Image feature recognition method and system applied to early kidney disease

Abstract

The invention relates to the technical field of image recognition, and discloses an image feature recognition method and system applied to the early stage of kidney diseases. According to the method, the features in the original ultrasonic image are extracted in parallel, and the micro texture change and the macroscopic form information shown in the abnormal area can be captured respectively, so that the limitation of a traditional single-scale feature extraction method in coping with diffusive and weak-saliency image features is overcome; the local texture feature map and the global structure feature map are analyzed, quantization and positioning of weak texture changes are achieved, priori knowledge is converted into objective indexes, pixel-level fusion is carried out on a texture abnormal map and a region importance weight map, and through the knowledge-guided and data-driven deep fusion method, the deep fusion of the texture abnormal map and the region importance weight map is achieved. It is ensured that the finally recognized abnormal area is not only a statistical outlier but also conforms to importance distribution of an image structure, and high-precision automatic recognition of weak and diffuse features in the kidney medical image is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to an image feature recognition method and system applied to early kidney disease. BACKGROUND

[0002] Ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI) and other technologies can non-invasively obtain kidney images, clearly presenting their morphological structure, tissue boundary and internal echo or signal distribution, among which ultrasound images are more widely used due to their real-time, convenient and non-radiation characteristics. Early image feature recognition of the kidney refers to a technical process of extracting, quantifying and judging feature changes existing in the kidney region based on such medical images, comprehensively utilizing image processing and pattern recognition technology. This process focuses on the visual attributes presented by the image itself, aiming to establish a mapping model from image data to feature description, to realize objective recognition and quantitative evaluation of abnormal regions in kidney images, and to provide stable and interpretable image basis for further analysis. Therefore, early feature recognition of kidney images is an important research direction in medical image analysis, and has a positive significance for promoting the combination of image intelligent processing technology and clinical application needs.

[0003] Since early medical images (such as ultrasound, CT) of kidney diseases are mostly characterized by diffuse, global and weak feature changes, such as very slight enhancement of renal cortical echo and slight blurring of the cortex and medulla boundary, such changes are extremely similar to normal tissue structure variations and image noise differences in vision, and lack obvious boundaries or focal regions, thereby bringing significant challenges to image recognition. The existing technology in identifying early images of kidney diseases either adopts an image analysis method with manually preset fixed parameters, or uses a general deep learning model, resulting in a mismatch between the image feature extraction method and the target characteristics, and the model focus deviates from the key region. This mismatch directly leads to insufficient sensitivity and specificity of the recognition of target features in early images, and further leads to poor recognition effect and low recognition accuracy. SUMMARY

[0004] The main purpose of the present application is to provide an image feature recognition method and system applied to early kidney disease, aiming to solve the technical problems in the prior art.

[0005] The present application provides an image feature recognition method applied to early kidney disease, comprising: Obtaining an original ultrasound image of the kidney, and performing feature extraction on the original ultrasound image to obtain a local texture feature map and a global structure feature map; Analyzing the local texture feature map to obtain a texture abnormality map and an abnormal region coordinate set; analyzing the global structure feature map to obtain a cortex-medulla boundary definition and a region importance weight map; performing pixel-level fusion on the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map; obtaining a comprehensive evaluation feature vector according to the spatially weighted anomaly map and the abnormal region coordinate set, and obtaining a high-confidence abnormal region set according to the comprehensive evaluation feature vector; screening each high-confidence abnormal region in the high-confidence abnormal region set through the cortex-medulla boundary definition to obtain a significant abnormal region set; analyzing the significant abnormal region set to obtain a comprehensive risk index, and generating a feature recognition result of the kidney image according to the comprehensive risk index.

[0006] Preferably, the step of analyzing the local texture feature map to obtain a texture anomaly map and an abnormal region coordinate set comprises: obtaining a plurality of neighborhood window corresponding scalar feature values from the local texture feature map, and concatenating the scalar feature values in order to obtain a corresponding multi-dimensional feature vector; obtaining a normal texture template of the kidney, and extracting a standard feature vector corresponding to each neighborhood window from the normal texture template of the kidney; calculating a Euclidean distance corresponding to each neighborhood window based on the standard feature vector and the multi-dimensional feature vector, and constructing an initial difference map based on the Euclidean distance; processing the initial difference map using an adaptive threshold segmentation algorithm to obtain a segmentation threshold corresponding to each neighborhood window, and comparing the Euclidean distance of each pixel with the segmentation threshold corresponding thereto to generate a binary mask image; performing connected region analysis on the binary mask image to obtain an abnormal region coordinate set; obtaining an abnormal saliency score of each pixel point according to the initial difference map and the abnormal region coordinate set, and constructing a texture anomaly map based on the abnormal saliency score.

[0007] Preferably, the step of analyzing the global structure feature map to obtain a cortex-medulla boundary definition and a region importance weight map comprises: performing Gaussian smoothing on the global structure feature map to obtain an enhanced structure map, and performing region growing and contour extraction on the enhanced structure map using a region growing algorithm to obtain a region identification map; extracting a pixel label value corresponding to each pixel point from the region identification map, and screening the pixel points based on the pixel label value to obtain an initial pixel set; An edge tracing algorithm is used to sort and connect all pixels in the initial pixel set to obtain the boundary line contour; The global structural feature map is processed using an edge gradient analysis method. By quantifying the average edge intensity of the boundary line contour, the clarity of the corticomedullary boundary is obtained. The importance weight value corresponding to each pixel is obtained based on the pixel label value, and a regional importance weight map is generated based on the importance weight value.

[0008] Preferably, the step of using a region growing algorithm to perform region growing and contour extraction on the enhanced structure map to obtain a region identifier map includes: Extract the pixel coordinate range from the enhanced structure map, and obtain the cortical seed point coordinates and medullary seed point coordinates based on the pixel coordinate range; Extract the cortical pixel intensity value and medullary pixel intensity value corresponding to the coordinates of the cortical seed point and the coordinates of the medullary seed point, respectively, from the enhanced structure map; Based on the cortical pixel intensity value and the medullary pixel intensity value, a region growing algorithm is used to iteratively expand the coordinates of the cortical seed point and the medullary seed point to obtain the cortical region coordinate set and the medullary region coordinate set. An initial binary labeled image is obtained based on the coordinate set of the cortical region and the coordinate set of the medullary region, and the initial binary labeled image is processed by morphological closing operation to obtain a binary segmentation image; An edge detection algorithm is used to extract the contours of the binary segmented image to obtain a contour coordinate set, and a region identification map is constructed based on the contour coordinate set, the cortical region coordinate set, and the medullary region coordinate set.

[0009] Preferably, the step of performing pixel-level fusion of the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map includes: Extract the anomaly saliency score of each pixel from the texture anomaly map, and obtain the local anomaly fluctuation value of each pixel based on the anomaly saliency score; The importance weight value of each pixel is extracted from the regional importance weight map, and the local abnormal fluctuation value is adaptively weighted and corrected based on the importance weight value to obtain the local salience enhancement value of each pixel; Obtain the core threshold parameter and slope control parameter, and obtain the intermediate fusion feature value of each pixel based on the core threshold parameter, slope control parameter and local saliency enhancement value; Obtain the fusion weight coefficients, and perform weighted fusion processing on the fusion weight coefficients, intermediate fusion feature values ​​and the anomaly significance scores to obtain spatial weighted anomaly intensity values; A spatially weighted anomaly map is constructed based on the spatially weighted anomaly intensity values.

[0010] Preferably, the step of obtaining a comprehensive evaluation feature vector based on the spatially weighted anomaly map and the anomaly region coordinate set, and obtaining a high-confidence anomaly region set based on the comprehensive evaluation feature vector, includes: Extract the set of anomaly intensity values ​​corresponding to each anomaly region from the spatially weighted anomaly map coordinate set, and obtain the region feature vector based on the set of anomaly intensity values; Pixel marker values ​​of all pixels within each abnormal region are extracted from the binarized mask image, and morphological feature vectors of each abnormal region are obtained based on the pixel marker values. Extract the center coordinates of each abnormal region from the set of abnormal region coordinates; The importance weight map of the region is processed using a skeletonization algorithm to obtain the renal cortex centerline, and the cortical proximity of each abnormal region is obtained based on the center coordinates and the renal cortex centerline. The region feature vector, morphological feature vector, and cortical proximity are concatenated to obtain a comprehensive evaluation feature vector; The confidence score of each anomalous region is obtained based on the comprehensive evaluation feature vector, and a set of high-confidence anomalous regions is constructed based on the confidence score.

