure-guided gray image, calculating a local density response and a direction gradient response, forming an enhanced feature map through restoration mapping and dynamic amplitude control, extracting a local spatial variation and a frequency residual signal based on the enhanced feature map, constructing a dynamic adjustment factor to perform amplitude response adjustment, and outputting a final feature map; and fusing and normalizing the final feature maps of the channels to generate an enhanced image. The application can significantly improve the expression ability of plaque structure details and contrast of the infrared polarization image.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement, specifically relating to a method for enhancing and storing infrared polarization images of carotid artery plaques. Background Technology

[0002] Infrared polarization image enhancement of carotid artery plaques is used to improve the optical contrast and structural boundary discernibility of plaque tissue. Current infrared polarization imaging procedures for determining the nature of carotid atherosclerotic plaques typically rely on two aspects: First, it is necessary to achieve high spatial resolution and high stability multi-polarization angle acquisition in the imaging window on the neck surface to suppress polarization distortion caused by surface specular reflection and random scattering; second, it is necessary to construct an image enhancement model in the post-processing stage to perform contrast remapping and boundary sharpening, as well as image grayscale distribution optimization, on details of the fibrous cap, lipid core, and local calcification areas, thereby supporting the determination of plaque composition and stability.

[0003] Most existing image enhancement models adjust the grayscale of images globally or locally through methods such as linear stretching, histogram equalization, gamma transform, and Retinex algorithm. Although these methods improve the visual quality of images to some extent, they still have the following shortcomings: On the one hand, the enhancement parameters are often fixed values ​​or manually set, lacking adaptability to different image scenes; on the other hand, noise amplification effects or excessive enhancement of local information may be introduced during the contrast enhancement process, resulting in distortion of image structural information. Summary of the Invention

[0004] To address the problems mentioned above in the background art, this invention designs a method and medium for enhancing infrared polarization images of carotid artery plaques. For the task of enhancing infrared polarization images of carotid artery plaques, it optimizes the image grayscale distribution, achieving an adjustment method that optimizes the image grayscale distribution while preserving image details. This improves the clarity and recognizability of supporting plaque composition and stability determination, and enhances the stability and accuracy of subsequent diagnostic tasks.

[0005] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps: A method for enhancing infrared polarization images of carotid artery plaques, comprising: Step 1: Obtain a single-channel grayscale image of the carotid artery plaque with infrared polarization as input, and calculate the edge map and gamma enhancement map based on the grayscale image to construct a three-channel input image; Step 2: Calculate the channel importance weight of each channel image, normalize the channel importance weights, and perform weighted fusion of the three channels of the input image to generate a single-channel fused image. Perform a convolution operation on the fused image to obtain the initial mapping value of the feature space. Step 3: Perform non-normalized density-driven weighted projection on the initial mapping values ​​of the feature space, fuse them to generate a single-channel structure-guided grayscale image, and construct the local density response tensor of the image based on the local pixel intensity differences of the structure-guided grayscale image; Step 4: Calculate the horizontal and vertical gradients of the guided grayscale image to construct the directional gradient intensity response. Remap the directional gradient intensity response and the local density response tensor to the original feature space through channel restoration mapping. Introduce a dynamic amplitude control factor to obtain an enhanced feature map. Step 5: Based on the enhanced feature map, calculate the local spatial change rate at each position of each channel, perform a two-dimensional Fourier transform on the enhanced feature map to construct a frequency residual signal, combine the change information of the spatial and frequency parts to construct a dynamic adjustment factor, adjust the response amplitude of the enhanced feature map, and output the final feature map. Step 6: Weighted fusion of the final feature maps of each channel and normalization through an activation function to optimize the grayscale distribution of the infrared polarization image of carotid artery plaques.

[0006] Preferably, the infrared polarized color image of the carotid artery plaque is converted into a single-channel grayscale image. The conversion is based on a weighted brightness mapping model to perform a weighted fusion operation on multiple color channels of the image to obtain the grayscale image. An edge response map is calculated based on the grayscale image. The edge response map is obtained by extracting and fusing the grayscale image gradients in the horizontal and vertical directions. The edge response map is used to distinguish the boundary position between carotid artery plaques and surrounding normal tissues. The grayscale image is enhanced by nonlinear mapping, and the enhanced image that highlights the local contrast of the image is output. The grayscale image, the edge response map and the enhanced image are stitched together in the channel dimension to construct a three-channel input image of the carotid artery plaque with infrared polarization after structural enhancement.

