Thyroid nodule ultrasonic image classification method and system based on feature extraction

Abstract

The invention relates to the technical field of image classification, in particular to a thyroid nodule ultrasonic image classification method and system based on feature extraction, and the method comprises the following steps: based on thyroid nodule ultrasonic image data, calling a pixel normalization function to adjust a gray scale range to a set interval, screening an isolated noise region, and carrying out the smooth filtering, and acquiring the ultrasonic image after smoothing filtering. According to the method, by adjusting the gray scale range and screening isolated noise areas, gray scale distribution is balanced, noise interference is reduced, image readability is improved, edge change is recognized by analyzing the gradient direction through a small-scale convolution kernel, texture features are extracted through a medium-scale filter, redundancy is reduced in combination with pooling operation, and the feature hierarchical expression ability is enhanced; the edge sharpness parameter is calculated, the feature weight is optimized, the edge sharpness and the classification adaptability are improved, the local feature segmentation precision is improved, the classification boundary is adjusted according to the similarity score, the classification loss measurement is optimized, the error offset is reduced, and the classification stability and accuracy are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to a method and system for classifying thyroid nodule ultrasound images by feature extraction. Background Art

[0002] The field of image classification technology includes methods and applications for feature extraction, analysis and classification of images of different categories. The core content of this technical field involves the acquisition of image data, feature selection, classification standard setting and implementation of classification methods, including preprocessing of target images, feature quantification, classification rule construction and classification result verification. In medical image analysis, image classification technology is widely used in the automatic identification and classification of different tissues or lesion areas. By analyzing the texture, edge, morphology and light intensity distribution in the image, normal tissues and abnormal lesions are distinguished. Systematic research in this technical field covers classification methods based on image statistical features, classification methods based on morphological features, and methods for comprehensive judgment based on multiple features. The classification accuracy is optimized through different classification standards to improve the accuracy and stability of medical image analysis.

[0003] Among them, the thyroid nodule ultrasound image classification method refers to a classification method that distinguishes different types of nodules by analyzing the characteristic information in the thyroid nodule ultrasound image. This method extracts tissue echo characteristics through the gray value distribution of the ultrasound image, and calculates the texture parameters in the image to characterize the uniformity and roughness of the tissue structure. This method uses morphological analysis to determine the boundary morphology and growth direction of the nodule, and combines statistical methods to extract edge gradient distribution characteristics. This method matches the extracted features through the established classification standards, and makes category judgments based on the set classification rules, thereby realizing the classification of thyroid nodule images.

[0004] The existing technology relies on the singleness of feature extraction in the classification process. The gray value distribution in ultrasound images is greatly disturbed by artifacts, which easily leads to insufficient image contrast, thus affecting the extraction of nodule features. For the calculation method of texture parameters, the existing method has weak adaptability to texture changes of different scales and it is difficult to fully reflect the uniformity and roughness of tissue structure. In the edge recognition process, due to the complex boundary morphology, the sole reliance on morphological analysis leads to insufficient ability to capture tiny edge changes, which reduces the recognition accuracy of some abnormal areas. The existing classification rules are highly fixed. When faced with different types of nodules, it is difficult to dynamically adjust the classification boundaries according to changes in feature distribution, resulting in some nodules with weaker edge features being misclassified. The limitations of the existing technology in feature extraction, feature quantification, and classification standard setting limit the classification accuracy and affect the reliability of medical image analysis. Summary of the invention

[0005] In order to solve the technical problems that in the existing classification process, it depends on the singularity of feature extraction, the gray value distribution in ultrasonic images is greatly interfered by artifacts, which easily leads to insufficient image contrast, thus affecting the extraction of nodule features. For the calculation method of texture parameters, the existing methods have weak adaptability to texture changes at different scales and are difficult to comprehensively reflect the uniformity and roughness of tissue structures. In the edge recognition process, due to the complex boundary morphology, simply relying on morphological analysis results in insufficient ability to capture tiny edge changes, reducing the recognition accuracy of some abnormal regions. The strong fixity of the existing classification rules makes it difficult to dynamically adjust the classification boundary according to the changes in feature distribution when facing different types of nodules, resulting in misclassification of some nodules with weak edge features. The limitations of the existing technology in aspects such as feature extraction, feature quantization, and classification standard setting limit the classification accuracy and affect the reliability of medical image analysis. The embodiments of the present invention provide a classification method and system for ultrasonic images of thyroid nodules with feature extraction. The technical solutions are as follows:

[0006] On the one hand, a classification method for ultrasonic images of thyroid nodules with feature extraction is provided. The method includes:

[0007] S1: Based on the ultrasonic image data of thyroid nodules, call the pixel normalization function to adjust the gray range to a set interval, screen out isolated noise regions and perform smoothing filtering to obtain the smoothed ultrasonic image;

[0008] S2: Based on the smoothed ultrasonic image, use a small-scale convolution kernel to analyze the pixel gradient direction, screen out edge change regions, apply a medium-scale filter to identify texture changes, reduce redundancy through pooling operations, and set global contrast parameters to evaluate brightness changes to obtain multi-level feature distribution values;

[0009] S3: Call the multi-level feature distribution values, calculate the nodule edge sharpness parameter, set the calculation rule of the mean square error of the edge gradient to obtain the nodule edge sharpness ratio, and adjust the feature weights according to the classification error feedback to obtain the image feature contribution ratio;

[0010] S4: Call the image feature contribution ratio, set the image area division standard, perform local nodule feature segmentation, and use cosine similarity analysis to screen out similar regions to obtain the nodule feature similarity score;

[0011] S5: Call the nodule feature similarity score, set the classification boundary adjustment rule, analyze the classification error deviation, optimize the measurement method of classification loss, and obtain the classification adjustment result of the thyroid nodule image.

[0012] As a further solution of the present invention, the smoothed and filtered ultrasound image includes a grayscale normalized image, a denoised image, and a smoothed and filtered image. The multi-level feature distribution values include edge change features, texture change features, and brightness change features. The nodule edge sharpness ratio includes edge sharpness parameters, mean square deviation of edge gradients, and feature weight ratios. The image feature contribution ratio includes region division criteria, local feature segmentation criteria, and similarity analysis parameters. The nodule feature similarity score includes local similarity scores, global similarity scores, and cosine similarity values. The thyroid nodule image classification adjustment result includes classification boundary parameters, error offset analysis values, and classification loss optimization parameters.