[0011] Preferably, the step of analyzing the set of significantly abnormal regions to obtain a comprehensive risk index includes: Extract the spatially weighted anomaly intensity value corresponding to all pixels in each of the significant anomaly regions from the set of significant anomaly regions, and obtain the global anomaly cumulative load based on the spatially weighted anomaly intensity value; Extract the area of ​​each significant abnormal region and its cortical proximity from the set of significant abnormal regions, and perform statistical analysis on the area of ​​all the abnormal regions to obtain the total area of ​​the abnormal regions. The total area of ​​renal parenchyma is obtained, and the total area of ​​renal parenchyma, the total area of ​​abnormal regions, and the global abnormal cumulative load are fused to obtain a composite abnormality index. Obtain the attenuation coefficient, and transform the cortical proximity using an exponential decay function based on the attenuation coefficient to obtain the spatial distribution risk coefficient; Obtain the normalized coefficient, and integrate the normalized coefficient, composite anomaly index, spatial distribution risk coefficient, and global anomaly cumulative load to obtain a comprehensive risk index.

[0012] This application also provides an image feature recognition system for early-stage kidney disease, comprising multiple modules for implementing the steps of the image feature recognition method for early-stage kidney disease described above.

[0013] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described image feature recognition method applied to the early stage of kidney disease.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described image feature recognition method for early-stage kidney disease.

[0015] The beneficial effects of this invention are as follows: This invention extracts features from the original ultrasound image through two parallel deep convolutional neural network branches, obtaining local texture feature maps and global structural feature maps. This effectively captures feature information at different scales in the image, including subtle texture changes and overall morphological structure, overcoming the limitations of traditional single-scale feature extraction methods in dealing with diffuse and weakly salient image features. Then, by analyzing the local texture feature maps, a texture anomaly map and anomaly region coordinate set are generated, realizing the quantification and localization of subtle texture changes in the image. Simultaneously, by analyzing the global structural feature map, the corticomedullary boundary clarity and regional importance weight maps are obtained, thus transforming prior information about image structure into computable weight parameters, enhancing the interpretability of the model. Finally, the texture anomaly map and regional importance weight maps are fused at the pixel level, through this knowledge-guided and data-driven approach... The dynamic deep fusion method ensures that the abnormal regions identified by the algorithm are not only statistically significant but also conform to the importance distribution of the image structure. Based on this, a comprehensive evaluation feature vector is obtained from the spatially weighted abnormal map and the coordinate set of abnormal regions, and input into the pre-trained region evaluation model to obtain a set of high-confidence abnormal regions. Then, the high-confidence abnormal regions are further filtered through global structural evidence, such as the clarity of the corticomedullary boundary, to obtain a set of significant abnormal regions. This mechanism modulates local judgments probabilistically through global evidence, which can effectively improve the system's ability to identify real abnormal regions in the image. Finally, by performing multi-dimensional analysis on the set of significant abnormal regions and generating a comprehensive risk index, a quantitative and interpretable kidney image recognition result is output. This method, which closely combines image features with prior knowledge, is beneficial to improving the accuracy and reliability of identifying weak and diffuse features in kidney ultrasound images. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of this application.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] like Figure 1 As shown, this application provides an image feature recognition method for early-stage kidney disease, comprising: S1. Obtain the original ultrasound image of the kidney, and extract features from the original ultrasound image to obtain a local texture feature map and a global structural feature map; S2. Analyze the local texture feature map to obtain a texture anomaly map and an anomaly region coordinate set; S3. Analyze the global structural feature map to obtain the clarity of the corticomedullary boundary and the weighting of regional importance; S4. Perform pixel-level fusion of the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map; S5. Obtain a comprehensive evaluation feature vector based on the spatially weighted anomaly map and the anomaly region coordinate set, and obtain a high-confidence anomaly region set based on the comprehensive evaluation feature vector; S6. Each high-confidence anomalous region in the high-confidence anomalous region set is screened based on the clarity of the corticomedullary boundary to obtain a set of significant anomalous regions; S7. Analyze the set of significant abnormal regions to obtain a comprehensive risk index, and generate feature recognition results of the kidney image based on the comprehensive risk index.

[0022] As described in steps S1-S7 above, this invention extracts features from the original ultrasound image through two parallel deep convolutional neural network branches: the first branch uses small-scale convolutional kernels to focus on local texture patterns and outputs a local texture feature map; the second branch uses large-scale convolutional kernels and pooling operations to capture the overall structure of the organ and outputs a global structure feature map. The original ultrasound image refers to a two-dimensional medical image acquired through an ultrasound scanning device, recording the internal structure and acoustic properties of the kidney in grayscale or color pixel matrix form. The local texture feature map refers to the image used to characterize the microstructure of kidney tissue (such as renal tubules and glomeruli). The regularity, uniformity, and roughness of the array are characteristic representations of texture attributes. The global structural feature map refers to the feature representation used to describe the overall morphology of the kidney, boundary integrity, and the spatial relationships and geometric attributes between the main internal structures (such as the cortex, medulla, and collecting system). This allows for the capture of micro-texture changes and macro-morphological information at different scales in the image, laying the foundation for subsequent analysis. This helps overcome the limitations of traditional single-scale feature extraction methods in dealing with diffuse and weakly salient image features. Then, by analyzing the local texture feature map, a texture anomaly map and anomaly region coordinate set are generated, realizing the micro-texture changes and macro-morphological information at different scales in the image. The quantification and localization of weak texture variations, where the texture anomaly map refers to the map in which the value of each pixel in the image represents the probability or significance of the tissue texture deviating from the normal reference standard at that location, and the anomaly region coordinate set refers to the set containing the geometric location and range information of all detected spatially continuous suspicious anomaly regions. Because traditional fixed-scale texture operators (such as LBP and GLCM) cannot adaptively capture the most discriminative weak information in images, and because general deep learning models are prone to overfitting under small sample training due to the scarcity of early image annotation data, resulting in insufficient capture of weak features, this invention... By conducting in-depth analysis of local texture feature maps, it is beneficial to solve the problem of insufficient capture of weak features in small samples in traditional methods. At the same time, by analyzing global structural feature maps, corticomedullary boundary clarity and regional importance weight maps are obtained. Corticomedullary boundary clarity is a quantitative indicator used to objectively measure whether the boundary between the renal cortex and medulla is clear and well-defined in images. Regional importance weight maps are spatial mapping maps that assign different confidence weights to different image regions based on prior image structural information. This transforms image structural information into computable objective parameters and enhances the interpretability of the model.

[0023] Since anomalies in images require both texture-level detection and structural importance evaluation, this invention generates a spatially weighted anomaly map by pixel-level fusion of a texture anomaly map and a region importance weight map. The spatially weighted anomaly map is a comprehensive feature map that integrates the saliency of texture anomalies and the structural importance of regions; high-value regions represent areas with both texture anomalies and high structural weights. This knowledge-guided and data-driven deep fusion method ensures that the identified anomaly regions are not only statistical outliers but also conform to the importance distribution of the image structure, thus significantly improving the reliability of the identification results. Based on this, according to the spatial... A comprehensive evaluation feature vector is obtained from the weighted anomaly map and the coordinate set of anomaly regions. The evaluation feature vector is a mathematical vector that fully describes the suspicious anomaly region through multiple dimensions. Then, the comprehensive evaluation feature vector of all anomaly regions is input into a pre-trained regional evaluation model for evaluation, resulting in a high-confidence anomaly region set containing regional coordinates, evaluation feature vectors, and confidence scores. The high-confidence anomaly region set refers to the set of anomaly regions with high confidence obtained after being screened by the machine learning model. Through this multi-dimensional feature fusion and machine learning classification method, the evaluation is improved from pixel level to region level, overcoming the limitations of single intensity features.

[0024] Furthermore, a set of significantly abnormal regions is obtained by secondary screening of high-confidence abnormal regions using global structural evidence, namely, the clarity of the corticomedullary boundary. First, the clarity of the corticomedullary boundary, a global quantitative indicator representing the overall structural state of the kidney, is used as a dynamic weighting factor and fused with the confidence score of each suspicious region. The core logic lies in establishing a correlation between local anomalies and global evidence. When the global boundary clarity is low, the confidence score of suspicious regions located in the renal cortex is systematically increased; conversely, when the global boundary clarity is high, the confidence scores of all suspicious regions are correspondingly decreased. Finally, the algorithm sorts regions according to the revised confidence scores and selects a subset of regions exceeding a preset threshold to form a set of significantly abnormal regions. This mechanism uses global evidence to generalize local judgments. The invention employs probabilistic modulation to enhance the sensitivity of identifying genuine abnormal regions while significantly suppressing false abnormal signals, thereby greatly improving the specificity of image recognition. Finally, by performing multi-dimensional analysis on the set of significant abnormal regions and generating a comprehensive risk index, the invention provides a holistic score reflecting the intensity, spatial range, and distribution pattern of abnormal features in the entire image. This comprehensive risk index can quantify the degree of abnormality in the entire image, facilitating the establishment of a standardized image evaluation process and outputting quantified and interpretable image recognition results. This transforms complex image representations into objective image analysis criteria. This invention innovatively combines image features with prior knowledge, enabling high-precision and high-reliability automatic identification of weak and diffuse features in renal medical images.

