Hypertensive fundus lesion detection system based on image recognition

Abstract

The application discloses a hypertension fundus lesion detection system based on image recognition, and comprises the following steps: performing format conversion, noise suppression and picture enhancement on a fundus image to generate a standard image; extracting structural features based on the standard image and forming a feature set; calculating a response value for the feature set to generate a response vector; performing mapping processing on the structural features and the response vector to form coupled feature data; generating a saliency map according to the position of the coupled feature in the image space; screening continuous regions in the saliency map to form a candidate lesion region map; extracting a composite feature vector from the candidate region, and generating lesion classification, position indexing and confidence information based on the composite feature to realize lesion recognition and position calibration.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a hypertensive retinopathy detection system based on image recognition. Background Technology

[0002] With the development of digital ophthalmic diagnosis and intelligent screening technologies, the automatic detection of hypertension-related lesions using fundus images has become an important application direction for intelligent medical imaging. Currently, lesion recognition generally relies on classification methods based on convolutional networks or traditional texture features to extract features and determine regions from input fundus images, assisting doctors in completing lesion screening and risk assessment.

[0003] Existing techniques for detecting fundus lesions still have significant limitations when dealing with image quality fluctuations, significant local structural changes, and large differences in lesion morphology. On the one hand, traditional image enhancement and feature extraction methods struggle to reliably acquire effective features under conditions of noise interference, overexposure, color shifts, and complex microvascular structures. This often results in blurred structural boundaries, loss of local texture, or insufficient response in lesion areas, leading to instability in subsequent identification processes. On the other hand, detection models based on single features or fixed representations are poorly adaptable to the irregularities and distributional differences in lesion regions. They cannot simultaneously reflect image structural information and local response levels, easily leading to missed detections, increased false positives, or localization errors.

[0004] Therefore, how to provide an image recognition-based detection system for hypertensive retinopathy is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an image recognition-based hypertensive retinopathy detection system. Through image preprocessing, structural feature extraction, entropy response construction, cross-vector mapping coupling, saliency generation, candidate region screening, and composite feature construction, the system achieves precise presentation and classification indexing of lesion regions.

[0006] The image recognition-based hypertensive retinopathy detection system according to an embodiment of the present invention includes the following steps: The image preprocessing module receives fundus images and performs format conversion, noise suppression, and image enhancement on the fundus images to generate standard images. The structural feature extraction module performs structural analysis on standard images based on improved Krautchouk moments to generate a set of structural features. The entropy response construction module calculates the response value of each structural feature in the structural feature set based on the Rényi entropy function, and generates a set of entropy response vectors. The feature coupling module maps the set of structural features to the set of entropy response vectors to generate coupled feature data. The saliency mapping module maps coupled feature data to a standard image space and generates a saliency map based on a preset response threshold. The candidate region extraction module judges the spatial continuity and distribution stability of the coupled feature data in the saliency map and generates a candidate lesion region map. The composite feature construction module extracts coupled feature data from the candidate lesion area map and generates a composite feature vector; The diagnostic result generation module generates classification identification information, image location indexing information, and confidence information based on composite feature vectors.