[0007] Preferably, based on the three-channel input image of the carotid artery plaque infrared polarization, the absolute average response value of each pixel in each channel is calculated, and the absolute average response value is used as the importance factor of each channel in the overall image. The importance factors of each channel are normalized to obtain the corresponding channel importance weight. The three-channel input image of the carotid artery plaque with infrared polarization is fused with the corresponding channel importance weights to generate a single-channel fused infrared polarization image of the carotid artery plaque. The fused image is then processed by a two-dimensional convolution operator to extract local gray-level structural features. The extracted local features are then subjected to channel dimensionality upscaling, and a tensor containing multiple feature channels is output as the initial feature space mapping value.

[0008] Preferably, based on the initial feature space mapping value, the feature space mapping value of each channel is weighted with the channel weight, and all weighted projections are added point by point at the pixel position to obtain a single-channel structure guided grayscale image; Based on the single-channel structure, an arbitrary target pixel in the grayscale image is used as the center position to establish a neighborhood range. For each pixel within the neighborhood range, its grayscale value is obtained, and the grayscale difference value between the grayscale value and the center pixel is calculated. A density response factor is constructed based on the grayscale difference value and a preset control parameter. The density response factor is weighted by an exponential function to calculate the grayscale difference. The density response factors of all neighboring pixels are superimposed to obtain the local density response tensor corresponding to the current target pixel.

[0009] Preferably, gradient calculations are performed on the single-channel structure-guided grayscale image in both the horizontal and vertical directions to calculate the grayscale change amplitude of each pixel in different directions, thus constructing the directional gradient intensity response of the pixel; the directional gradient intensity response is jointly mapped with the local density response tensor of the corresponding pixel, and a combined feature reflecting the local density change and structural directional intensity is formed based on the fusion relationship between the two; a one-to-one channel convolution operation is performed on the fused result to generate an enhanced feature map of the image.

[0010] Preferably, based on any pixel in each channel of the enhanced feature map, a local neighborhood region is constructed with that pixel as the center, the difference between the current pixel value and the average value of all pixels in the neighborhood region is calculated, and the difference is used to construct the local spatial change rate of the current pixel in the enhanced feature map; Perform frequency domain transformation on each channel of the enhanced feature map to obtain the spectral form, then use a smoothing filter function to construct a fuzzy spectrum map, and calculate the difference amplitude between the original spectrum map and the fuzzy spectrum map to construct a frequency residual signal. The median of the local spatial rate of change and the frequency residual signal is calculated within the local neighborhood. The response offset is calculated based on the difference between the current pixel and the corresponding median. The variance of the local spatial rate of change and the frequency residual signal is calculated within the local neighborhood to obtain a normalization factor. The response offset and the normalization factor are combined using a nonlinear function to construct a dynamic amplitude adjustment factor. The enhanced feature map and the dynamic amplitude adjustment factor are multiplied point by point along the pixel dimension to obtain the final feature map.

[0011] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, and stores an image enhancement program on the storage medium, wherein the carotid artery plaque infrared polarization image enhancement program can implement the carotid artery plaque infrared polarization image enhancement steps described above.

[0012] Compared with the prior art, the present invention has the following technical effects: The carotid artery plaque infrared polarization image enhancement processing method proposed in the present invention integrates local structural information and frequency residual signals to construct a space-frequency joint dynamic amplitude adjustment mechanism, which can effectively improve the gray-scale distribution contrast and boundary clarity of the plaque area while maintaining the details and edge structure integrity of the carotid artery plaque infrared polarization image. Secondly, by constructing a synergistic adjustment factor between the local spatial rate of change and frequency domain features through non-normalized density-driven response, the problems of excessive edge enhancement, noise amplification, and structural information distortion in traditional global enhancement methods are avoided. This method has higher image adaptability and diagnostic usability, and significantly improves the accuracy and robustness of subsequent patch component identification and stability determination. Attached Figure Description

[0013] Figure 1 A flowchart of a method for enhancing infrared polarization images of carotid artery plaques; Figure 2 This is a flowchart of the initial feature map extraction process in the image processing of carotid artery plaques; Figure 3 Flowchart for extracting the local density response tensor from carotid artery plaque images; Figure 4 Flowchart for constructing enhanced feature maps for carotid artery plaque images; Figure 5 A flowchart showing the final output image after processing carotid artery plaques; Figure 6 A comparison of the effects of infrared polarization image processing on carotid artery plaques before and after processing. Detailed Implementation

[0014] 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.