[0025] In one embodiment, step S2, which involves analyzing the local texture feature map to obtain a texture anomaly map and anomaly region coordinate set, includes: S21. Obtain scalar feature values ​​corresponding to multiple neighborhood windows based on the local texture feature map, and concatenate the scalar feature values ​​in order to obtain the corresponding multidimensional feature vector; S22. Obtain the normal texture template of the kidney, and extract the standard feature vector corresponding to each neighborhood window from the normal texture template of the kidney. S23. Calculate the Euclidean distance corresponding to each neighborhood window based on the standard feature vector and the multidimensional feature vector, and construct an initial difference map based on the Euclidean distance; S24. The initial difference map is processed using an adaptive threshold segmentation algorithm to obtain the segmentation threshold corresponding to each neighborhood window, and the Euclidean distance of each pixel is compared with its corresponding segmentation threshold to generate a binary mask image. S25. Perform connected component analysis on the binarized mask image to obtain the coordinate set of abnormal regions; S26. Obtain the anomaly saliency score for each pixel based on the initial difference map and the anomaly region coordinate set, and construct a texture anomaly map based on the anomaly saliency score.

[0026] As described in steps S21-S26 above, this invention enhances the saliency of edges and details in the local texture feature map by performing Gaussian difference filtering on the feature map, thus obtaining an edge-enhanced texture feature map. Since subtle texture changes in kidney images are usually manifested as subtle and diffuse signal differences, Gaussian difference filtering, as a multi-scale edge enhancement operator, can suppress low-frequency background noise and highlight high-frequency detail information (such as weak texture irregularities), thereby solving the problem that traditional methods cannot adaptively capture key weak features at a fixed scale. Next, for each pixel in the edge-enhanced texture feature map, a fixed-size window (e.g., 16x16 pixels) is defined with that pixel as the center, and the image area covered by this window is defined as the neighborhood to be analyzed for the current pixel. Then, a series of scalar feature values ​​for quantifying texture features within each neighborhood window are calculated. Here, scalar feature values ​​refer to numerical values ​​extracted from image data to quantify and describe a single texture attribute. The scalar feature values ​​are then concatenated in order to obtain a multi-dimensional feature vector corresponding to each neighborhood window. Here, the multi-dimensional feature vector refers to a vector used to comprehensively describe the local area. The mathematical vector representing the overall texture characteristics of the domain, where each dimension carries different aspects of texture information, is obtained as follows: 1. First, extract the pixel intensity value of the target pixel from the edge-enhanced texture feature map; 2. Within each neighborhood window, perform local binary mode encoding on each pixel (including the target pixel and its surrounding pixels): obtain the grayscale values ​​of the target pixel and its surrounding pixels respectively, and compare its grayscale value with the grayscale values ​​of its surrounding pixels one by one, with the target pixel as the center. If the grayscale value of the surrounding pixels is greater than or equal to the grayscale value of the target pixel, it is marked as "1", otherwise it is marked as "0", thereby generating a set of binary encoding sequences. Then, convert the binary encoding sequences into corresponding decimal values, and calculate the mean, variance, and information entropy based on the decimal values ​​corresponding to all pixels in the neighborhood window. These three statistics together constitute the grayscale statistical components. The grayscale statistical components refer to a set of statistics used to quantify the overall texture distribution pattern of each neighborhood window. Through these statistics, the complex high-dimensional histogram information can be compressed into a few highly representative numerical indicators.Third, it is necessary to analyze the spatial relationships between pixels within each local neighborhood window. For this, a set of standardized offset vectors needs to be pre-defined to define the rules for forming "pixel pairs." Each offset vector uniquely corresponds to a spatial analysis direction. For example, vector (1, 0) corresponds to 0° (horizontal to the right), and vector (1, 1) corresponds to 45° (lower right diagonal). Next, a two-dimensional statistical matrix is ​​obtained. This matrix is ​​a blank matrix whose number of rows and columns is determined by the number of gray levels in the image. All elements in the matrix are zero, and the number of rows and columns equals the possible gray levels of the image. Subsequently, for each of the above-mentioned pre-defined offset directions... (i.e., the direction represented by each offset vector), the following process is executed independently: traverse every position within the local neighborhood window that can form a pixel pair (i.e., according to the current offset vector, the first pixel and its specific neighboring pixels must be within the predefined window boundary). For each pixel pair found, obtain the grayscale values ​​of these two pixels from the edge-enhanced texture feature map, and increment the count by one at the corresponding coordinate position in the two-dimensional statistical matrix to obtain the cumulative co-occurrence frequency matrix. After completing the traversal of the entire local neighborhood window, divide each count value in the cumulative co-occurrence frequency matrix by the total number of pixel pairs to normalize it into a representation of different... The joint probability distribution matrix of grayscale combinations is generated. Then, based on this matrix, the probability value at each position is multiplied by the square of the difference between the row and column numbers at that position. These multiplications are then summed across all positions to obtain the texture contrast, which directly reflects the drasticness of grayscale changes. Simultaneously, the sum of the squares of the probability values ​​at each position is calculated to obtain the texture uniformity, which reflects the uniformity and regularity of the image texture; a higher value indicates a more uniform texture distribution. Next, for each probability value at each position in the matrix, the absolute value of the difference between the corresponding row and column numbers is calculated, and this absolute value is added to 1 to obtain... The distance penalty coefficient is used to calculate the quotient of the probability value and the distance penalty coefficient. The quotients at all positions are summed to obtain the local consistency. The value of the local consistency reflects the consistency of gray values ​​in the local area. The higher the value, the smoother the texture. Finally, the row index (representing the gray level of the first pixel) and column index (representing the gray level of the second pixel) of the matrix are regarded as two random variables. The probability value is regarded as the joint probability of the occurrence of this gray level combination. Correlation analysis is performed on the two gray level variables to obtain the texture correlation. The larger the value, the stronger the directionality and regularity of the image texture. It is used to measure the degree of linear dependence between row elements and column elements in the matrix.Texture contrast, texture uniformity, local consistency, and texture correlation constitute the texture feature index set for each direction. After performing the above calculations in four directions (0°, 45°, 90°, 135°), the arithmetic mean of the calculation results for each texture index in different directions is taken, ultimately generating a set of four rotation-invariant texture spatial relationship components. The obtained multidimensional feature vector is a mathematical vector used to comprehensively describe the overall texture characteristics of a local region. It contains information in multiple dimensions, covering pixel intensity values, gray-level statistical components, and texture spatial relationship components. The gray-level statistical components consist of the mean, variance, and information entropy, used to quantify the overall texture distribution pattern of each neighborhood window. The texture spatial relationship components include texture contrast, texture uniformity, local consistency, and texture correlation. These are rotation-invariant indices obtained by averaging after multi-directional calculations, used to reflect the spatial relationship characteristics between pixels. Each dimension of this vector carries different aspects of texture information, thereby comprehensively capturing the texture features of the local region, which is beneficial for the effective detection and analysis of potential texture abnormalities in kidney images.

[0027] Then, a normal kidney texture template is obtained. This template is a statistical benchmark representing standard normal texture features, trained using a large number of clinically validated healthy kidney ultrasound images. It stores a standard feature vector at each spatial location. This standard feature vector is a reference vector defined at a specific spatial location, representing the standard texture pattern that healthy kidney tissue should have at that location. For each neighborhood window, the Euclidean distance between its multidimensional feature vector and the corresponding standard feature vector in the normal kidney texture template is calculated. This Euclidean distance directly quantifies the overall texture features of each neighborhood window. The overall deviation of normal baseline texture features is determined, and the Euclidean distances corresponding to all neighboring windows are arranged according to their original image spatial positions to generate an initial difference map of the same size as the original image. In this difference map, the grayscale value of each pixel is proportional to the Euclidean distance of its centered neighboring window. The higher the brightness, the greater the difference between the texture of that area and the normal kidney texture. This method of comparing high-dimensional feature vectors window by window can transform diffuse and global weak abnormal changes into a visualized and quantifiable spatial difference map, which is beneficial for highlighting potential abnormal areas and providing a data basis for subsequent quantitative analysis.

[0028] Next, a sliding window technique is used to traverse each pixel in the initial difference map. For the local neighborhood of each pixel, the Otsu method is used to automatically calculate the optimal segmentation threshold for that local region: by maximizing the inter-class variance within the region, the local threshold that can most effectively distinguish the foreground (potential anomalies) from the background (normal tissue) is found, thus overcoming the defect that a globally fixed threshold is difficult to adapt to the contrast changes in different regions of the image. Then, the value of the target pixel in the initial difference map is compared with this dynamically calculated local threshold. If the value is higher than the local threshold, it is determined to be an abnormal pixel and marked as white (value 1) in the mask; otherwise, it is marked as black (value 0). This process is performed pixel by pixel, and finally a binary mask image is generated. This method of obtaining an adaptive threshold through local statistical modeling can significantly enhance the signal sensitivity in low-contrast regions, while suppressing noise interference in uniform background regions, thereby accurately separating statistically significant abnormal pixel clusters. This is beneficial to solving the problem that weak abnormal signals are difficult to reliably segment in non-uniform backgrounds in existing technologies.