[0013] As a further solution of the present invention, the steps of the smoothed and filtered ultrasound image are specifically as follows:

[0014] S101: Based on the thyroid nodule ultrasound image data, analyze the pixel grayscale range, perform a linear transformation according to the normalization scale factor, adjust the pixel grayscale value to the set interval, and obtain the normalized grayscale image data;

[0015] S102: Based on the normalized grayscale image data, detect isolated noise regions, calculate the mean and standard deviation of pixel grayscales within the local window, screen abnormal pixels according to the noise judgment threshold, call the mean grayscale of adjacent pixels to replace the abnormal pixels, and adjust the local abnormal grayscale distribution to obtain the image data after noise removal;

[0016] S103: For the image data after noise removal, calculate the weighted mean of pixels within the local window, perform a smoothing operation according to the weighting coefficient, update the ultrasound image data, and obtain the smoothed and filtered ultrasound image.

[0017] As a further solution of the present invention, the steps of the multi-level feature distribution values are specifically as follows:

[0018] S201: Based on the smoothed and filtered ultrasound image, call a small-scale convolution kernel to calculate the pixel gradient value, screen the edge region according to the change of the gradient direction, and obtain the edge gradient distribution value;

[0019] S202: For the edge gradient distribution value, analyze the amplitude of gray-scale change within the local window, screen the texture region according to the gradient change rate, call the set threshold to judge the pixel points whose gray-scale difference exceeds the standard, update the pixel matrix and adjust the local gray-scale distribution to obtain the texture change coefficient;

[0020] S203: According to the texture change coefficient, identify the gray-scale difference within the pooling window, call the global contrast parameter to evaluate the brightness change, and obtain the multi-level feature distribution value.

[0021] As a further solution of the present invention, the texture change coefficient adopts the formula:

[0022]

[0023] Among them, TVC represents the texture change coefficient, N represents the total number of pixel points in the local window, ΔG i represents the gradient change amount of the i-th pixel point, θ represents the set gradient change threshold, G i represents the gray value of the i-th pixel point, G0 represents the gray value of the pixel point at the window center, and ∈ is a constant to avoid division-by-zero errors.

[0024] As a further solution of the present invention, the steps of the image feature contribution ratio are specifically as follows:

[0025] S301: Call the multi-level feature distribution value, analyze the gray change in the nodule edge area, identify the edge sharpness based on the pixel gradient direction, and obtain the nodule edge sharpness parameter;

[0026] S302: For the nodule edge sharpness parameter, calculate the mean square deviation of the pixel gradient values within the local window, screen the pixel points that meet the feature adjustment according to the gradient change, call the edge sharpness ratio calculation rule, adjust the weights of the abnormal gradient points, update the pixel gradient data, and obtain the local gradient sharpness ratio;

[0027] S303: Based on the local gradient sharpness ratio, identify the classification error feedback amount, call the feature weight adjustment coefficient, update the feature contribution ratio according to the weight deviation, and obtain the image feature contribution ratio.

[0028] As a further solution of the present invention, the steps of the nodule feature similarity score are specifically as follows:

[0029] S401: Call the image feature contribution ratio, set the region division standard according to the gray change in the feature region, analyze the pixel gradient range, and obtain the region division parameter;

[0030] S402: Based on the region division parameter, call the pixel gradient information within the local window, perform nodule local feature segmentation according to the feature change boundary, screen the pixel points that meet the division standard, and obtain the nodule local segmentation feature;

[0031] S403: Based on the nodule local segmentation feature, analyze the cosine similarity between the region feature vectors, screen the similar regions, and obtain the nodule feature similarity score.

[0032] As a further solution of the present invention, the nodule feature similarity score uses the formula:

[0033]

[0034] Among them, S represents the nodule feature similarity score, A kThe k-th eigenvector component representing region A, B k The k-th eigenvector component representing region B, and M represents the number of dimensions of the eigenvector.

[0035] As a further solution of the present invention, the steps of adjusting the classification result of the thyroid nodule image are specifically as follows:

[0036] S501: Invoke the nodule feature similarity score, calculate the mean value of the eigenvectors of the category samples, analyze the deviation of the category boundary based on the mean value difference, screen the regions where the deviation exceeds the threshold, and obtain the classification boundary adjustment parameter;

[0037] S502: Based on the classification boundary adjustment parameter, calculate the classification error deviation value of the category region, invoke the error feedback data, screen the sample points with errors greater than the threshold, correct the category discrimination weight according to the error distribution gradient, adjust the category discrimination weight distribution ratio, optimize the classification loss calculation method, and obtain the optimized classification loss parameter;

[0038] S503: Based on the optimized classification loss parameter, analyze the error adjustment amount, invoke the correction factor to update the category division boundary, and adjust the boundary coefficient according to the feature distribution to obtain the adjusted result of the thyroid nodule image classification.

[0039] On the other hand, the electric vehicle status monitoring system is used to execute the above-mentioned electric vehicle status monitoring method, and the system includes:

[0040] The ultrasonic image processing module analyzes the gray range based on the ultrasonic image data of the thyroid nodule, adjusts the pixel gray value, identifies the gradient change rate to screen noise points, invokes local smoothing to calculate the pixel equalization degree, and obtains the smoothed and normalized image;

[0041] The edge detection module sets a gradient change threshold based on the smoothed and normalized image to screen edge points, identifies the pixel density to obtain the edge region, calculates the direction change rate to evaluate the edge sharpness, invokes local pixel statistics to analyze the edge gray uniformity, identifies the contrast mutation points to screen the boundary transition zone, and invokes the morphological stability to calculate the boundary symmetry degree to obtain the nodule edge gradient distribution;

[0042] The nodule contour analysis module invokes the contour tracking data to calculate the connectivity based on the nodule edge gradient distribution, analyzes the edge mutation points to screen the contour curvature change, and obtains the nodule contour shape parameters;

[0043] The region feature calculation module calculates the region gray mean and variance based on the nodule contour shape parameters, invokes morphological measurement to identify the region extension degree, and obtains the nodule region feature parameters;

[0044] Based on the characteristic parameters of the nodule region, the classification structure optimization module calculates the classification boundary value, calls the feature matching to analyze the category similarity, corrects the classification error range, optimizes the classification decision parameters, and obtains the classification adjustment result of the thyroid nodule image.