[0007] Optionally, the generated set of structural features includes: The standard image is divided into sections based on a fixed number of rows and columns. For each non-overlapping rectangular region, the gray values ​​of adjacent pixels are read row by row, the increase or decrease relationship of the gray values ​​of adjacent pixels is compared, and recorded as "+", "-" or "0" in turn. The recorded results are arranged in row order to form a symbol sequence grid. Perform a sign substitution operation on the symbol sequence grid, replacing the sign "+" with a positive number, the sign "-" with a negative number, and the sign "0" with a zero value; Set a position index for each row of the symbolic numerical grid. ,in An integer incrementing from zero, representing the horizontal pixel position within the current structural block; Let the width of the current structure block be... , representing the total number of pixels contained in the row; let the projection order be . , indicating the current expansion level of the Krawtchouk basis functions; Will Assign the value to the constant 1, and The initial value is defined according to the formula. ,in This is a probability parameter, with a value between 0 and 1; For all orders In the location index The Krawtchouk basis function values ​​are calculated using a recursive formula, and each calculation result is written into the corresponding column of the horizontal basis function table in ascending order of order. Write all recursive results into the horizontal basis function table, and then... OK, List organization; Each column in each structural block is processed in the same way as the horizontal basis function table is generated, and the vertical pixel index is set to . The column width is Repeat the recursive calculation process and sort the corresponding results according to... OK, Columns are written into the vertical basis function table; For each pixel position Read the pixel value at the location from the symbolic numerical grid, and read the basis function values ​​of the corresponding order from the horizontal and vertical basis function tables respectively. Multiply the three values ​​as the moment product value of the current projection point and record them in the moment coefficient accumulation table of the corresponding order. After traversing all pixel positions, the product of moments of all orders in each structural block is accumulated and used as the local Krawtchouk moment coefficients corresponding to the structural block. The local moment coefficients of all structural blocks are written into the contiguous storage area in the order of the structural blocks to form a first-level moment value sequence. The first-level moment value sequence is further concatenated into an overall moment feature vector according to the scanning order, and divided into N moment value segments of fixed length. The difference comparison of corresponding terms is performed on each pair of adjacent moment value segments in turn. If the value of the previous segment is greater than that of the next segment, it is recorded as "1", less than it is recorded as "-1", and equal to it is recorded as "0", thus generating a difference sign sequence. The local Krautchouk moment coefficient sequence, the global moment feature vector, and the difference sign sequence are combined sequentially to form a set of structural features.

[0008] Optionally, the set of generated entropy response vectors includes: Read each structural feature vector sequentially from the set of structural features, and represent the current structural feature vector as a vector containing m real components. And establish an index number for the structural feature vector in memory; For each component in the current structural feature vector, calculate the adjacent difference sequentially, for indices from 1 to... Calculate the difference at each position Set a fixed difference threshold :when At position i, record the sign "+", when At position i, record the symbol "-", when Less than or equal to At position i, record the symbol "0"; Arrange all symbols in index order to obtain the incremental direction symbol sequence; Perform numerical mapping on the incremental direction symbol sequence, replacing the symbol "+" at each position with the value 1, the symbol "-" with the value -1, and the symbol "0" with the value 0, generating a direction value sequence, and storing it in a one-dimensional array in the original index order; Set interval length parameter For positive integers, starting from the first element of the direction value sequence, proceed in order with each... Each element is used to divide adjacent intervals. For each complete interval, ... Iterate through each element and count the number of times the value 1 appears. The number of times the value 0 appears is recorded. The number of times the value -1 appears is recorded. and will Write the results into a ternary array in a fixed order as the statistical results for that interval; Calculate the total value for the ternary statistical results for each interval. Perform a weight calculation for each component: Divide by Obtain the proportion ,Will Divide by Obtain the proportion ,Will Divide by Obtain the proportion Generate a weight vector And write them into the weight vector array in interval order; The cumulative weight vector array is processed by keeping the weight vector of the first interval unchanged. Starting from the second interval, the weight vector of each interval is added to the weight vector of the immediately preceding interval according to the components to calculate the cumulative weight vector. The cumulative weight vector is then written into the cumulative weight array in interval order. Set the order parameter of the Rényi entropy function, apply the Rényi entropy function to each cumulative weight vector in the cumulative weight array, perform exponentiation, summation and logarithmic operations, and calculate the corresponding entropy response value; Write the entropy response value of each interval into a one-dimensional array according to the interval index to obtain the entropy response vector corresponding to the current structural feature vector; Repeat the above entropy response vector generation operation for all structural feature vectors in the structural feature set. Store the entropy response vector corresponding to each structural feature vector into the entropy response vector storage area in the order of the structural feature vectors in the structural feature set to form an entropy response vector set.