[0029] Next, connected component analysis is performed on the binarized mask image: First, a scan-line-based connected component labeling algorithm is used to traverse all pixels in the binarized mask image. This algorithm scans the image line by line, checking the connection relationship between each foreground pixel (white, value 1) and its adjacent pixels in the previous row, and assigning the same unique identifier to interconnected pixels, thereby identifying all independent connected regions. Subsequently, for each labeled connected region, the arithmetic mean of the coordinates of all pixels within that region is calculated, that is, the average of the x and y coordinates of all pixels is calculated to obtain the center coordinates of each connected region, and then the center coordinates of the region are found. The maximum and minimum coordinate values ​​of each connected region in the x and y directions are obtained, and based on these, a minimum rectangle that can completely enclose the region is obtained, i.e., the bounding rectangle parameters of each connected region are obtained. Finally, the unique identifier, center coordinates, and bounding rectangle parameters (top left corner coordinates, width, and height, etc.) of each connected region are combined into a structured dataset. The structured datasets of all connected regions constitute the coordinate set of the anomalous region. Through connected region analysis, discrete anomalous pixels can be aggregated into meaningful image anomalous objects, and their spatial localization and subsequent quantitative analysis are provided by geometric parameterization.

[0030] Finally, for each connected region, the pixel value of each pixel is extracted from the initial difference map, and the arithmetic mean of all pixels within it is calculated to obtain the overall average deviation of the region. The arithmetic mean of all connected regions is then normalized to obtain the anomaly intensity score for each region. Next, each pixel in the initial difference map is traversed. If a pixel belongs to a connected region, its pixel value is fused with the anomaly intensity score of that region: first, the sum of the anomaly intensity score and "1" is calculated, then the product of this sum and the pixel value is calculated to obtain the anomaly significance score. This operation preserves the original pixel-level detail differences and achieves adaptive enhancement based on the overall anomaly level of the region, resulting in generally higher pixel values ​​and preserved internal variations within high-anomaly regions. Since connected region analysis has already been performed... The adaptive thresholding filter removes most random noise and negligible differences. Therefore, pixels not included in any connected regions are considered clinically insignificant anomalies and should be suppressed in subsequent analysis. Thus, if a pixel does not belong to any anomalous region identified by connected region analysis (i.e., it is considered background or noise in the binarized mask image), its anomaly saliency score is directly set to 0 or a very small baseline value close to 0 (e.g., 0.001). Finally, the anomaly saliency scores of all pixels are arranged according to their original coordinate positions to generate a texture anomaly map of the same size as the original image. This achieves adaptive extraction and highlighting of diffuse weak texture features, solves the problem of mismatch between existing feature extraction strategies and weak diffuse image features, and is beneficial to improving the ability to identify image anomalies.

[0031] In one embodiment, step S3, which involves analyzing the global structural feature map to obtain a corticomedullary boundary clarity and regional importance weight map, includes: S31. The global structural feature map is subjected to Gaussian smoothing to obtain an enhanced structural map, and the enhanced structural map is subjected to region growing and contour extraction using a region growing algorithm to obtain a region identification map. S32. Extract the pixel label value corresponding to each pixel from the region identification map, and sieve the pixels based on the pixel label value to obtain an initial pixel set; S33. Use an edge tracing algorithm to sort and connect all pixels in the initial pixel set to obtain the boundary line outline; S34. The global structural feature map is processed using the edge gradient analysis method. By quantifying the average edge intensity of the boundary line contour, the clarity of the corticomedullary boundary is obtained. S35. Obtain the importance weight value corresponding to each pixel point according to the pixel label value, and generate a regional importance weight map based on the importance weight value.

[0032] As described in steps S31-S35 above, this invention obtains an enhanced structure map by performing Gaussian smoothing on the global structure feature map to suppress noise and enhance region consistency. Subsequently, based on the enhanced structure map, the system uses image structure characteristics to initialize the segmentation process: it automatically sets initial seed points for the cortex and medulla in the outer region (expected cortical location) and the near-center region (expected medullary location) of the enhanced structure map, respectively, as the starting point for region growth. Then, using a region growth algorithm, it iteratively checks the 8 neighboring pixels centered on the initial seed point. If the intensity difference between the neighboring pixels and the current seed region is small... If a pixel is added to the current region at a preset threshold (the threshold is set based on image contrast and prior knowledge), the pixel is added until no new pixels can be added. After the region growth is completed, morphological operations (such as closing operations) are used to fill the holes inside the region, and edge detection operators (such as Canny) or active contour models are used to extract the precise contour of each region. All pixels belonging to the renal cortex are marked with a specific value (such as 1), all pixels belonging to the renal medulla are marked with another value (such as 2), and all pixels in the background region (pixels not covered by growth and with an intensity lower than a specific threshold) are marked with 0, thereby completing the segmentation and labeling, and constructing a region labeling map based on this.

[0033] Next, based on the generated region label map, each pixel in the region label map is traversed, and the pixel label values ​​of all pixels in its eight neighboring regions are checked. The judgment logic is: if the current pixel is labeled as renal cortex (value 1), and at least one of its eight neighboring pixels is labeled as renal medulla (value 2), then the current pixel is determined to be a boundary line pixel. After traversing the entire image using this rule, all the identified boundary line pixels together constitute the initial pixel set of the boundary line. Then, an edge tracking algorithm (such as the Moore-Neighborhood tracking algorithm) is used to sort and connect the pixels in the initial pixel set to form one or more continuous, closed boundary line contours. Here, the boundary line contour refers to the precise geometric expression of the boundary line between the two regions of renal cortex and renal medulla in the image. This method of region growing and contour extraction explicitly models the structural features of the image, combining feature extraction with structural constraints, and avoiding the limitations of general deep learning models. To address the issue of neglecting structural features due to a lack of prior knowledge, this paper proposes a method for analyzing the corticomedullary junction. For each pixel on the boundary contour, the Sobel operator is applied to obtain the gradient magnitude of that pixel based on the global structural feature map. Then, the arithmetic mean of the gradient magnitudes of all pixels on the boundary contour is calculated to obtain the average gradient magnitude. Finally, by querying a large database of normal samples, the typical range of gradient magnitudes at the healthy kidney boundary is determined, yielding the maximum and minimum gradient magnitudes. Based on the average gradient magnitude, the maximum and minimum gradient magnitudes, a linear scaling formula is used to normalize the average gradient magnitude, resulting in the corticomedullary junction sharpness. The corticomedullary junction sharpness is a standardized index used to objectively quantify the sharpness of the edge at the corticomedullary junction. A value closer to 1 indicates a sharper boundary, while a value closer to 0 indicates a more blurred boundary. This method of transforming edge sharpness perception into an objective quantitative index is beneficial for providing repeatable and quantifiable objective evidence for image analysis.

[0034] Then, each pixel in the region identification map is traversed, and a preset importance weight value is assigned to it based on its pixel label value: for all pixels labeled "renal cortex", a first weight value (e.g., 0.8) is assigned, giving it the highest weight to ensure that subsequent analysis focuses on these high-risk areas; for all pixels labeled "renal medulla", a second weight value (e.g., 0.5) is assigned, giving it a moderate weight to capture its potential contribution without excessive interference; for all pixels labeled "background", a third weight value (0) is assigned to completely exclude any interference from these non-organic areas to subsequent calculations, thus obtaining a region importance weight map with the same size as the original image. In this way, image structural information is directly embedded into the feature space, guiding subsequent fusion and recognition algorithms to concentrate computational resources on key image regions, effectively enhancing the feature focusing ability and recognition specificity of weak features.

[0035] In one embodiment, step S31, which involves pixel-level fusion of the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map, includes: S311. Extract the pixel coordinate range from the enhanced structure map, and obtain the cortical seed point coordinates and medullary seed point coordinates according to the pixel coordinate range; S312. Extract the cortical pixel intensity value and medullary pixel intensity value corresponding to the coordinates of the cortical seed point and the coordinates of the medullary seed point from the enhanced structure map, respectively. S313. Based on the cortical pixel intensity value and the medullary pixel intensity value, the region growing algorithm is used to iteratively expand the coordinates of the cortical seed point and the medullary seed point to obtain the cortical region coordinate set and the medullary region coordinate set. S314. Obtain an initial binary labeled image based on the coordinate set of the cortical region and the coordinate set of the medullary region, and process the initial binary labeled image using morphological closing operation to obtain a binary segmentation image; S315. An edge detection algorithm is used to extract the contour of the binary segmented image to obtain a contour coordinate set, and a region identification map is constructed based on the contour coordinate set, the cortical region coordinate set, and the medullary region coordinate set.