[0015] Please see Figure 1 The infrared polarization image enhancement processing method for carotid artery plaques in this embodiment of the application has the following specific steps.

[0016] Step 1: Obtain a single-channel grayscale image of the carotid artery plaque infrared polarization image as input. Calculate the edge map and gamma enhancement map based on the grayscale image. Then, stitch the grayscale image, edge map, and gamma enhancement map together in the channel dimension to construct a three-channel input image.

[0017] In this embodiment, 100 carotid artery plaque images were acquired using infrared polarization imaging to construct a carotid artery plaque infrared polarization image dataset. The images in the dataset were cropped to a uniform 512×512 fixed resolution and used to train the carotid artery plaque infrared polarization image enhancement processing model proposed in this invention. Specifically, the carotid artery plaque infrared polarization image dataset is divided into a training set and a test set in a 7:3 ratio. The training set is used to train the carotid artery plaque infrared polarization image enhancement model, outputting a carotid artery plaque infrared polarization image with optimized grayscale distribution. Then, the test set is used to test the trained carotid artery plaque infrared polarization image enhancement model. The test results are judged based on the loss function value. The test results directly represent the enhancement effect of the carotid artery plaque infrared polarization image; the smaller the loss function value, the better the enhancement effect. The method for constructing the carotid artery plaque infrared polarization image enhancement model is as follows: Figure 1 S1 to S6 are shown.

[0018] Furthermore, in step one, a single-channel grayscale image is used as input. The specific steps are as follows: In this embodiment, the color images of carotid artery plaques in the dataset are converted to grayscale using a weighted summation-based brightness conversion model to obtain a single-channel grayscale base image. The weighted summation-based brightness conversion model is as follows: R1, G1, and B1 are the three components of image color. This is a grayscale image of a single-channel carotid artery plaque. , where H and W are both 512.

[0019] The edge response map is calculated based on the grayscale image. The edge response map is obtained by extracting and fusing the grayscale image gradients in the horizontal and vertical directions. The edge response map is used to distinguish the boundary position between the carotid artery plaque and the surrounding normal tissue. Nonlinear mapping enhancement is performed on the grayscale image to output an enhanced image that highlights the local contrast of the image. The grayscale image, edge response map and enhanced image are stitched together in the channel dimension to construct a structurally enhanced three-channel input image.

[0020] In this embodiment, grayscale images Edge operator calculations are performed to obtain the edge map. The gradient map is obtained by first extracting the horizontal and vertical edges using the Sobel operator and then fusing the magnitudes; then the grayscale image is processed. Perform a nonlinear gamma transform to obtain the pre-enhanced image. Finally, at each pixel location, the edge map, pre-enhanced map, and single-channel carotid plaque grayscale image are stitched together along the channel dimension to obtain a structurally enhanced three-channel infrared polarization input image of the carotid plaque. The mathematical model for implementing the above method is as follows: ; Here, Concat is a stitching function that stitches the edge map, pre-enhanced map, and single-channel carotid artery plaque grayscale image, which are three one-dimensional images, into a three-dimensional image. , .

[0021] Step 2: Based on the three-channel input image, calculate the channel importance weight of each channel image, normalize the channel importance weights, and perform weighted fusion of the three channels of the input image according to the normalized channel weights to generate a single-channel fused image. Perform a convolution operation on the fused image to obtain the initial mapping value of the feature space.

[0022] Furthermore, in step two, the channel importance weight of each channel image is calculated, and the specific steps are as follows.