[0009] Optionally, generating coupling feature data includes: Read the structural feature vectors sequentially from the set of structural features according to their index numbers, and represent each structural feature vector as a sequence of real number components. Read the entropy response vectors with the same index number from the set of entropy response vectors. Each entropy response vector contains a fixed number of entropy response values. Set the number of components in the current structural feature vector to . The number of components in the entropy response vector is set to ,according to and The numerical relationship constructs a size of Position index mapping matrix; Perform a linear transformation operation on each column of the position index mapping matrix, read the corresponding values ​​sequentially from the first row to the last row, divide each read value by the sum of all values ​​in that column to obtain the column normalization result, write the normalization result into the column vector to form the column normalized position sequence; For each column of normalized position sequence, perform an element-by-element multiplication operation with the corresponding component of the entropy response vector, using the entropy response component of the corresponding column as the multiplier to obtain the product sequence; Read all product sequences in column order, accumulate all column values ​​in each row, and write the accumulated result to the corresponding position of the coupled feature vector to obtain the complete coupled feature vector of the current structure feature vector; Repeat the above steps for generating complete coupled feature vectors for all structural feature vectors in the structural feature set; All generated coupled feature vectors are written sequentially into the coupled feature data storage area according to the original index order of the structural feature set, forming a coupled feature data set.

[0010] Optionally, generating a saliency plot includes: The coupled feature data set is organized into a two-dimensional coupled feature array according to the number of rows and columns of blocks used when the structural feature extraction module generates the structural feature set. Each element in the array corresponds to a coupled feature vector. For each coupled feature vector in the two-dimensional coupled feature array, starting from the first component of the vector, all components are read sequentially, and a summation operation is performed on all components. The summation value is divided by the number of components to obtain the response value corresponding to the vector. All response values ​​are written into the response value array according to the row and column indices of the array elements. Based on the block size of the standard image, starting from the top left corner of the standard image, and according to the row and column indices of the response value array, determine the starting row number, starting column number, row span, and column span of the corresponding pixel block in the standard image for each response value. For each response value, read the corresponding pixel block position, and write the response value into all pixel positions covered by the pixel block in a row-by-column manner, so that all pixel blocks are filled with the corresponding response value, and generate an intermediate saliency map. Perform maximum and minimum value calculations on all pixels in the intermediate saliency map to obtain normalized pixel values, and write all normalized pixel values ​​into the normalized saliency map according to their original pixel positions; Set a fixed response threshold A comparison operation is performed on each pixel in the normalized saliency map. When the normalized pixel value is greater than or equal to... When the normalized pixel value is less than 1, the output value at that pixel position is set to 1. When the output value is set to 0, the final saliency image is formed.

[0011] Optionally, generating candidate lesion region maps includes: Each pixel value in the saliency map is loaded into the saliency matrix according to its two-dimensional coordinates; pixels with a value of 1 in the matrix are set as foreground pixels, pixels with a value of 0 are set as background pixels, and an initial foreground pixel label matrix is ​​generated. Perform a fixed four-neighbor connected region scan on the foreground pixel labeling matrix, and traverse each foreground pixel in each row and each column in turn: if there are pixels that have been labeled as foreground pixels in its four adjacent positions above, below, left and right, then assign the current pixel and its adjacent pixels the same connected number; if there are no adjacent foreground pixels, then assign a new number to the current pixel. After completing the traversal, a region index matrix containing the numbers of all connected regions is generated. Perform pixel counting on each of the non-zero numbered regions in the region index matrix and count the number of foreground pixels contained in each numbered region; set a pixel count threshold T, record all pixel coordinates of all numbered regions with a pixel count less than T in the elimination list, and clear the corresponding number value in the region index matrix to zero; For each region number in the reserved area, read all pixel coordinates corresponding to that number from the region index matrix, obtain the minimum row number, maximum row number, minimum column number, and maximum column number respectively, and determine a rectangular border whose coverage ranges from the minimum row number to the maximum row number and from the minimum column number to the maximum column number. Using the region number as the key, the border coordinate information and the set of pixel coordinates of each reserved region are written into the lesion candidate region structure; All structures are uniformly written into the candidate region map matrix, and all pixels belonging to the candidate region are marked as 1 and other pixels as 0 in the image coordinate system in binary form to generate the candidate lesion region map.