[0036] As described in steps S311-S315 above, this invention can clarify the spatial dimensional boundaries of the enhanced structure map by extracting the pixel coordinate range from the enhanced structure map, providing a coordinate reference for the subsequent division of the selection area for cortical and medullary seed points, and avoiding the selection of seed points exceeding the effective range of the image. Based on the typical spatial distribution pattern of kidney structure in ultrasound images (in healthy kidney images, the renal cortex is mostly distributed in the outer layer of kidney tissue, and the renal medulla is mostly concentrated in the inner near-central region of kidney tissue), the selection areas for cortical seed point coordinates and medullary seed point coordinates can be divided according to the pixel coordinate range: the area in the image located at the two edges in the horizontal direction and in the middle in the vertical direction is defined as the outer region of the image (matching the structural distribution position of the renal cortex), and the area in the image located in the middle in both the horizontal and vertical directions is defined as the inner near-central region of the image (matching the structural distribution position of the renal medulla). Then, 3-5 pixel coordinates are randomly selected in the outer region of the image as the cortical seed point coordinates, and 3-5 pixel coordinates are randomly selected in the inner near-central region of the image as the medullary seed point coordinates.

[0037] Next, the pixel intensity values ​​corresponding to the cortical seed point coordinates and medullary seed point coordinates are extracted from the enhanced structure map to obtain the cortical pixel intensity values ​​and medullary pixel intensity values. The set of cortical seed points is taken as the cortical starting region, and the set of medullary seed points is taken as the medullary starting region. The region growing algorithm is used for iterative expansion: for the 8 neighboring pixels of each pixel in the starting region, the difference between the intensity value of the neighboring pixels and the average intensity value of all pixels in the current starting region is calculated. If the difference is less than a preset threshold, the coordinates and intensity value of the neighboring pixels are merged into the current starting region. This iterative expansion process is repeated until no new pixels can be merged into any region, resulting in the cortical region coordinate set (containing the pixel coordinates and intensity values ​​of all pixels merged into the cortical starting region) and the medullary region coordinate set (containing the pixel coordinates and intensity values ​​of all pixels merged into the medullary starting region).

[0038] Then, an initial binary labeled image is generated based on the coordinate sets of the cortical and medullary regions, and a morphological closing operation is performed on the initial binary labeled image: using 3×3 rectangular structuring elements, first, the tiny gaps between pixels are filled by dilation, and then the original shape of the region edges is restored by erosion, thereby filling the voids inside the cortical and medullary regions to obtain a void-free binary segmentation image; then, contour extraction is performed on the binary segmentation image: using an active contour model, the initial contour is set as the outer rectangle of each region (determined based on the maximum and minimum coordinates of the region pixel set), and the elasticity coefficient is set to 0.3-0.5. Through iterative optimization, a set containing the contour coordinates of the cortical and medullary regions is obtained, i.e., the contour coordinate set; finally, the label values ​​corresponding to all pixel coordinates in the cortical region coordinate set and the cortical region contour coordinate set are set to 1, the label values ​​corresponding to all pixel coordinates in the medullary region coordinate set and the medullary region contour coordinate set are set to 2, and the label values ​​corresponding to the remaining pixels are set to 0, thus obtaining the region label map. This image segmentation method, which relies on the inherent intensity difference and spatial distribution characteristics of the renal cortex and medulla in ultrasound or CT images, is beneficial to ensure the accurate identification of image structures.

[0039] In one embodiment, step S4, which involves pixel-level fusion of the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map, includes: S41. Extract the anomaly saliency score of each pixel from the texture anomaly map, and obtain the local anomaly fluctuation value of each pixel based on the anomaly saliency score; S42. Extract the importance weight value of each pixel from the regional importance weight map, and perform adaptive weighted correction on the local abnormal fluctuation value based on the importance weight value to obtain the local salience enhancement value of each pixel; S43. Obtain the core threshold parameter and slope control parameter, and obtain the intermediate fusion feature value of each pixel based on the core threshold parameter, slope control parameter and local saliency enhancement value; S44. Obtain the fusion weight coefficient, and perform weighted fusion processing on the fusion weight coefficient, intermediate fusion feature value and the anomaly significance score to obtain the spatial weighted anomaly intensity value. S45. Construct a spatially weighted anomaly map based on the spatially weighted anomaly intensity values.

[0040] As described in steps S41-S45 above, this invention extracts the anomaly saliency score of each pixel from the texture anomaly map. For each pixel, the standard deviation of the set consisting of its own anomaly saliency score and the anomaly saliency scores of its eight neighboring pixels (including top, bottom, left, right, and four diagonal directions) is calculated. This yields the local anomaly fluctuation value for each pixel. The local anomaly fluctuation value refers to the feature value that quantifies the dispersion and drastic change of anomaly intensity within a small local region where the pixel is located. Since subtle anomalies in images often manifest as non-uniform patterns in a diffuse background, relying solely on the pixel's anomaly saliency score is insufficient to distinguish between real anomalies and uniform noise. The method of calculating the local standard deviation... This method can convert the anomalous significance score, which represents the absolute anomalous intensity of a single pixel, into a local anomalous fluctuation value, which represents the degree of variation in anomalous intensity within the local region where the pixel is located. By quantifying this degree of local variation, it can effectively capture the edge region of lesions (where the anomalous scores of adjacent pixels often differ significantly, resulting in a high standard deviation) and scattered microlesions (where these isolated anomalous points form a strong contrast with the surrounding normal pixels, also producing a high standard deviation). At the same time, this value can suppress interference signals in uniform regions (whether it is normal tissue or uniform noise, the pixel values ​​within the region change gradually, and the calculated standard deviation is low), thereby enhancing the algorithm's ability to identify image anomalous features.

[0041] Next, the importance weight value of each pixel is extracted from the regional importance weight map. Then, the enhancement coefficient is obtained. The enhancement coefficient is determined through offline optimization, mainly using a historical kidney ultrasound dataset with expert annotations. The enhancement coefficient is used as a hyperparameter, and a grid search is performed within a preset range (e.g., 0.1 to 5.0). For each candidate enhancement coefficient value, the performance of the generated local saliency enhancement map is evaluated, and the enhancement coefficient that achieves the best performance is finally selected as the final parameter. Then, the local abnormal fluctuation value is adaptively weighted and corrected based on the importance weight value: first, the product of the enhancement coefficient and the importance weight value is calculated, then the sum of this product and "1" is calculated to obtain the adaptive weighting factor, and then the product of the adaptive weighting factor and the local abnormal fluctuation value is calculated to obtain the local saliency enhancement value. The local saliency enhancement value refers to a local abnormality intensity index after being weighted and modulated by spatial prior knowledge. By adaptively amplifying the local abnormal fluctuation signal according to the importance of the image region where the pixel is located, it is beneficial to focus computational resources and attention on key image regions, significantly improve the perception ability of weak abnormal signals in these regions, and suppress interference from secondary regions.

[0042] Then, the core threshold parameter and slope control parameter are obtained. The core threshold parameter is pre-set based on the statistical distribution of local saliency enhancement values ​​in historical data (such as the median or a specific quantile of all pixel values), representing the theoretical critical point for distinguishing background from potential anomalies. The slope control parameter determines the steepness of the function's transition near the core threshold parameter, and its value is determined through offline optimization. The aim is to make the output change sharply near the core threshold parameter, and then use the formula " "Calculate the intermediate fusion feature value, where F represents the intermediate fusion feature value, β represents the slope control parameter, S represents the local significance enhancement value, and γ represents the core threshold parameter. The intermediate fusion feature value refers to the anomaly confidence score after nonlinear compression and probabilistic transformation. The core threshold parameter in the formula, as a learnable threshold, represents the critical point that distinguishes between 'background noise' and 'potential anomalies.' The term '(S-γ)' aims to align the center of the data distribution to zero. When the local significance enhancement value is greater than the core threshold parameter, the input is positive, and the Sigmoid output is greater than 0.5, indicating that the point is more likely to be an anomaly. When the local significance enhancement value is less than the core threshold parameter, the input is negative, and the output is less than 0.5, indicating that the point is more likely to be background noise. The slope control parameter controls the slope of the Sigmoid function at zero (i.e., near the core threshold parameter). A large slope control parameter will make the function curve very steep, which means that even if the local significance enhancement value is greater than the core threshold parameter, the local significance enhancement value will be less than the core threshold parameter ... core threshold parameter is a learnable threshold parameter that represents the critical point that distinguishes between 'background noise' and 'potential anomalies.' The core threshold parameter is a learnable threshold parameter that represents the critical point that distinguishes between 'background noise' and 'potential anomalies.' The core threshold parameter is a learnable threshold parameter that represents the critical point that distinguishes between 'background noise' and 'potential anomalies.' The core threshold parameter is a learnable threshold parameter that represents the critical point that distinguishes between 'background noise' and 'potential anomalies.' The core threshold Even a slight change in the saliency enhancement value relative to the core threshold parameter can cause a dramatic jump between 0 and 1 in the output intermediate fusion feature value. This effect greatly enhances the contrast of the image near the threshold, making the edges of abnormal regions clearer and the separation of the background and the target more thorough. The Sigmoid function ultimately non-linearly compresses the scaled and translated input values ​​into the interval (0, 1). This makes the output intermediate fusion feature value no longer a simple intensity value, but can be interpreted as the probability or confidence that the pixel belongs to an abnormal region. This transformation suppresses low saliency values ​​(e.g., suppressing 0.1 to 0.05) on the one hand, and saturates high saliency values ​​(e.g., enhancing 3.0 to 0.95) on the other hand. Finally, based on the intermediate fusion feature values ​​corresponding to all pixels in the image, an intermediate fusion feature map with the same size as the original image, less noise, more prominent targets, and all values ​​within a uniform range is constructed.