[0045] The beneficial effects brought by the technical solution provided by the embodiment of the present invention at least include:

[0046] By adjusting the gray range and screening the isolated noise regions, the overall gray distribution of the image is made more balanced, the interference of isolated noise to feature analysis is reduced, and the readability of the ultrasonic image is improved. Analyzing the pixel gradient direction using a small-scale convolution kernel helps to accurately identify the edge change regions, and identifying texture changes through a medium-scale filter can retain the key structural features while reducing the influence of irrelevant details. Combining the pooling operation reduces data redundancy, making the feature extraction more targeted, and optimizing the evaluation method of brightness changes through the global contrast parameter, thereby enhancing the hierarchical expression ability of image features. Calculating the edge sharpness parameter and setting the mean square error calculation rule enhance the characterization ability of the nodule edge clarity. Combining the classification error feedback to adjust the feature weights makes the feature contribution ratio more in line with the classification requirements. Setting the region division standard and applying the similarity analysis make the local feature segmentation of the nodule more accurate, improving the internal structure consistency of the region. Adjusting the classification boundary according to the similarity score and optimizing the measurement method of classification loss reduce the classification error deviation, making the classification decision more stable. The overall solution improves the classification accuracy and stability of the ultrasonic image by optimizing the feature extraction method, enhancing the feature contrast, reasonably distributing the feature weights, and optimizing the classification rules. Description of the Drawings

[0047] Figure 1 It is a schematic diagram of the working process of the present invention;

[0048] Figure 2 It is a detailed flowchart of S1 of the present invention;

[0049] Figure 3 It is a detailed flowchart of S2 of the present invention;

[0050] Figure 4 It is a detailed flowchart of S3 of the present invention;

[0051] Figure 5 It is a detailed flowchart of S4 of the present invention;

[0052] Figure 6 It is a detailed flowchart of S5 of the present invention;

[0053] Figure 7 It is a system flowchart of the present invention. Detailed Embodiments

[0054] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0055] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0056] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0057] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0058] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0059] See also Figure 1 The embodiment of the present invention provides a method for classifying thyroid nodules by ultrasound images based on feature extraction. The processing flow of the method may include the following steps:

[0060] S1: Based on the ultrasound image data of thyroid nodules, the pixel normalization function is called to adjust the grayscale range to the set interval, the isolated noise area is screened and smoothed, and the ultrasound image after smoothing is obtained;

[0061] S2: Based on the smoothed filtered ultrasound image, a small-scale convolution kernel is used to analyze the pixel gradient direction, the edge change area is screened, a medium-scale filter is applied to identify texture changes, redundancy is reduced through pooling operations, global contrast parameters are set to evaluate brightness changes, and multi-level feature distribution values ​​are obtained;

[0062] S3: Call the multi-level feature distribution value, calculate the nodule edge sharpness parameter, set the mean square error calculation rule of the edge gradient, obtain the nodule edge sharpness ratio, adjust the feature weight according to the classification error feedback, and obtain the image feature contribution ratio;

[0063] S4: Invoke the image feature contribution ratio, set the image region division criterion, perform local nodule feature segmentation, apply cosine similarity analysis to screen for similar regions, and obtain the nodule feature similarity score;

[0064] S5: Invoke the nodule feature similarity score, set the classification boundary adjustment rule, analyze the classification error offset, optimize the measurement method of the classification loss, and obtain the classification adjustment result of the thyroid nodule image.

[0065] The smoothed ultrasound image includes a gray-scale normalized image, a denoised image, and a smoothed filtered image. The multi-level feature distribution values include edge change features, texture change features, and brightness change features. The nodule edge sharpness ratio includes edge sharpness parameters, edge gradient mean square errors, and feature weight ratios. The image feature contribution ratio includes region division criteria, local feature segmentation criteria, and similarity analysis parameters. The nodule feature similarity score includes local similarity scores, global similarity scores, and cosine similarity values. The classification adjustment result of the thyroid nodule image includes classification boundary parameters, error offset analysis values, and classification loss optimization parameters.

[0066] Specifically, as Figure 2 shown, the steps of the smoothed ultrasound image are specifically as follows:

[0067] S101: Based on the thyroid nodule ultrasound image data, analyze the pixel gray-scale range, perform a linear transformation according to the normalization scale factor, and adjust the pixel gray-scale value to the set interval to obtain the normalized gray-scale image data;

[0068] When analyzing the pixel gray-scale range of the image and performing normalization processing, first obtain the ultrasound image data of the thyroid nodule. This data is the gray-scale value of each pixel point (for example, represented by an integer value from 0 to 255). The first step performed is to calculate the minimum and maximum gray-scale values of the entire image. For example, the gray-scale value range of the image is [45, 210]. Perform a linear transformation according to the set target gray-scale value range. For example, the target range is set to [0, 255], and perform normalization processing through the formula:

[0069]

[0070] where G is the pixel gray-scale value of the original image, G min is the minimum gray-scale value of the original image, G max is the maximum gray-scale value of the original image, G target min and G target max are the minimum and maximum values of the target range respectively. Through this formula, the gray-scale value of each pixel in the original image is converted to the target range to complete gray-scale normalization. For example, for a certain gray-scale value G = 45, after processing;

[0071] Its new grayscale value is:

[0072] In this way, through linear transformation, the data range of the entire image has been adjusted to [0, 255].