[0012] Optionally, generating composite feature vectors includes: Load the candidate lesion region map into a pixel value matrix, where each pixel position with a value of 1 represents a pixel within the candidate lesion region. Read the two-dimensional row and column coordinates of the positions with pixel values ​​of 1 one by one and store them into the lesion pixel coordinate list. The system receives a set of coupled feature data generated by the feature coupling module. This matrix corresponds one-to-one with the standard image space, and each matrix position stores a coupled feature vector. For each coordinate position in the list of lesion pixel coordinates, the corresponding coupled feature vector is read from the coupled feature data set, written into the lesion feature vector sequence in the order of reading, and the source coordinates of each vector are recorded. The lesion feature vector sequence is classified into regions according to the spatial continuity of the source coordinates. Coordinates with continuous row and column numbers are clustered and divided into independent lesion region groups. All coupled feature vectors in each group form a feature subset. For all coupled feature vectors in each feature subset, the mean is calculated along each dimension to obtain the average response value in each dimension of the subset, which is denoted as the mean vector of the subset. For each coupled feature vector in the same feature subset, calculate the squared Euclidean distance between it and the mean vector, and average all the squared Euclidean distances to obtain the variance index of the subset. Perform range calculation on the response values ​​in each dimension of the feature subset, find the maximum and minimum values ​​of each dimension, calculate the difference and arrange them in dimensional order to form the response range vector of the subset. The mean vector, variance index, and response range vector are concatenated in a predetermined order to form a composite feature vector for the lesion region.

[0013] Optionally, the dataset for generating diagnostic results includes: Each composite feature vector in the composite feature vector set is loaded sequentially according to its index, and each composite feature vector is assigned a unique number to establish a mapping table between the number and the original lesion area. For each composite feature vector, numerical features are extracted in order of vector dimension to form an input vector, which is then input into the trained classification model. Forward computation is performed, and a probability distribution vector containing class probability values ​​is output. In the probability distribution vector, obtain the target category index with the largest value, and write the category identifier corresponding to the index into the classification identifier information field; at the same time, use the maximum value as the confidence value of the corresponding lesion area and write it into the confidence information field. Based on the number of the composite feature vector, the coordinate information of the lesion region associated with the number is found in the original candidate lesion region map, and the row and column number, width and height parameters of the upper left corner of the border are extracted to generate image location indexing information. The classification identification information, image location indexing information, and confidence information are written into the result structure in a unified format, and all result structures are arranged in numerical order to form a diagnostic result dataset.

[0014] The beneficial effects of this invention are: This invention introduces a multi-dimensional feature modeling method that combines structural feature extraction with Rényi entropy response construction, enabling simultaneous analysis of local and global information in fundus images. It can accurately capture the response features of lesion areas under conditions of noise, uneven illumination, or subtle lesions, significantly improving the stability and robustness of feature expression.

[0015] This invention integrates structural features and entropy response across space through feature coupling and saliency mapping mechanisms, and maps the results back to the standard image space to form a lesion saliency map. This achieves a visual correspondence between image structural information and response intensity, significantly improving the localization accuracy of lesion areas.