[0043] Finally, since the intermediate fusion feature map and the texture anomaly map carry different types of complementary information: the intermediate fusion feature map greatly highlights the anomaly regions with significant local contrast at important image locations and effectively suppresses background noise, but its drawback is that it may lose some original true anomaly intensity details or introduce edge artifacts due to over-enhancement and thresholding; conversely, the texture anomaly map retains the original, unmodified pixel-level anomaly intensity information, has the highest fidelity, and can reflect subtle gradient changes in anomalies, but its drawback is that it contains more noise and lacks spatial selectivity. Therefore, it is necessary to obtain the fusion weight coefficient, which is used to precisely control the contribution ratio of the two information sources. The fusion weight coefficient is determined through offline optimization, that is, the performance of different fusion weight coefficient values ​​is tested on historical labeled datasets, and the value that makes the final comprehensive risk index most correlated with the expected output is selected; if the fusion weight coefficient is close to 1, it indicates that the fusion result is highly dependent on the enhanced saliency features, which is suitable for pursuing high specificity and clarity. In scenarios where the core lesions are clearly highlighted, if the fusion weight coefficient is close to 0, it indicates greater trust in the original abnormal data, suitable for scenarios that pursue high sensitivity and avoid missing weak abnormal signals. Usually, the fusion weight coefficient is an intermediate value, indicating that both information are indispensable. The weight coefficient corresponding to the abnormality significance score can be obtained based on the difference between "1" and the fusion weight coefficient. Based on the intermediate fusion feature value, fusion weight coefficient, abnormality significance score and its corresponding weight coefficient, the spatial weighted abnormality intensity value corresponding to each pixel is calculated through a weighted summation algorithm. Based on the spatial weighted abnormality intensity values ​​corresponding to all pixels, a spatial weighted abnormality map with the same size as the original image is constructed. It has both the detail fidelity of the original texture abnormality map and incorporates the significance enhancement effect based on spatial prior. It is conducive to forming a highly reliable feature that is highly sensitive to diffuse weak abnormalities and can accurately locate key image regions, thereby solving the problem of insufficient recognition performance caused by the mismatch between feature expression and image characteristics in traditional methods.

[0044] In one embodiment, step S5, which involves obtaining a comprehensive evaluation feature vector based on the spatially weighted anomaly map and the anomaly region coordinate set, and obtaining a high-confidence anomaly region set based on the comprehensive evaluation feature vector, includes: S51. Extract the set of abnormal intensity values ​​corresponding to each abnormal region from the spatial weighted anomaly map coordinate set, and obtain the region feature vector based on the set of abnormal intensity values. S52. Extract the pixel marker values ​​of all pixels in each abnormal region from the binarized mask image, and obtain the morphological feature vector of each abnormal region based on the pixel marker values; S53. Extract the center coordinates of each abnormal region from the set of abnormal region coordinates; S54. The importance weight map of the region is processed by the skeletonization algorithm to obtain the renal cortex center line, and the cortical proximity of each abnormal region is obtained according to the center coordinates and the renal cortex center line. S55. The region feature vector, morphological feature vector and cortical proximity are concatenated to obtain a comprehensive evaluation feature vector; S56. Obtain the confidence score of each abnormal region based on the comprehensive evaluation feature vector, and construct a set of high-confidence abnormal regions based on the confidence scores.

[0045] As described in steps S51-S56 above, this invention accurately locates the corresponding region in the spatially weighted anomaly map for each anomaly region coordinate in the anomaly region coordinate set. Then, it traverses all pixels within that region, extracting the spatially weighted anomaly intensity value for each pixel to form an original set of anomaly intensity values. Based on this set of anomaly intensity values, it obtains the region feature vector corresponding to each anomaly region. The region feature vector is a set of values ​​used to quantify the distribution characteristics of anomaly intensity within an anomaly region, including the mean anomaly intensity, the maximum anomaly intensity, and the standard deviation of the anomaly intensity. The specific acquisition method is as follows: 1. Calculate the arithmetic mean of the spatially weighted anomaly intensity values ​​corresponding to all pixels within the anomaly region, i.e., obtain the mean anomaly intensity. This value reflects the average anomaly degree of the region; a higher mean anomaly intensity indicates a greater overall difference between the region and normal tissue. 2. Find the maximum value among the spatially weighted anomaly intensity values ​​corresponding to all pixels within the anomaly region, i.e., obtain the maximum anomaly intensity. This strategy is used to capture the most significant anomalies within anomaly regions, helping to identify suspicious areas that, while having a low average anomaly level, contain local extrema, thus avoiding the missed detection of weak but concentrated anomalous signals. Third, the standard deviation of the spatially weighted anomaly intensity values ​​corresponding to all pixels within the anomaly region is calculated, yielding the anomaly intensity standard deviation. This value measures the dispersion or fluctuation of the anomaly intensity values ​​within the region. A larger anomaly intensity standard deviation indicates an uneven distribution of anomalies within the region, potentially containing complex textures or blurred boundaries, while a smaller anomaly intensity standard deviation indicates a more homogeneous distribution. This multi-faceted statistical feature extraction strategy can characterize the anomaly intensity distribution characteristics of each anomaly region, providing richer and more discriminative information input for subsequent machine learning models. This enables them to more accurately distinguish target anomaly regions with specific statistical distribution patterns from false signals, thereby solving the problem of misjudgment caused by simple threshold segmentation or single features in existing technologies due to image noise or normal tissue variations.

[0046] Next, for each anomalous region coordinate in the anomalous region coordinate set, its corresponding region in the binary mask image is precisely located. Then, the pixel marker values ​​corresponding to all pixels within this anomalous region are extracted from the binary mask image. All pixels with a marker value of 1 collectively define the precise shape and extent of the anomalous region. By counting the total number of pixels with a marker value of 1, the area of ​​the anomalous region can be obtained. Subsequently, image processing algorithms (such as boundary-tracking algorithms) are used to extract the outer contour of the anomalous region. By accumulating the sum of the Euclidean distances between all adjacent pixel pairs on the contour, the perimeter of the anomalous region is calculated. Then, the ratio of the perimeter to the area is calculated to obtain the shape complexity of the anomalous region. Finally, the formula " "Calculate the circularity of the abnormal region, where C represents the circularity of the abnormal region, A represents the area of ​​the abnormal region, and P..." 2 The circularity measure of an abnormal region quantifies how close its shape is to a perfect circle. The core concept of this operation lies in leveraging the inherent morphological differences between the target abnormal region and image noise for identification: target abnormal regions tend to form relatively continuous, compact, and smooth-bounded areas (approximately circular or elliptical) in images, exhibiting moderate area, low perimeter-to-area ratio, and high circularity. Conversely, false signals generated by image noise, artifacts, or variations in normal regions typically exhibit small areas, fragmented and rugged shapes, and highly irregular boundaries, with significantly higher perimeter-to-area ratios and lower circularity. Therefore, by concatenating the area, shape complexity, and circularity of the abnormal region into a morphological feature vector corresponding to each abnormal region, the morphological feature vector provides a powerful and independent basis for subsequent classifiers, independent of intensity and spatial features. This significantly enhances the algorithm's ability to suppress false signals (improving specificity) and effectively solves the fundamental problem of low recognition accuracy caused by the lack of effective morphological filtering in existing technologies.

[0047] Then, based on the typical spatial distribution characteristics of target abnormal regions in kidney images—that an abnormal region with similar texture and morphological features is more likely to be a target abnormal region if it is adjacent to or located near the renal cortex centerline than a similar signal located in the renal medulla or peripheral region—this invention extracts the center coordinates of each abnormal region from the abnormal region coordinate set. Next, based on a regional importance weight map, for the target region formed by pixels whose importance weight value is assigned the first weight value (representing the renal cortex), a skeletonization algorithm (such as the Zhang-Suen iterative thinning algorithm) is applied to process it, continuously eroding the region edges until a… A central axis, only one pixel wide, is established to preserve the original topology and geometry of the renal cortex. This path, composed of a series of continuous pixel coordinates, is the renal cortical central line. Then, the Euclidean distance from the center coordinates of the abnormal region to each pixel on the renal cortical central line is calculated. By iterating through all pixels on the central line and calculating their Euclidean distances to the center coordinates of the abnormal region, the minimum value is taken as the cortical proximity. This value quantifies the spatial proximity of the abnormal region to key renal structures, transforming the prior knowledge of the "typical spatial distribution characteristics of the target abnormal region" into a computable indicator. This provides an important basis for subsequent classification and effectively improves the model's specificity.