[0023] Based on the three-channel input image of the carotid artery plaque infrared polarization, the absolute average response value of each pixel in each channel is calculated. The absolute average response value is used as the importance factor of each channel in the overall image. The importance factors of each channel are normalized to obtain the corresponding channel importance weight.

[0024] Furthermore, such as Figure 2 As shown, in this embodiment, the three-channel input image is a three-channel image obtained by stitching together an edge map, a pre-enhanced map, and a single-channel carotid artery plaque grayscale image. Within its spatial dimension, the importance factor of each channel is calculated by averaging the absolute values ​​of all pixels in that channel. The specific implementation model is as follows: ; in, Let H be the importance factor of the i-th channel, H be the height of the infrared polarization three-channel input image of the carotid artery plaque, W be the width of the infrared polarization three-channel input image of the carotid artery plaque, and x and y be the pixel coordinates. for; The importance factor of each channel is normalized to obtain the channel importance weight of each channel image. The specific implementation model is as follows: .

[0025] The three-channel input image of the carotid artery plaque with infrared polarization is fused with the corresponding channel importance weights to generate a single-channel fused infrared polarization image of the carotid artery plaque. The fused image is then processed by a two-dimensional convolution operator to extract local gray-level structural features. The extracted local features are then subjected to channel dimensionality upscaling, and a tensor containing multiple feature channels is output as the initial feature space mapping value.

[0026] In this embodiment, firstly, the channels of the three-channel input image are weighted and fused based on the importance weights of each channel in the infrared polarization image of the carotid artery plaque. Then, the multi-channel infrared polarization images of the carotid artery plaque are fused into a single-channel fused image using a weighted summation method for feature extraction in step three. The weighted fusion ensures that the contribution ratio of various feature maps in the single-channel construction depends on their corresponding weight coefficients. The mathematical model is as follows: ; in, Infrared polarization images of carotid artery plaques fused from a single channel; Furthermore, based on the single-channel fused image, a two-dimensional convolution operation with a kernel size of 3×3 is used to perform feature mapping on the fused image, extracting the local gray-level structural features of the infrared polarization image of the carotid artery plaque; finally, this convolution operation is used to achieve dimensionality-upgrading mapping of the feature space, obtaining the initial mapping value of the feature space, which serves as the input to the subsequent image enhancement model; specifically, the initial mapping value of the feature space... The acquisition implementation model is as follows: ; Where C represents the number of channels in the feature map, with each channel corresponding to a feature extracted by a convolutional kernel, and the initial mapping value in the feature space. This is the initial feature map used in the image processing of carotid artery plaques.

[0027] Step 3: Perform non-normalized density-driven weighted projection on the initial mapping values ​​of the feature space, and fuse them to generate a single-channel structure-guided grayscale image. The weighting is the average number of channels of the initial mapping values ​​of the feature space. Construct the local density response tensor of the image based on the local pixel intensity differences of the structure-guided grayscale image.

[0028] Furthermore, such as Figure 3 As shown, in step three, a single-channel structure-guided grayscale image is generated. The specific steps are as follows.

[0029] Based on the initial feature space mapping value, the feature space mapping value of each channel is weighted with the channel weight, and all weighted projections are added point by point at the pixel position to obtain a single-channel structure guided grayscale image.

[0030] In this embodiment, non-normalized density-driven weighted projection is performed. First, the initial feature space mapping value of each channel is obtained. Where c = 1, 2, ..., C; initial mapping values ​​for the feature space of each channel. Multiply by the corresponding channel fusion weight The weighted projections are obtained, and then all the weighted projections are summed point by point at the pixel position to obtain a single-channel structure-guided grayscale image. ; Preferably, in this implementation, the channel fusion weights need to meet the normalization condition. The channel fusion weights Set to mean w c =1 / C, where channel C is a three-channel system.

[0031] Furthermore, in step three, the local density response tensor of the image is constructed, and the specific steps are as follows.

[0032] Based on the single-channel structure, an arbitrary target pixel in the grayscale image is used as the center position to establish a neighborhood range. For each pixel within the neighborhood range, its grayscale value is obtained, and the grayscale difference value between the grayscale value and the center pixel is calculated. A density response factor is constructed based on the grayscale difference value and a preset control parameter. The density response factor is weighted by an exponential function to calculate the grayscale difference. The density response factors of all neighboring pixels are superimposed to obtain the local density response tensor corresponding to the current target pixel.