[0073] S102: Based on the normalized grayscale image data, detect isolated noise regions, calculate the mean and standard deviation of pixel grayscale values within a local window, screen out abnormal pixels according to the noise judgment threshold, call the mean grayscale value of adjacent pixels to replace the abnormal pixels, adjust the local abnormal grayscale distribution, and obtain the image data after noise removal;

[0074] During the noise removal process, first, according to the normalized grayscale image data, select an appropriate local window (such as a 3×3 window) for traversal. For each window, calculate the mean and standard deviation of the pixel grayscale values within the window. Suppose the grayscale values within a certain window are [80, 82, 85, 78, 90, 91, 95, 80, 100];

[0075] Its mean is:

[0076] The standard deviation is:

[0077] By calculating the mean and standard deviation, abnormal pixels are judged. If the grayscale value of a certain pixel deviates too much from the mean and exceeds the set threshold (such as the set threshold is 2 times the standard deviation), then this pixel is considered abnormal. For example, the grayscale value of a certain pixel is 50, its deviation from the mean 86.56, and the difference is greater than the set threshold 2×6.5 = 13. Therefore, it is judged as noise. Replace the grayscale value of this abnormal pixel with the mean of the neighborhood pixels. The grayscale values of the neighborhood pixels are [82, 85, 78, 90, 91, 95, 80, 100], and the mean is:

[0078]

[0079] Therefore, the grayscale value of the abnormal pixel is replaced with 84.625. Through this operation, the abnormal pixel is repaired, the noise is effectively removed, the local grayscale distribution of the image is adjusted, and thus the image data after noise removal is obtained.

[0080] S103: For the image data after noise removal, calculate the weighted mean of pixels within a local window, perform a smoothing operation according to the weighting coefficient, update the ultrasonic image data, and obtain the ultrasonic image after smoothing filtering;

[0081] After noise removal, when performing smoothing processing on the image, first set a weighting coefficient. For example, use a Gaussian weighting coefficient to smooth the pixel grayscale. Suppose a 3×3 Gaussian weighting matrix is used, and the weights are:

[0082] For each pixel, the gray values of the pixel and its neighboring pixels are weighted and averaged with the weighting coefficients. Taking a certain pixel as an example, assume the gray value of this pixel is 90, and the neighboring gray values are [85, 92, 88, 87, 90, 92, 86, 91, 89]. After applying the weighting coefficients, the smoothed value is calculated as follows:

[0083]

[0084] According to this formula, the gray value of each pixel will be adjusted according to the weighting coefficients, so as to achieve the effect of smoothing filtering. The entire ultrasonic image will be updated by this method to generate a smoothed ultrasonic image after filtering, which removes noise and maintains the smooth transition of image details.

[0085] Specifically, as Figure 3 shown, the steps of the multi-level feature distribution value are specifically as follows:

[0086] S201: Based on the ultrasonic image after smoothing filtering, call a small-scale convolution kernel to calculate the pixel gradient value, screen the edge area according to the change of the gradient direction, and obtain the edge gradient distribution value;

[0087] First, select a small-scale convolution kernel, such as a 3×3 Sobel operator, to calculate the pixel gradient value of the image. For each pixel, the convolution operation multiplies the gray values of the pixel and its surrounding 8 pixels with the weights of the convolution kernel and sums them to obtain the gradient value of this point. For example, assume the gray values of the current pixel and its neighborhood are [80, 82, 85, 78, 90, 91, 95, 80, 100]. If the Sobel operator is selected for convolution calculation, the calculation process will involve the following matrix operations:

[0088]

[0089] This calculation process will obtain the gradient in the horizontal direction of the image. Similar operations will be applied to the vertical direction to obtain the vertical gradient value. Combining the horizontal and vertical gradients, the Pythagorean theorem can be used to calculate the total gradient value of each pixel:

[0090]

[0091] In practical applications, the edge area appears as a place with a higher total gradient value. Therefore, according to the calculated total gradient value, the edge area can be screened out. For example, when the total gradient value is greater than a certain set threshold (for example, the set threshold is 50), then this area is considered as the edge area. Through the edge gradient distribution value, the edge area in the image can be accurately identified, thus realizing the edge detection of the image.

[0092] S202: Analyze the amplitude of gray-scale variation within the local window for the edge gradient distribution values, screen the texture regions based on the gradient change rate, calculate the texture change coefficient, call the set threshold to determine the pixel points with gray-scale differences exceeding the standard, update the pixel matrix and adjust the local gray-scale distribution to obtain the texture change coefficient;

[0093] The texture change coefficient is calculated using the formula:

[0094]

[0095] where TVC represents the texture change coefficient, N represents the total number of pixel points in the local window, ΔG i represents the gradient change amount of the i-th pixel point, θ represents the set gradient change threshold, G i represents the gray-scale value of the i-th pixel point, G0 represents the gray-scale value of the pixel point at the center of the window, and ∈ is a constant to avoid division-by-zero errors;

[0096] Detailed explanation of the formula and the derivation process of the formula calculation:

[0097] N: This parameter is determined according to the size of the local window selected in the actual calculation area. For example, assuming the local window is a 3×3 area, then N = 9;

[0098] ΔG i : The gradient change amount of each pixel point can be calculated through the gray-scale values of adjacent pixel points. Assuming that the change in gray-scale value is obtained by calculating the image gradient, the specific calculation method is the gray-scale difference between the pixel point and its surrounding pixel points. For example, if the gray-scale difference between the i-th pixel point and its neighborhood is 3, then ΔG i = 3;

[0099] θ: The gradient change threshold is generally based on experience or set to a standard value to distinguish whether there is significant texture change. Assuming the set threshold is 5, that is, θ = 5;

[0100] G i and G0: The gray-scale values are from the pixels of the image and are obtained by the image processing module. For example, the gray-scale value G i of the i-th pixel point = 120, while the gray-scale value G0 of the center of the window = 100;

[0101] ∈: This constant takes a very small value to avoid division-by-zero errors. Assuming ∈ = 0.01;

[0102] Calculating the gradient change amount ΔG i : For each pixel point, calculate the gray-scale difference between it and its surrounding pixel points. In this example, assuming the gray-scale value of the i-th pixel point is 120 and the difference in neighborhood gray-scale values is 3, then ΔG i = 3;

[0103] Calculate the difference between the gradient change and the threshold: Calculate |ΔG i - θ|. In this example, assume θ = 5, |ΔG i - θ| = |3 - 5| = 2;

[0104] Calculate the gradient change ratio: Calculate In this example, assume the gray value G i of the i-th pixel is 120 and the gray value G0 of the window center is 100, and ∈ = 0.01,

[0105] Calculate the texture variation coefficient (TVC): Assume there are a total of 9 pixels in the local window, then N = 9, and the final texture variation coefficient TVC is calculated as:

[0106]

[0107] The calculated texture variation coefficient TVC = 0.2998, which represents the intensity of texture variation within this local window. A higher TVC value means there are obvious texture differences in this area, indicating that the texture features in the area have changed significantly.