[0048] Finally, the regional feature vector, morphological feature vector, and cortical proximity corresponding to each abnormal region are concatenated into a comprehensive evaluation feature vector. This comprehensive evaluation feature vector is then input into a pre-trained regional evaluation model. This model is a lightweight machine learning model trained on historical data to evaluate whether a region is abnormal, outputting a scalar value between 0 and 1 to obtain the confidence score for each abnormal region. The confidence score is a reliability index used to quantify the probability that the abnormal region is the target abnormal region. Finally, the coordinate information of each abnormal region, its corresponding comprehensive evaluation feature vector, and the confidence score are combined into a data item. All abnormal region data items are sorted from high to low confidence scores to form a high-confidence abnormal region set. This method, which combines multi-dimensional feature fusion with machine learning discrimination, can achieve accurate capture and identification of weak and diffuse features in kidney images. It helps to solve the problem of low sensitivity and poor specificity in target abnormality identification caused by the mismatch between feature extraction strategies and target abnormal features in existing technologies.

[0049] In one embodiment, step S7, which involves analyzing the set of significantly anomalous regions to obtain a comprehensive risk index, includes: S71. Extract the spatial weighted anomaly intensity value corresponding to all pixels in each significant anomaly region from the set of significant anomaly regions, and obtain the global anomaly cumulative load based on the spatial weighted anomaly intensity value; S72. Extract the area of ​​each significant abnormal region and the cortical proximity from the set of significant abnormal regions, and perform statistical analysis on the area of ​​all the abnormal regions to obtain the total area of ​​the abnormal regions. S73. Obtain the total area of ​​renal parenchyma, and perform fusion processing on the total area of ​​renal parenchyma, the total area of ​​abnormal regions and the global abnormal cumulative load to obtain a composite abnormality index. S74. Obtain the attenuation coefficient, and transform the cortical proximity using an exponential decay function based on the attenuation coefficient to obtain the spatial distribution risk coefficient. S75. Obtain the normalized coefficient, and integrate the normalized coefficient, composite anomaly index, spatial distribution risk coefficient and global anomaly cumulative load to obtain a comprehensive risk index.

[0050] As described in steps S71-S75 above, the present invention obtains the sum of the spatially weighted abnormal intensity values ​​corresponding to all pixels in each of the significant abnormal regions in the significant abnormal region set, thereby obtaining the global abnormal cumulative load. The global abnormal cumulative load refers to the sum of the abnormal signals carried by all significant abnormal regions in the kidney image. This index quantifies the total intensity of abnormal signals in the image as a whole. Since there may be global abnormal signals in the image that do not have obvious local protrusion features but are composed of a large number of scattered weak abnormal regions, by accumulating these distributed weak abnormal values, diffuse abnormal features that are easily ignored by traditional methods can be captured. This solves the problem that the existing technology only focuses on local salient features and cannot effectively evaluate global and distributed abnormalities.

[0051] Next, the area of ​​each significantly abnormal region is extracted from the set of significantly abnormal regions, and the total area of ​​abnormal regions is calculated. Then, based on the region identification map, the total number of pixels in the renal cortex and renal medulla is precisely counted to obtain the total area of ​​the renal parenchyma, which is the sum of the functional renal cortex and renal medulla, excluding the collecting system, cysts, or background outside the kidney. Then, the ratio of the total area of ​​abnormal regions to the total area of ​​the renal parenchyma is calculated to obtain the percentage of abnormal regions, reflecting the spatial extent of the abnormal region distribution. Finally, the ratio of the global cumulative abnormal load to the total area of ​​the renal parenchyma is calculated. This ratio can be understood as the "average abnormal load per unit area of ​​renal parenchyma," which normalizes the total abnormal intensity and eliminates bias caused by differences in kidney size. Then, add "1" to this ratio to form an intensity weighting factor. The purpose of adding "1" is to ensure that even if the global cumulative load of anomalies is very small, the weighting factor is still a positive number not less than 1, thus ensuring that the composite anomaly index will not be less than the simple area proportion due to the introduction of the intensity term. Finally, calculate the product of the anomaly area proportion and the intensity weighting factor to obtain the composite anomaly index. The composite anomaly index is a comprehensive index that simultaneously reflects the coverage of the anomaly area and the intensity of the anomaly signal. Through this data processing method, a composite feature with greater discriminative power than a single area or a single intensity index is created, which can amplify the contribution of the synergistic effect of the two, making the index more sensitive to typical anomaly patterns in the image, and effectively solving the problem that existing methods cannot adaptively fuse multi-dimensional weak information.

[0052] The spatial clustering of abnormal regions in an image is a crucial feature for distinguishing between target anomalies and false signals. The spatial distribution of target anomaly regions typically exhibits non-random clustering, while false positive regions caused by image noise or variations in normal regions tend to be more randomly dispersed. Therefore, this invention extracts the cortical proximity of each significant anomaly region from the set of significant anomaly regions and calculates the arithmetic mean based on all cortical proximity values ​​to obtain the average cortical proximity. The average cortical proximity quantifies the degree of clustering of the anomaly region distribution (the smaller the average cortical proximity, the more clustered the distribution). Next, an attenuation coefficient is obtained. The attenuation coefficient is a preset parameter, and its optimal value is typically determined through grid search or optimization algorithms on a historical image dataset containing labeled target anomalies and normal regions. The aim is to ensure that the calculated spatial distribution risk coefficient achieves the desired outcome. The optimal differentiation effect between the standard group and the control group is used to control the sensitivity of spatial clustering to the risk coefficient. Then, the product of the average nearest neighbor distance and the attenuation coefficient is calculated, and the negative of the product is taken as the exponent. Then, the natural exponential function is calculated to obtain the spatial distribution risk coefficient. By calculating the spatial distribution risk coefficient, this invention essentially constructs a highly efficient "clustering detector". It can keenly capture and quantify the key differences of the above distribution patterns, thereby distinguishing the diffuse abnormal spatial patterns with target characteristics from the randomly dispersed false positive signals. This significantly enhances the model's specificity for identifying target abnormal regions and effectively improves identification accuracy. More importantly, this method transforms the prior knowledge of "typical spatial distribution characteristics of target abnormal regions" into a computable index, solving the defect that general processing models cannot explicitly utilize such structured prior knowledge.

[0053] Subsequently, the normalization coefficient is obtained. The normalization coefficient is a preset constant determined by calculating the maximum value of the unnormalized composite risk index across all historical samples or the reciprocal of a specific quantile (e.g., the 99th percentile). Its function is to linearly scale the calculated composite risk index to a standardized range of 0 to 1, thus facilitating the setting of a uniform clinical decision threshold. Then, the composite risk index is calculated using the formula "I=ɑ*B*R*(1+log(1+G))", where I represents the composite risk index, B represents the composite anomaly index, R represents the spatial distribution risk coefficient, G represents the global anomaly cumulative load, and α represents the normalization coefficient. Since the global anomaly cumulative load is a cumulative value, its value can be very large and its distribution wide (long-tailed distribution). A very small number of samples with extremely high global anomaly cumulative loads will completely dominate the multiplication result, overwhelming the results of other samples. The contributions of the composite anomaly index and the spatial distribution risk coefficient are considered. Therefore, by taking the natural logarithm of "(1+G)" and then adding 1 to the result, it is ensured that even if the global cumulative anomaly load is 0, the result of this term is still 1, thus guaranteeing the mathematical stability of the formula. Next, the product of the composite anomaly index, the spatial distribution risk coefficient, and the logarithmic term is calculated. Through the synergistic effect of multiplication, it can be ensured that a high-risk score will only be generated when the evidence of the three dimensions of anomaly coverage, spatial distribution pattern, and anomaly intensity are all significantly present. This simulates the comprehensive decision-making logic of clinicians, greatly enhances the specificity of the model, and reduces false positives. Finally, the product of this result and the normalization coefficient is calculated to obtain the final comprehensive risk index, which can generate a stable, standardized risk assessment value with clear identification significance, which is beneficial to improving the accuracy and consistency of target abnormal region identification in kidney images.