[0033] In this embodiment, firstly, any pixel in the single-channel structure-guided grayscale image is selected as the center position, and a 3×3 neighborhood window is constructed. All pixels within this neighborhood window are considered as the neighborhood pixel set. For each pixel in this neighborhood, the grayscale value of the current neighboring pixel is read, and the grayscale value of the center pixel is subtracted from it to obtain a grayscale difference value. Then, the absolute value of this grayscale difference value is taken as the degree of grayscale deviation. This absolute value is multiplied by a negative coefficient to obtain a product value. An exponential operation is performed on the above product value to obtain a weight. Finally, the weights of all neighboring pixels are summed to obtain the final local density response tensor of the image. The specific mathematical model is as follows: ; in, For pixels within the neighborhood of pixel (i,j), To control the sensitivity of the response to grayscale differences, The grayscale value of the center pixel. This represents the grayscale value of each pixel within the neighborhood. This represents the local density response value. Among them, the sensitivity of the control response to grayscale differences The value is between [0.5, 2.5], and the initial value is... To balance the sensitivity of differential responses within the neighborhood with the smoothness of information fusion, the overall contrast and neighborhood variance of the grayscale image are dynamically adjusted based on the structure. The mathematical model is as follows: ; The grayscale variance within a 3×3 neighborhood window of the grayscale image guiding the structure. The mean value of all pixels within a 3×3 neighborhood window of the structure-guided grayscale image; The summation part reflects the first The degree of gray-level similarity between neighboring pixels and the center pixel determines the density response tensor value of the final image. The larger the value, the more concentrated the gray-level values ​​around the pixel, and the stronger the density response of the infrared polarization image of the carotid artery plaque. In this embodiment, the process is repeated for each pixel in the infrared polarization image of the carotid artery plaque, and finally a density response map of the same size as the original image is obtained.

[0034] Step 4: Calculate the horizontal and vertical gradients of the guided grayscale image to construct the directional gradient intensity response. Remap the directional gradient intensity response and the local density response tensor to the original feature space through channel restoration mapping. Introduce a dynamic amplitude control factor to obtain an enhanced feature map.

[0035] Furthermore, such as Figure 4 As shown, in step four, an enhanced feature map of the image is constructed, and the specific steps are as follows.

[0036] In this embodiment, the Sobel operator is first used to calculate the grayscale change amplitude of each pixel in the horizontal and vertical directions of the single-channel structure-guided grayscale image, respectively. These are then used as the horizontal and vertical gradient components. The squares of these two components are added together and then the square root is taken to construct the directional gradient intensity response. This process reflects the intensity of the pixel in the gradient direction. The specific implementation model for constructing the directional gradient intensity response is as follows: ; in, The directional gradient intensity response for each pixel; and These are the horizontal Sobel operator and the vertical Sobel operator, respectively.

[0037] Multiply the local density response tensor of the pixel by the logarithm of the gradient intensity response value plus one, and fuse the changes in gray density in the image with the changes in structural orientation intensity. Perform a one-to-one convolution operation on the fused image to convert the single-channel image into a C-channel spatial image and obtain the enhanced feature map of the image. During implementation, the kernel size of the convolution is 1, the stride is 1, and zero-padding is used. The mathematical model is as follows: ; in, Enhanced feature map of infrared polarization image of carotid artery plaque. .

[0038] Furthermore, in this embodiment, during the mapping to the original feature space, the enhanced response value at each location is multiplied with the weights in the corresponding convolutional kernel, and the results are accumulated to generate a new output value. The final output feature map size is consistent with the original feature, and the number of channels is restored to the preset feature dimension. The process involves injecting enhanced information driven by gray-level density into the original feature channel space. In this process, the changes in gray-level density in the image are fused with the intensity of changes in structural orientation, which can better highlight the detailed changes at the edge of the infrared polarization image of the carotid artery plaque and exhibit a stronger response in areas with dense gray-level density and abrupt gradient changes.