[0108] S203: According to the texture variation coefficient, identify the gray difference within the pooling window, call the global contrast parameter to evaluate the brightness change, and obtain the multi-level feature distribution value;

[0109] After the texture area is determined, further global contrast evaluation is carried out to identify the gray difference within the pooling window and evaluate the brightness change of the image. The global contrast parameter is calculated by statistically analyzing the gray value distribution of the entire image, and the standard deviation is used as the contrast metric. Assume the gray value range of the image is [0, 255], and the standard deviation σ of the image gray can be calculated. Assume the calculation result is 20, indicating that the gray change range of the image is relatively wide. According to the global contrast parameter, the gray differences of each pooling window can be compared and screened. If the gray change of a certain pooling window exceeds the set contrast threshold (for example, the set threshold is 15), it is considered that there is a significant brightness change in this area. Calculate the feature distribution value of each pooling window, and further analyze the multi-level features of the image based on this value. Assume the gray difference of a certain pooling window is 18, which exceeds the contrast threshold of 15, then the brightness change of this area can be identified based on this difference, and the multi-level feature distribution value can be obtained accordingly, thereby realizing the refined recognition of image features.

[0110] Specifically, as Figure 4 shown, the steps of the image feature contribution ratio are specifically as follows:

[0111] S301: Call the multi-level feature distribution values, analyze the gray-scale change in the nodule edge region, identify the edge sharpness based on the pixel gradient direction, and obtain the nodule edge sharpness parameter;

[0112] By analyzing the gray-scale change in the nodule edge region, first extract the pixel data in this region, calculate the gray-scale value change pixel by pixel, and judge the sharpness of its edge based on the change rate of the gray scale. Obtain the gray-scale value distribution in the nodule edge region. The gray-scale value will be calculated based on the local change of the pixel, and the corresponding gradient value represents the sharpness of the image at this position. By calculating the change direction of the gradient, the edge sharpness in the image can be identified, thus providing a necessary basis for subsequent image feature analysis. For example, when analyzing a pulmonary nodule in a chest CT image, extract the gray-scale distribution of the pixels around the nodule, and then calculate its gradient value. If the gray-scale change at the nodule edge is large, it means the edge is relatively sharp, otherwise it is blurred. The edge sharpness parameter is calculated through the above gradient change.

[0113] S302: For the nodule edge sharpness parameter, calculate the mean square deviation of the pixel gradient values in the local window, screen the pixel points that meet the feature adjustment according to the gradient change, call the edge sharpness ratio calculation rule, adjust the weight of the abnormal gradient points, update the pixel gradient data, and obtain the local gradient sharpness ratio;

[0114] First, by selecting a local window area, calculate the mean square deviation of the pixel gradient values in this area to obtain the gradient fluctuation degree in this area. By comparing with the gradient value, screen out the pixel points that meet the feature adjustment standard, and then perform weight adjustment. For the pixel points with abnormal gradients, perform dynamic adjustment according to the edge sharpness ratio calculation rule. For example, if the gradient value of a certain pixel point is relatively low compared with its surrounding points, it means that its weight in the image analysis needs to be increased to avoid misclassification. By adjusting the weight, recalculate the gradient weight of the pixel point and update its gradient data. In the calculation rule of the edge sharpness ratio, it involves calculating a specific weight value of the edge region by comparing the edge sharpness with the surrounding region. Suppose the gradients of some pixels in the local window are 0.2, while the gradient values of other pixels are 0.8, and the mean square deviation is 0.1. By calculation, some abnormal values are found, and their weights are adjusted and updated to more reasonable values.

[0115] S303: Based on the local gradient sharpness ratio, identify the classification error feedback amount, call the feature weight adjustment coefficient, update the feature contribution ratio according to the weight deviation, and obtain the image feature contribution ratio;

[0116] Identify the classification error feedback amount. For the error, further calculations are performed according to the feature weight adjustment coefficient. According to the weight deviation between the current feature and the target category, the feature contribution ratio is dynamically adjusted. The feature contribution ratio is an index to measure the importance of each feature in the classification decision. By adjusting the weight coefficient, the classification result can be optimized. For example, in the analysis process of lung nodules, assuming that the image feature contribution of a certain type of nodule is too large, resulting in a high classification error, the contribution ratio of this feature will be adjusted at this time, reducing the weight of this feature and weakening its influence on the classification result, thereby improving the classification accuracy. For the setting of the feature weight adjustment coefficient, the initial coefficient value is first set according to the experimental data. For example, the initial coefficient is set to 1.0. When the deviation is greater than a certain set threshold (such as 0.5), the coefficient is adjusted to 0.8 until the error reaches a smaller range, and the image feature contribution ratio is output. The accuracy of the optimized classification result will increase accordingly.

[0117] Specifically, as Figure 5 shown, the steps of the nodule feature similarity score are specifically as follows:

[0118] S401: Invoke the image feature contribution ratio, set the region division standard according to the gray value change of the feature region, analyze the pixel gradient range, and obtain the region division parameters;

[0119] First, it is necessary to set the division standard for the feature region in the image. Based on this, the image is segmented and analyzed. To complete the task, first, the gray value of each pixel point in the region is analyzed. By setting the standard (for example, when the gray value difference is less than a certain threshold, it is considered to belong to the same region), by calculating the gray value change range, the corresponding region division parameters are obtained. The parameters determine the size, shape, and distribution law of the image region. In actual operation, the gray value of the image can be statistically analyzed. For example, the mean, variance, minimum value, and maximum value of the gray value of each pixel point are calculated, and then the division standard is set through the gray value change range. For example, when the deviation of the gray mean value of a region from the mean value of its neighboring region exceeds the set threshold (such as 5), it is considered that the region needs to be re-divided. In the calculation process, the gray value can be measured by the pixel gradient change, and a linear or non-linear function is used to judge the region change. Determining the boundary of different feature regions in the image is crucial. Through the set division standard, the division parameters of each region are obtained and refined to form the region division standard, which is convenient for the subsequent execution of image segmentation.