[0054] This application also provides an image feature recognition system for early-stage kidney disease, comprising: The feature extraction module is used to acquire the original ultrasound image of the kidney and extract features from the original ultrasound image to obtain a local texture feature map and a global structural feature map. The texture analysis module is used to analyze the local texture feature map to obtain a texture anomaly map and a set of coordinates of anomaly regions; The structural analysis module is used to analyze the global structural feature map to obtain the clarity of the corticomedullary boundary and the weighted map of regional importance. The feature fusion module is used to perform pixel-level fusion of the texture anomaly map and the region importance weight map to obtain a spatially weighted anomaly map; The regional assessment module is used to obtain a comprehensive assessment feature vector based on the spatially weighted anomaly map and the anomaly region coordinate set, and to obtain a high-confidence anomaly region set based on the comprehensive assessment feature vector; The region filtering module is used to filter each high-confidence anomalous region in the high-confidence anomalous region set based on the clarity of the corticomedullary boundary, thereby obtaining a set of significant anomalous regions. The risk assessment module is used to analyze the set of significant abnormal regions, obtain a comprehensive risk index, and generate feature recognition results of the kidney image based on the comprehensive risk index.

[0055] It also includes several other modules, all of which are used to implement the steps of the image feature recognition method applied to the early stage of kidney disease.

[0056] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described image feature recognition method applied to the early stage of kidney disease.

[0057] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described image feature recognition method for early-stage kidney disease.

[0058] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0060] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An image feature recognition method applied to early stage of kidney disease, characterized in that, The method comprises the following steps: obtain an original ultrasound image of a kidney, and perform feature extraction on the original ultrasound image to obtain a local texture feature map and a global structure feature map; analyze the local texture feature map to obtain a texture anomaly atlas and an abnormal region coordinate set; analyze the global structure feature map to obtain a cortex-medulla boundary definition and a region importance weight map; fuse the texture anomaly atlas and the region importance weight map at the pixel level to obtain a spatially weighted anomaly atlas; obtain a comprehensive evaluation feature vector according to the spatially weighted anomaly atlas and the abnormal region coordinate set, and obtain a high-confidence abnormal region set according to the comprehensive evaluation feature vector; screen each high-confidence abnormal region in the high-confidence abnormal region set through the cortex-medulla boundary definition to obtain a significant abnormal region set; analyze the significant abnormal region set to obtain a comprehensive risk index, and generate a feature recognition result of the kidney image according to the comprehensive risk index.

2. The image feature recognition method for early stage of kidney disease according to claim 1, wherein, The step of analyzing the local texture feature map to obtain a texture anomaly atlas and an abnormal region coordinate set comprises the following steps: obtain a plurality of neighborhood window corresponding scalar feature values from the local texture feature map, and sequentially splice the scalar feature values to obtain a corresponding multi-dimensional feature vector; obtain a normal texture template of the kidney, and extract a standard feature vector corresponding to each neighborhood window from the normal texture template of the kidney; calculate the Euclidean distance corresponding to each neighborhood window based on the standard feature vector and the multi-dimensional feature vector, and construct an initial difference map based on the Euclidean distance; process the initial difference map using an adaptive threshold segmentation algorithm to obtain a segmentation threshold corresponding to each neighborhood window, and compare the Euclidean distance of each pixel with the segmentation threshold corresponding thereto to generate a binary mask image; perform connected region analysis on the binary mask image to obtain an abnormal region coordinate set; obtain an abnormal saliency score of each pixel point according to the initial difference map and the abnormal region coordinate set, and construct a texture anomaly atlas based on the abnormal saliency score. 3.The image feature recognition method for early stage of kidney disease according to claim 1, wherein, The step of analyzing the global structure feature map to obtain a cortex-medulla boundary definition and a region importance weight map comprises the following steps: perform Gaussian smoothing processing on the global structure feature map to obtain an enhanced structure map, and perform region growing and contour extraction on the enhanced structure map using a region growing algorithm to obtain a region identification map; extract a pixel label value corresponding to each pixel point from the region identification map, and screen the pixel points based on the pixel label value to obtain an initial pixel set; sort and connect all pixel points in the initial pixel set using an edge tracking algorithm to obtain a boundary contour; process the global structure feature map using an edge gradient analysis method to obtain a cortex-medulla boundary definition by quantifying the average edge strength of the boundary contour; obtain an importance weight value corresponding to each pixel point according to the pixel label value, and generate a region importance weight map based on the importance weight value.

4. The image feature recognition method for early stage of kidney disease according to claim 3, characterized in that, The step of performing region growing and contour extraction on the enhanced structure map by using a region growing algorithm to obtain a region identification map comprises: extracting a pixel coordinate range from the enhanced structure map, and obtaining cortex seed point coordinates and medulla seed point coordinates according to the pixel coordinate range; extracting cortex pixel intensity values and medulla pixel intensity values corresponding to the cortex seed point coordinates and the medulla seed point coordinates from the enhanced structure map; based on the cortex pixel intensity values and the medulla pixel intensity values, performing iterative expansion on the cortex seed point coordinates and the medulla seed point coordinates by using a region growing algorithm to obtain a cortex region coordinate set and a medulla region coordinate set; obtaining an initial binary labeled image according to the cortex region coordinate set and the medulla region coordinate set, and processing the initial binary labeled image by using a morphological closing operation to obtain a binary segmentation image; performing contour extraction on the binary segmentation image by using an edge detection algorithm to obtain a contour coordinate set, and constructing a region identification map according to the contour coordinate set, the cortex region coordinate set and the medulla region coordinate set. 5.The image feature recognition method for early stage of kidney disease according to claim 1, wherein, The step of performing pixel-level fusion on the texture anomaly atlas and the region importance weight map to obtain a spatially weighted anomaly atlas comprises: extracting an anomaly saliency score of each pixel point from the texture anomaly atlas, and obtaining a local anomaly fluctuation value of each pixel point according to the anomaly saliency score; extracting an importance weight value of each pixel point from the region importance weight map, and performing adaptive weighted correction on the local anomaly fluctuation value based on the importance weight value to obtain a local saliency enhancement value of each pixel point; obtaining a core threshold parameter and a slope control parameter, and obtaining an intermediate fusion feature value of each pixel point according to the core threshold parameter, the slope control parameter and the local saliency enhancement value; obtaining a fusion weight coefficient, and performing weighted fusion processing on the fusion weight coefficient, the intermediate fusion feature value and the anomaly saliency score to obtain a spatially weighted anomaly intensity value; constructing a spatially weighted anomaly atlas based on the spatially weighted anomaly intensity value. 6.The image feature recognition method for early stage of kidney disease according to claim 1, wherein, The step of obtaining a comprehensive evaluation feature vector according to the spatially weighted anomaly atlas and the anomaly region coordinate set, and obtaining a high-confidence anomaly region set according to the comprehensive evaluation feature vector comprises: extracting an anomaly intensity value set corresponding to each anomaly region in the anomaly region coordinate set from the spatially weighted anomaly atlas, and obtaining a region feature vector according to the anomaly intensity value set; extracting a pixel label value of all pixel points in each anomaly region from a binary mask image, and obtaining a morphological feature vector of each anomaly region based on the pixel label value; extracting a center coordinate of each anomaly region from the anomaly region coordinate set; processing the region importance weight map by using a skeletonization algorithm to obtain a renal cortex center line, and obtaining a cortex proximity of each anomaly region according to the center coordinate and the renal cortex center line; splicing the region feature vector, the morphological feature vector and the cortex proximity to obtain a comprehensive evaluation feature vector; splicing the region feature vector, the morphological feature vector and the cortex proximity to obtain a comprehensive evaluation feature vector; According to the comprehensive evaluation feature vector, a confidence score of each abnormal region is obtained, and a high-confidence abnormal region set is constructed based on the confidence score. 7.The image feature recognition method for early stage of kidney disease according to claim 1, wherein, The step of analyzing the set of significant abnormal regions to obtain a comprehensive risk index comprises: Extracting spatially-weighted abnormal intensity values corresponding to all pixel points in each significant abnormal region from the set of significant abnormal regions, and obtaining a global abnormal cumulative load according to the spatially-weighted abnormal intensity values; Extracting an abnormal region area and a cortical proximity corresponding to each significant abnormal region from the set of significant abnormal regions, and statistically analyzing all the abnormal region areas to obtain a total abnormal region area; Obtaining a total kidney parenchyma area, and performing fusion processing on the total kidney parenchyma area, the total abnormal region area, and the global abnormal cumulative load to obtain a composite abnormal index; Obtaining an attenuation coefficient, and transforming the cortical proximity based on the attenuation coefficient using an exponential attenuation function to obtain a spatial distribution risk coefficient; Obtaining a normalization coefficient, and performing integrated fusion on the normalization coefficient, the composite abnormal index, the spatial distribution risk coefficient, and the global abnormal cumulative load to obtain a comprehensive risk index.

8. An image feature recognition system applied to early stage of kidney disease, characterized in that, The computer program product comprises a plurality of modules for implementing the steps of the method of any one of claims 1 to 7. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor implements the steps of the method of any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images