[0039] Step 5: Based on the enhanced feature map, calculate the local spatial change rate at each position of each channel. Perform a two-dimensional Fourier transform on the enhanced feature map to construct a frequency residual signal. Combine the change information of the spatial and frequency components to construct a dynamic adjustment factor. Adjust the response amplitude of the enhanced feature map and output the final feature map, as shown below. Figure 5 As shown.

[0040] Furthermore, in step five, a dynamic adjustment factor is constructed by combining the change information of both spatial and frequency components. The specific steps are as follows.

[0041] Based on any pixel in each channel of the enhanced feature map, a local neighborhood region is constructed centered on that pixel. The difference between the current pixel value and the average value of all pixels in the neighborhood region is calculated. Using this difference, the local spatial change rate of the current pixel in the enhanced feature map is constructed. A frequency domain transformation is performed on each channel of the enhanced feature map to obtain the spectral form. A blurred spectrum map is then constructed using a smoothing filter function. The difference amplitude between the original spectrum map and the blurred spectrum map is calculated to construct a frequency residual signal. The median of the local spatial change rate and the frequency residual signal is calculated within the local neighborhood. The response offset is calculated based on the difference between the current pixel and the corresponding median. The variance of the local spatial change rate and the frequency residual signal is calculated within the local neighborhood to obtain a normalization factor. The response offset and the normalization factor are combined using a nonlinear function to construct a dynamic amplitude adjustment factor.

[0042] In this embodiment, the feature map is enhanced. It consists of C channels, each channel corresponds to a type of feature response and has the same spatial size; for each channel, taking any pixel in the current channel as the center, extract a local neighborhood region composed of several pixels around it, and the size of the neighborhood region can be set to 3×3; then, calculate the difference between the value of the center pixel and the average value of all pixels in its neighborhood as the local spatial change rate at that position. ;in, Let be the local spatial change rate of the c-th channel at point (i,j) in the neighborhood. The local spatial change rate reflects the intensity of structural change of the pixel in the enhanced feature map. Let be the enhanced feature map value of the c-th channel at pixel position (i,j). Let (i,j) be the neighborhood region centered at (i,j), and (p,q) be the pixel index within the neighborhood. Furthermore, the input enhanced feature map is subjected to discrete Fourier transform to obtain the frequency domain representation, and then Gaussian filtering is performed. The difference between the original spectrum and the fuzzy spectrum is calculated to obtain the frequency residual signal map. ; in, This is the frequency residual signal map in the current channel frequency domain. Specifically, it is obtained through the spectrum of the original enhanced feature map. Subtract smooth spectrum The absolute value result preserves details and edge frequency responses in the enhanced feature map; Furthermore, within the neighborhood window of each pixel position (i,j), the local median of the structural response and frequency response are calculated respectively. Furthermore, the local spatial rate of change and frequency residual signal map value corresponding to each pixel point within the neighborhood window are calculated to construct the amplitude dynamic adjustment factor. Preferably, the neighborhood window range is 3×3 window. ; in, This is the dynamic amplitude adjustment factor for the c-th channel at point (i,j) in the local neighborhood; Let be the frequency residual signal value of pixel (i,j) in the neighborhood of the c-th channel. It is a very small positive value, taking the value 1×e -6 ; Let be the median of the local spatial rate of change corresponding to the pixels in the local neighborhood. Let Var be the median of the frequency residual signal values ​​corresponding to the pixels in the local neighborhood, and Var be the variance calculation function in the local neighborhood. The local spatial rate of change at point (i,j) within the neighborhood. This represents the frequency residual signal map value of pixel (i,j) in the neighborhood.

[0043] In this embodiment, the numerator of the amplitude dynamic adjustment factor represents the cooperative offset of the structural response and frequency response relative to their respective medians. A positive enhancement trend occurs when the offset directions are consistent. The denominator uses a covariance normalization form based on the Mahalanobis distance structure to automatically suppress the enhancement amplitude in regions with severe local fluctuations or high noise. Furthermore, through... The amplitude dynamic adjustment factor is mapped to the (0,1) interval to increase the amplitude only in the true double strong response region.