[0120] S402: Based on the region division parameters, invoke the pixel gradient information within the local window, perform nodule local feature segmentation according to the feature change boundary, screen the pixel points that meet the division standard, and obtain the nodule local segmentation features;

[0121] Further analyze the local windows in the image, extract the pixel gradient information therein, and perform local feature segmentation of nodules. First, according to the set region division parameters, the image is cut into multiple small local windows. For each local window, calculate the pixel gradient information within the window. This task can be completed by calculating the gray level change rate of pixel points. For example, calculate the gradient within each local window by using the Sobel operator or the Laplacian operator. The gradient value reflects the intensity of feature changes within the region. Based on the change boundary of nodule features, use the gradient information to perform feature segmentation within the local window. If the gradient change exceeds a certain set threshold (such as 10), it is considered that there is an obvious boundary change in this region, which is the boundary of the nodule. By analyzing and extracting the gradient information of each local window, filter out the pixel points that meet the division criteria. By traversing each local window in the image in a loop, calculate the gray level of each pixel point and analyze its gradient change to determine whether the pixel point meets the feature division criteria of the nodule region. Finally, extract the pixel points that meet the criteria as the local features of the nodule.

[0122] S403: According to the local segmentation features of nodules, analyze the cosine similarity between regional feature vectors, filter out similar regions, and obtain the nodule feature similarity score;

[0123] The nodule feature similarity score uses the formula:

[0124]

[0125] where S represents the nodule feature similarity score, A k represents the k-th feature vector component of region A, B k represents the k-th feature vector component of region B, and M represents the number of dimensions of the feature vector;

[0126] Detailed explanation of the formula and the derivation process of formula calculation:

[0127] This formula is used to calculate the cosine similarity between regional feature vectors A and B, which is specifically realized through summation operation, squaring, square root operation, and division operation;

[0128] S: The regional feature similarity score, which measures the similarity between region A and region B, and the value range is [0,1]. The closer the value is to 1, the higher the similarity;

[0129] A k : The k-th feature vector component of region A, which is obtained by segmenting the image of region A and extracting the features (such as gray level, texture, etc.) of each region. Suppose the texture feature of region A is monitored and analyzed, A k =(0.5, 0.7, 0.4) (for simplicity of example, a three-dimensional feature vector is taken);

[0130] B k : The k-th eigenvector component of region B. The texture features of region B are obtained in a similar manner. Assume the feature of region B is B k =(0.6, 0.6, 0.5);

[0131] M: The dimension number of the eigenvector, representing the number of features of regions A and B in the same dimension. In this example, it is set as a 3-dimensional eigenvector;

[0132] First, calculate the absolute value of the difference between the eigenvectors of regions A and B item by item:

[0133] |A1 - B1| = |0.5 - 0.6| = 0.1;

[0134] |A2 - B2| = |0.7 - 0.6| = 0.1;

[0135] |A3 - B3| = |0.4 - 0.5| = 0.1;

[0136] Then, calculate the sum of the absolute differences:

[0137] Next, calculate the sum of squares of the eigenvectors of regions A and B:

[0138]

[0139] Then, calculate the square root value:

[0140]

[0141] Then, substitute the above results into the formula:

[0142]

[0143] Finally, the nodule feature similarity score S is approximately 0.321. This value indicates that the feature similarity between region A and region B is low, because a low similarity score means that the two regions have significant differences in texture or extracted features.

[0144] Specifically, as Figure 6 shown, the steps of the thyroid nodule image classification adjustment result are specifically as follows:

[0145] S501: Call the nodule feature similarity score, calculate the mean value of the eigenvectors of the category samples, analyze the category boundary shift based on the mean value difference, screen the regions where the shift exceeds the threshold, and obtain the classification boundary adjustment parameters;

[0146] First, it is necessary to divide the feature regions by setting standards, and based on this, perform image analysis and segmentation. When setting the standards for gray-scale changes, it is necessary to consider the degree of gray-scale change of each pixel point in the image, calculate the gray-scale value range of each pixel point, and determine whether it belongs to the same region according to the difference between this value and the mean value of neighboring pixels. For example, statistical quantities such as the mean value and variance can be used to calculate the difference between the gray-scale value of a pixel point and its neighborhood. If the mean difference between the pixel gray-scale values in a certain region and its neighborhood is greater than the set threshold (such as 5), it is considered that there are significant changes in this region and it should be divided into different regions. To further accurately divide the regions, the range of gray-scale value changes or gradient changes is used for definition. When the gray-scale change of a pixel exceeds the set threshold (such as 10), it can be determined that there is an obvious boundary change in a certain region of the image, thus serving as the basis for the division standard. For example, in some medical images, the delineation of regional boundaries is particularly important. Such standards help to distinguish the characteristics of nodules from the surrounding tissues. By setting the threshold, the parameters of each region are obtained, thus providing an accurate division basis for subsequent image segmentation operations.