[0016] This invention uses a candidate region extraction and composite feature construction module to perform spatial connectivity analysis and region screening on the saliency map, extract representative composite feature vectors, and combine them with a classification model to generate classification labels, location indexes, and confidence information. This achieves integrated processing of lesion detection, classification, and visualization annotation, and has the beneficial effects of high recognition accuracy, strong adaptability, and high degree of automation. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the image recognition-based hypertensive fundus lesion detection system proposed in this invention; Figure 2 This is a schematic diagram illustrating the calculation of the improved Krautchouk moment coefficients proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 - Figure 2 The image recognition-based hypertensive retinopathy detection system includes the following steps: The image preprocessing module performs format conversion, noise suppression, and image enhancement on fundus images to generate standard images; The structural feature extraction module performs structural analysis on standard images based on improved Krautchouk moments to generate a set of structural features. The entropy response construction module calculates the response value of each structural feature in the structural feature set based on the Rényi entropy function, and generates a set of entropy response vectors. The feature coupling module maps the set of structural features to the set of entropy response vectors to generate coupled feature data. The saliency mapping module maps coupled feature data to a standard image space and generates a saliency map based on a preset response threshold. The candidate region extraction module judges the spatial continuity and distribution stability of the coupled feature data in the saliency map and generates a candidate lesion region map. The composite feature construction module extracts coupled feature data from the candidate lesion area map and generates a composite feature vector; The diagnostic result generation module receives composite feature vectors and generates classification identification information, image location indexing information, and confidence information based on the composite feature vectors.

[0020] In this embodiment, generating a standard image includes: Receive fundus image files, load the images as raw pixel matrices, and read their row and column dimensions and color channel format; Perform format conversion processing on the original pixel matrix. If it is a color image, extract the red, green and blue channel data and reassemble them into a unified three-channel matrix. If it is a grayscale image, retain the single-channel data unchanged. The pixel matrix after format unification is subjected to size standardization processing, and the image is adjusted to the preset standard size using a bilinear interpolation algorithm; Noise suppression processing is performed on standard-sized images by using a fixed-size sliding window to extract local region pixels, calculating the median and replacing the center pixel value of the window to form a denoised image; To enhance the denoised image, first perform a linear contrast stretching operation to map pixel values ​​to a preset brightness range, then perform local histogram equalization to improve local grayscale differences, and finally output a standard image.

[0021] In this embodiment, generating the structural feature set includes: The standard image is divided into multiple non-overlapping rectangular regions with a fixed number of rows and columns. The gray values ​​of adjacent pixels in each rectangular region are read row by row. The increase or decrease relationship of the gray values ​​of adjacent pixels is compared and recorded as "+", "-" or "0" in turn. The recorded results are arranged in row order to form a symbol sequence grid. Perform a sign substitution operation on the sign sequence grid, replacing the sign "+" with a positive number, the sign "-" with a negative number, and the sign "0" with a zero value, to form a sign numerical grid that can be used for numerical operations; Set a position index for each row of the symbolic numerical grid. ,in An integer incrementing from zero, representing the horizontal pixel position within the current structural block; Let the width of the current structure block be... , representing the total number of pixels contained in the row; let the projection order be . , used to indicate the current expansion level of the Krawtchouk basis functions; Will Assign the value to the constant 1, and The initial value is defined according to the formula. ,in This is a probability parameter, with a value between 0 and 1, used to adjust the shape of the function; For all orders In the location index The following recursive formula is used to calculate the Krawtchouk basis function values, and each calculation result is written into the corresponding column of the horizontal basis function table in ascending order of order: ; in, Indicates horizontal position Place, No. The Krawtchouk basis function values ​​of order 1. : These are the basis function values ​​for the first and second orders, respectively. These are probability parameters used in the basis functions to adjust the shape of the discrete distribution. The width of the current structure block; All recursive results are written into the horizontal basis function table, categorized by... OK, List organization; Each column in each structure block is processed in the same way, with the vertical pixel index set to . The column width is Repeat the above recursive calculation process and sort the corresponding results according to... OK, Columns are written into the vertical basis function table; For each pixel position Read the pixel value at the location from the symbolic numerical grid, and read the basis function values ​​of the corresponding order from the horizontal and vertical basis function tables respectively. Multiply the three values ​​as the moment product value of the current projection point and record them in the moment coefficient accumulation table of the corresponding order. After traversing all pixel positions, the product of moments of all orders in each structural block is accumulated and used as the local Krawtchouk moment coefficient corresponding to that structural block. The local moment coefficients of all structural blocks are written into the contiguous storage area in the order of the structural blocks to form a first-level moment value sequence. The first-level moment value sequence is further concatenated into an overall moment feature vector according to the scanning order, and divided into multiple moment value segments with a fixed length. The difference comparison of corresponding terms is performed on each pair of adjacent moment value segments in turn. If the value of the previous segment is greater than that of the next segment, it is recorded as "1", less than it is recorded as "-1", and equal to it is recorded as "0", generating a difference sign sequence. The local Krautchouk moment coefficient sequence, the global moment feature vector, and the difference sign sequence are combined sequentially to form a structural feature set, which is then passed as the output of the structural feature extraction module to the entropy response construction module.