[0044] In this embodiment, the enhanced feature map is multiplied point-by-point along the pixel dimension by the amplitude adjustment factor to adjust the response amplitude and complete the final enhanced output of the feature map. The mathematical model is as follows: ; in, This represents the final enhanced eigenvalue after dynamic adjustment of the fusion structure and frequency.

[0045] Step 6: Weightedly fuse the final feature maps of each channel and normalize them using an activation function to obtain the final output image, thereby optimizing the grayscale distribution of the infrared polarization image of carotid artery plaques.

[0046] Further, in step six, the final output image is a three-channel image. The enhanced feature maps of multiple channels are weighted and summed according to the channel direction, and a bias term and activation function are introduced to compress the value range of the fusion result, ultimately forming a three-channel output infrared polarization image of carotid artery plaques. The mathematical model is as follows: ; in, It is obtained by weighted fusion of enhanced feature values ​​from multiple channels at each pixel (i,j). The weighting coefficients corresponding to channel c. For bias terms, specifically, the overall output offset after linear weighting is adjusted by bias terms.

[0047] In this embodiment, the model constructed in steps one to six above is trained using the training set in the dataset. During the training process, the carotid artery plaque infrared polarization image of the input dataset is 512×512×3, and the output image is calculated. With target image The difference between them yields the loss value, and the mathematical model is as follows: ; Through loss value Weighting coefficients and bias terms for channel c By performing differentiation, the gradients of the weight coefficients and bias terms are calculated; then, the weight coefficients and bias terms corresponding to channel c are updated using the gradient descent algorithm. The mathematical model is: and ;in, The learning rate; The goal is to minimize the loss function. The initial learning rate is set to 0.001, which decays exponentially and non-linearly with the number of iterations. The training iterations are set to 100, and the batch size for each training iteration is set to 10. After 100 training iterations, the final enhanced infrared polarization image of the carotid artery plaque is output.

[0048] In an optional embodiment, this application also provides a computer-readable storage medium storing a carotid artery plaque infrared polarization image enhancement program for steps one to six. The program includes computer-executable instructions, which, when executed by a computer's processor, can be used to implement any of the aforementioned carotid artery plaque infrared polarization image enhancement methods. The image enhancement program can be developed in a computing platform using the Python programming language and built based on the PyTorch framework, and can be deployed in an image processing device with computing capabilities.

[0049] like Figure 6 , Figure 6 In the infrared polarization image of the carotid artery plaque in region A, the overall grayscale distribution is relatively concentrated, with some areas exhibiting grayscale stacking, resulting in significant low contrast. This is particularly evident in the unclear grayscale boundary between the background and the main plaque area. Figure 6 As shown in B, the method of the present invention effectively improves the overall grayscale dynamic range of the image, making the grayscale distribution of the enhanced image more uniform, the distinction between bright and dark areas more obvious, and improving the visual perception contrast of the lesion area.

Claims

1. A method for enhancing infrared polarization images of carotid artery plaques, characterized in that, include: A single-channel grayscale image of a carotid artery plaque with infrared polarization is acquired as input. An edge map and a gamma enhancement map are calculated based on the grayscale image to construct a three-channel input image. Calculate the channel importance weight of each channel image, normalize the channel importance weights, and perform weighted fusion of the three channels of the input image to generate a single-channel fused image. Perform a convolution operation on the fused image to obtain the initial mapping value of the feature space. Non-normalized density-driven weighted projection is performed on the initial mapping values ​​of the feature space, and a single-channel structure-guided grayscale image is generated by fusing them. The local density response tensor of the image is constructed based on the local pixel intensity differences of the structure-guided grayscale image. Among them, non-normalized density-driven weighted projection is to weight the initial feature space mapping value of each channel with the channel weight, and add all weighted projections point by point at the pixel position. The horizontal and vertical gradients of the guiding grayscale image are calculated to construct the directional gradient intensity response. The directional gradient intensity response and the local density response tensor are then remapped to the original feature space through channel restoration mapping to obtain the enhanced feature map. Specifically, gradient calculations are performed on the single-channel structure-guided grayscale image in both the horizontal and vertical directions to calculate the grayscale change amplitude of each pixel in different directions, thus constructing the directional gradient intensity response of the pixel. The directional gradient intensity response is then jointly mapped with the local density response tensor of the corresponding pixel. Based on the fusion relationship between the two, a combined feature reflecting the local density change and the structural directional intensity is formed. A one-to-one channel convolution operation is performed on the fused result to generate an enhanced feature map of the image. Based on the enhanced feature map, the local spatial change rate of each channel at each position is calculated. A two-dimensional Fourier transform is performed on the enhanced feature map to construct a frequency residual signal. The change information of the spatial and frequency parts are combined to construct a dynamic adjustment factor. The response amplitude of the enhanced feature map is adjusted, and the final feature map is output. The final feature maps of each channel are weighted and fused, and then normalized using an activation function to optimize the grayscale distribution of the infrared polarization image of carotid artery plaques.