[0147] S502: Adjust the parameters based on the classification boundary, calculate the classification error offset value of the category region, call the error feedback data, screen the sample points with errors greater than the threshold, correct the category discrimination weight according to the error distribution gradient, adjust the category discrimination weight allocation ratio, optimize the classification loss calculation method, and obtain the optimized classification loss parameters;

[0148] Apply the local window analysis technique in the image to extract the pixel gradient information therein, and then perform the segmentation of local features. To complete the task, first, the image needs to be cut into multiple local windows according to the region division parameters. The pixel gradient change within each window reflects the feature change within this region. The Sobel operator is used to calculate the horizontal and vertical gradients of each local window of the image, which can reflect the edge strength within this region of the image. Assume that within a certain local window, the pixel gray-scale value changes relatively violently and the gradient is greater than the set threshold (such as 10), then it is considered that the local region is the boundary of the nodule. When analyzing the gradient information of each local window, it is necessary to screen the pixel points that meet the standards according to the boundary conditions of nodule feature changes. For example, if the gradient change in a certain region is greater than the set boundary change threshold, then by calculating the gradient information and screening out the pixel points that meet the standards, the local features of the nodule are finally extracted. This process depends on the intensity of gradient changes within the image. The gradient analysis of all local windows should traverse each window of the image through a loop to analyze and judge the gray-scale and gradient of each pixel point, and select the regions that meet the nodule feature requirements.

[0149] S503: Analyze the error adjustment amount based on the optimized classification loss parameters, call the correction factor to update the category division boundary, adjust the boundary coefficient according to the feature distribution, and obtain the classification adjustment result of the thyroid nodule image;

[0150] Analyze the similarity between features, calculate the cosine similarity of regional feature vectors to screen for similar regions. Convert the local features of each region into feature vectors, which are composed of pixel gradients, grayscale means, texture information, etc. To reduce the computational complexity, the local feature vectors of the image can be dimensionally reduced by the principal component analysis (PCA) method, retaining the most discriminative feature components, and calculate the cosine similarity between different regional feature vectors. The formula is as follows:

[0151] where cosine similarity represents the cosine similarity, A and B are the feature vectors of two regions, the dot product operation (·) reflects the similarity between the vectors, ∥A∥ and ∥B∥ are the norms of the vectors. Through calculation, the closer the obtained cosine similarity value is to 1, the more similar the features of these two regions are. For example, assume that the feature vectors of two regions are A = [0.4, 0.6, 0.1] and B = [0.5, 0.7, 0.3] respectively. The similarity value can be calculated through the formula. If its cosine similarity value is greater than the set threshold (e.g., 0.9), it is considered that these two regions have high similarity in features, so as to screen out the regions similar to the nodule features. Through such analysis, the nodule feature similarity score is finally obtained, and this score can provide a basis for subsequent classification and diagnosis.

[0152] As Figure 7 shown, a thyroid nodule ultrasound image classification system for feature extraction, the system includes:

[0153] The ultrasound image processing module is based on the thyroid nodule ultrasound image data, analyzes the grayscale range, adjusts the pixel grayscale value, identifies the gradient change rate to screen out noise points, calls local smoothing to calculate the pixel equalization degree, and obtains a smoothed and normalized image;

[0154] The edge detection module is based on the smoothed and normalized image, sets a gradient change threshold to screen out edge points, identifies the pixel density to obtain the edge region, calculates the direction change rate to evaluate the edge sharpness, calls local pixel statistics to analyze the edge grayscale uniformity, identifies the contrast mutation points to screen out the boundary transition zone, and calls morphological stability to calculate the boundary symmetry degree to obtain the nodule edge gradient distribution;

[0155] The nodule contour analysis module is based on the nodule edge gradient distribution, calls the contour tracking data to calculate the connectivity, analyzes the edge mutation points to screen out the contour curvature changes, and obtains the nodule contour morphological parameters;

[0156] The regional feature calculation module is based on the nodule contour morphological parameters, calculates the regional grayscale mean and variance, calls morphological measurement to identify the regional stretching degree, and obtains the nodule regional feature parameters;

[0157] The classification structure optimization module calculates the classification boundary value based on the nodule region feature parameters, calls the feature matching to analyze the category similarity, corrects the classification error range, optimizes the classification determination parameters, and obtains the classification adjustment result of the thyroid nodule image.

[0158] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A feature extraction method for classifying thyroid nodules ultrasound images, characterized in that: The following steps are involved: S1: Based on the ultrasound image data of thyroid nodules, the pixel normalization function is called to adjust the grayscale range to the set interval, the isolated noise area is screened and smoothed, and the ultrasound image after smoothing is obtained; S2: Based on the smoothed filtered ultrasound image, a small-scale convolution kernel is used to analyze the pixel gradient direction, the edge change area is screened, a medium-scale filter is applied to identify texture changes, redundancy is reduced through pooling operations, a global contrast parameter is set to evaluate brightness changes, and multi-level feature distribution values ​​are obtained; S3: calling the multi-level feature distribution value, calculating the nodule edge sharpness parameter, setting the mean square error calculation rule of the edge gradient, obtaining the nodule edge sharpness ratio, adjusting the feature weight according to the classification error feedback, and obtaining the image feature contribution ratio; S4: calling the image feature contribution ratio, setting the image region division standard, performing nodule local feature segmentation, applying cosine homogeneity analysis to screen homogeneous regions, and obtaining a nodule feature homogeneity score; S5: calling the nodule feature homogeneity score, setting classification boundary adjustment rules, analyzing classification error offset, optimizing the classification loss measurement method, and obtaining thyroid nodule image classification adjustment results.

2. The method for classifying thyroid nodules by ultrasonic image extraction based on feature extraction according to claim 1, characterized in that: The smoothed filtered ultrasound image includes a grayscale normalized image, a denoised image, and a smoothed filtered image. The multi-level feature distribution values ​​include edge change features, texture change features, and brightness change features. The nodule edge sharpness ratio includes an edge sharpness parameter, an edge gradient mean square error, and a feature weight ratio. The image feature contribution ratio includes a region division standard, a local feature segmentation criterion, and a homogeneity analysis parameter. The nodule feature homogeneity score includes a local homogeneity score, a global homogeneity score, and a cosine homogeneity value. The thyroid nodule image classification adjustment result includes a classification boundary parameter, an error offset analysis value, and a classification loss optimization parameter.