[0022] The Krawtchouk basis function recursive formula originates from the discrete orthogonal polynomial theory in mathematics. Originally proposed by Krawtchouk in the 20th century, it describes a set of orthogonal functions over a finite domain and is widely used in discrete signal processing and image analysis. This application, based on this classical algorithm, incorporates the discrete pixel arrangement characteristics of image structural blocks and introduces pixel position indices in the image. , width of structural blocks order and control parameters The original recurrence relation is parametrically transformed to fit the specific structural distribution of rows and columns in the image. The formula takes the following form: ; All variables are dimensionless discrete parameters. Indicates the first Krawtchouk basis functions of order at position The function value at a given location represents the structural response of the image at that location, expressed in pure numbers. Both sides of the formula are Krautchouk basis function values, and its structure is a weighted recursive relationship between basis functions. The weight coefficients are unitless algebraic expressions, which conforms to the mathematical rationality of discrete transformation processing in image space. Therefore, the physical dimensions of the overall formula are consistent, satisfying the dimensional conservation requirements in the construction of orthogonal functions, and it has good engineering applicability and mathematical interpretability.

[0023] In this embodiment, the set of entropy response vectors to be generated includes: Read each structural feature vector sequentially from the set of structural features, and represent the current structural feature vector as a vector containing m real components. And create an index number for the feature vector of this structure in memory; For each component in the current structural feature vector, calculate the adjacent difference sequentially, for indices from 1 to... Calculate the difference at each position Set a fixed difference threshold ,when At position i, record the sign "+", when At position i, record the symbol "-", when Less than or equal to At position i, record the symbol "0" and arrange all symbols in index order to obtain the incremental direction symbol sequence; Perform numerical mapping on the incremental direction symbol sequence, replacing the symbol "+" at each position with the value 1, the symbol "-" with the value -1, and the symbol "0" with the value 0, generating a direction value sequence, and storing it in a one-dimensional array in the original index order; Set interval length parameter For positive integers, starting from the first element of the direction value sequence, proceed in order with each... Each element is used to divide adjacent intervals. For each complete interval, ... Iterate through each element and count the number of times the value 1 appears. The number of times the value 0 appears is recorded. The number of times the value -1 appears is recorded. and will Write the results into a ternary array in a fixed order as the statistical results for that interval; Calculate the total value for the ternary statistical results for each interval. Perform a weight calculation on each component, and... Divide by Obtain the proportion ,Will Divide by Obtain the proportion ,Will Divide by Obtain the proportion Generate a weight vector And write them into the weight vector array in interval order; The cumulative weight vector array is processed by adding the weight vector of the first interval to each interval, keeping the weight vector of the first interval unchanged. Starting from the second interval, the weight vector of each interval is added to the weight vector of the immediately preceding interval according to its components to calculate the cumulative weight vector. And write the cumulative weight vector into the cumulative weight array in interval order; Set the order parameter of the Rényi entropy function Apply the Rényi entropy function to each cumulative weight vector in the cumulative weight array, performing exponentiation, summation, and logarithmic operations to calculate the corresponding entropy response value: ; Write the entropy response value of each interval into a one-dimensional array according to the interval index to obtain the entropy response vector corresponding to the current structural feature vector; Repeat the above process for all structural feature vectors in the structural feature set, and store the entropy response vector corresponding to each structural feature vector into the entropy response vector storage area in the order of the structural feature vectors in the structural feature set to form an entropy response vector set.