2. The method for enhancing carotid artery plaque infrared polarization images according to claim 1, characterized in that, Infrared polarized color images of carotid artery plaques are converted into single-channel grayscale images. The conversion is based on a weighted brightness mapping model that performs a weighted fusion operation on multiple color channels of the image to obtain the grayscale image. An edge response map is calculated based on the grayscale image. The edge response map is obtained by extracting and fusing the grayscale image gradients in the horizontal and vertical directions. The edge response map is used to distinguish the boundary position between carotid artery plaques and surrounding normal tissues. Gamma enhancement is performed on the grayscale image to output an enhanced image that highlights the local contrast of the image. The grayscale image, edge response map and enhanced image are stitched together in the channel dimension to construct a three-channel input image of the carotid artery plaque with infrared polarization after structural enhancement.

3. The method for enhancing carotid artery plaque infrared polarization images according to claim 2, characterized in that, Based on the three-channel input image of the carotid artery plaque infrared polarization, the absolute average response value of each pixel in each channel is calculated. The absolute average response value is used as the importance factor of each channel in the overall image. The importance factors of each channel are normalized to obtain the corresponding channel importance weight. The three-channel input image of the carotid artery plaque with infrared polarization is fused with the corresponding channel importance weights to generate a single-channel fused infrared polarization image of the carotid artery plaque. The fused image is then processed by a two-dimensional convolution operator to extract local gray-level structural features. The extracted local features are then subjected to channel dimensionality upscaling, and a tensor containing multiple feature channels is output as the initial feature space mapping value.

4. The method for enhancing carotid artery plaque infrared polarization images according to claim 3, characterized in that, Based on the single-channel structure, an arbitrary target pixel in the grayscale image is used as the center position to establish a neighborhood range. For each pixel in the neighborhood range, its grayscale value is obtained, and the grayscale difference value between the grayscale value and the center pixel is calculated. Based on the grayscale difference value and combined with preset control parameters, a density response factor is constructed. The density response factor is processed by exponential operation and superposition to obtain the local density response tensor corresponding to the current target pixel. The control parameters are dynamically adjusted by combining the ratio of the grayscale variance value to the grayscale average value of all pixels in the local region with the initial sensitivity parameter.

5. The method for enhancing carotid artery plaque infrared polarization images according to claim 1, characterized in that, Based on any pixel in each channel of the enhanced feature map, a local neighborhood region is constructed with that pixel as the center. The difference between the current pixel value and the average value of all pixels in the neighborhood region is calculated. The difference is used to construct the local spatial change rate of the current pixel in the enhanced feature map. Perform frequency domain transformation on each channel of the enhanced feature map to obtain the spectral form, then use a smoothing filter function to construct a fuzzy spectrum map, and calculate the difference amplitude between the original spectrum map and the fuzzy spectrum map to construct a frequency residual signal. The median of the local spatial rate of change and the frequency residual signal is calculated within the local neighborhood. The response offset is calculated based on the difference between the current pixel and the corresponding median. The variance of the local spatial rate of change and the frequency residual signal is calculated within the local neighborhood to obtain a normalization factor. The response offset and the normalization factor are combined using a nonlinear function to construct a dynamic amplitude adjustment factor. The enhanced feature map and the dynamic amplitude adjustment factor are multiplied point by point along the pixel dimension to obtain the final feature map.

6. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and the storage medium stores an image enhancement program, which can implement the carotid artery plaque infrared polarization image enhancement processing method as described in any one of claims 1 to 5.

Images