3. The feature-extracted thyroid nodule ultrasound image classification method according to claim 1, characterized in that: The step of smoothing the filtered ultrasonic image is specifically as follows: S101: Based on the ultrasound image data of the thyroid nodule, analyzing the pixel grayscale range, performing linear transformation according to the normalization scale factor, adjusting the pixel grayscale value to a set range, and obtaining normalized grayscale image data; S102: Based on the normalized grayscale image data, detect isolated noise areas, calculate the grayscale mean and standard deviation of pixels in a local window, filter abnormal pixels according to a noise judgment threshold, call the grayscale mean of adjacent pixels to replace abnormal pixels, adjust the local abnormal grayscale distribution, and obtain image data after noise removal; S103: Calculate the weighted mean of pixels in a local window for the image data after noise removal, perform a smoothing operation according to a weighting coefficient, update the ultrasonic image data, and obtain a smoothed filtered ultrasonic image.

4. The method for classifying thyroid nodules by ultrasonic image extraction based on feature extraction according to claim 1, characterized in that: The steps of multi-level feature distribution value are specifically as follows: S201: Based on the smoothed and filtered ultrasound image, a small-scale convolution kernel is called to calculate a pixel gradient value, and an edge area is screened according to a change in gradient direction to obtain an edge gradient distribution value; S202: Analyze the grayscale variation in the local window according to the edge gradient distribution value, select the texture area according to the gradient change rate, call the set threshold to determine the pixel points where the grayscale difference exceeds the standard, update the pixel matrix and adjust the local grayscale distribution to obtain the texture variation coefficient; S203: According to the texture variation coefficient, the grayscale difference in the pooling window is identified, and the global contrast parameter is called to evaluate the brightness change to obtain a multi-level feature distribution value.

5. The method for classifying thyroid nodules by ultrasonic image extraction based on feature extraction according to claim 4, characterized in that: The texture variation coefficient adopts the formula: Among them, TVC represents the texture variation coefficient, N represents the total number of pixels in the local window, and ΔG i represents the gradient change of the i-th pixel, θ represents the set gradient change threshold, G i represents the gray value of the i-th pixel, G0 represents the gray value of the pixel at the center of the window, and ∈ is a constant to avoid division by zero errors.

6. The feature-extracted thyroid nodule ultrasound image classification method according to claim 1, characterized in that: The step of image feature contribution ratio is specifically as follows: S301: calling the multi-level feature distribution value, analyzing the grayscale change of the nodule edge area, identifying the edge sharpness according to the pixel gradient direction, and obtaining the nodule edge sharpness parameter; S302: for the nodule edge sharpness parameter, calculate the mean square error of the pixel gradient value in the local window, select the pixel points that meet the feature adjustment according to the gradient change, call the edge sharpness ratio calculation rule, adjust the weight of the abnormal gradient point, update the pixel gradient data, and obtain the local gradient sharpness ratio; S303: According to the local gradient sharpness ratio, the classification error feedback amount is identified, the feature weight adjustment coefficient is called, the feature contribution ratio is updated according to the weight deviation, and the image feature contribution ratio is obtained.

7. The feature-extracted thyroid nodule ultrasound image classification method according to claim 1, characterized in that: The steps of nodule feature homogeneity scoring are specifically as follows: S401: calling the image feature contribution ratio, setting a region division standard according to the grayscale change of the feature region, analyzing the pixel gradient range, and obtaining region division parameters; S402: Based on the region division parameters, pixel gradient information in the local window is called, local feature segmentation of nodules is performed according to feature change boundaries, pixel points meeting the division criteria are screened, and local segmentation features of nodules are obtained; S403: Analyze the cosine homogeneity between regional feature vectors based on the nodule local segmentation features, select similar regions, and obtain nodule feature homogeneity scores.

8. The feature-extracted thyroid nodule ultrasound image classification method according to claim 7, characterized in that: The nodule feature homogeneity score is scored using the formula: Among them, S represents the nodule feature homogeneity score, A k represents the kth eigenvector component of region A, B k represents the kth eigenvector component of region B, and M represents the dimension of the eigenvector.

9. The method for classifying thyroid nodules by ultrasonic image extraction based on feature extraction according to claim 1, characterized in that: The steps of classifying and adjusting the results of the thyroid nodule images are specifically as follows: S501: calling the nodule feature homogeneity score, calculating the feature vector mean of the category samples, analyzing the category boundary offset according to the mean difference, screening the area where the offset exceeds the threshold, and obtaining the classification boundary adjustment parameter; S502: Based on the classification boundary adjustment parameter, the classification error offset value of the category area is calculated, the error feedback data is called, the sample points whose errors are greater than the threshold are screened, the category distinction weight is corrected according to the error distribution gradient, the category distinction weight distribution ratio is adjusted, the classification loss calculation method is optimized, and the optimized classification loss parameter is obtained; S503: Analyze the error adjustment amount according to the optimized classification loss parameter, call the correction factor to update the category division boundary, adjust the boundary coefficient according to the feature distribution, and obtain the thyroid nodule image classification adjustment result.

10. A feature extraction system for classifying thyroid nodules ultrasound images, characterized in that: According to the feature extraction method for thyroid nodule ultrasound image classification according to any one of claims 1 to 9, the system comprises: The ultrasound image processing module analyzes the grayscale range, adjusts the pixel grayscale value, identifies the gradient change rate to filter out noise points, and uses local smoothing to calculate pixel balance to obtain a smoothed normalized image based on the ultrasound image data of thyroid nodules. The edge detection module is based on the smoothed normalized image, sets a gradient change threshold to filter edge points, identifies pixel density to obtain edge areas, calculates direction change rate to evaluate edge sharpness, calls local pixel statistics to analyze edge grayscale uniformity, identifies contrast mutation points to filter boundary transition zones, calls morphological stability to calculate boundary symmetry, and obtains nodule edge gradient distribution; The nodule contour analysis module uses contour tracking data to calculate connectivity based on the nodule edge gradient distribution, analyzes edge mutation points to screen contour curvature changes, and obtains nodule contour morphological parameters; The regional feature calculation module calculates the regional grayscale mean and variance based on the nodule contour morphological parameters, calls the morphological measurement to identify the extension degree of the region, and obtains the nodule regional feature parameters; The classification structure optimization module calculates the classification boundary value based on the nodule region characteristic parameters, calls feature matching to analyze the category homogeneity, corrects the classification error range, optimizes the classification judgment parameters, and obtains the thyroid nodule image classification adjustment result.

Images