[0024] In this embodiment, generating coupling feature data includes: Read the structural feature vectors sequentially from the set of structural features according to their index numbers, and represent each structural feature vector as a sequence of real number components. Read the entropy response vectors with the same index number from the set of entropy response vectors. Each entropy response vector contains a fixed number of entropy response values. Set the number of components in the current structural feature vector to . The number of components in the entropy response vector is set to ,according to and The numerical relationship constructs a size of The position index mapping matrix; each row of the matrix corresponds to a component in the structural feature vector, and each column corresponds to an entropy value in the entropy response vector. The difference between the row index and the column index in the matrix is ​​written into the matrix in absolute value order to represent the relative positional relationship between the structural feature component and the entropy response component. Perform a linear transformation operation on each column of the position index mapping matrix, read the corresponding values ​​sequentially from the first row to the last row, divide each read value by the sum of all values ​​in that column to obtain the column normalization result, write the normalization result into the column vector to form the column normalized position sequence; For each column of normalized position sequence, perform an element-wise multiplication operation with the corresponding component of the entropy response vector, using the entropy response component of the corresponding column as the multiplier to obtain a product sequence, which is used to represent the coupling weight of the structural feature component on the position trajectory. Read all product sequences in column order, accumulate all column values ​​in each row, and write the accumulated result to the corresponding position of the coupled feature vector to obtain the complete coupled feature vector of the current structural feature vector; this process establishes a one-to-one mapping relationship between each component in the structural feature vector and all components in the entropy response vector. Repeat the above processing steps for all structural feature vectors in the structural feature set, and write all the generated coupled feature vectors into the coupled feature data storage area in the original index order of the structural feature set to form a coupled feature data set. The coupled feature dataset is taken as the output of the feature coupling module and passed to the saliency mapping module as input data.

[0025] Construct a size of The location index mapping matrix includes: Defining the source of dimension: setting the number of components in the structural feature vector as... This indicates that the vector contains A number of consecutively numbered structural feature components; the number of components in the entropy response vector is set to... , indicating that it contains Entropy response index; Create an empty matrix: Initialize a matrix of size in memory. OK, A two-dimensional array of columns, where each element corresponds to a structural feature component. With entropy response components One-to-one combination ; Fill position value: For each element in the matrix, calculate its row index. With column index The absolute value of the difference between them, i.e., the execution Write this value into the corresponding matrix cell to represent the relative distance between the current structural feature component and the entropy response component; Complete construction: After all cells are filled, a result is obtained where each element is a non-negative integer. The mapping matrix is ​​used to subsequently construct normalized position weights and perform mapping fusion of structural features and entropy response; The core function of this matrix is ​​to quantify the relative positional dependence between structural feature components and entropy response components, thereby enabling subsequent coupled calculations to have spatial interpretability and structural consistency.

[0026] In this embodiment, generating a saliency map includes: The coupled feature data is organized into a two-dimensional coupled feature array according to the number of rows and columns used when setting up the structural feature set, where each element in the array corresponds to a coupled feature vector. For each coupled feature vector in the two-dimensional coupled feature array, starting from the first component of the vector, all components are read sequentially, and a summation operation is performed on all components. The summation value is divided by the number of components to obtain the response value corresponding to the vector. All response values ​​are written into the response value array according to the row and column indices of the array elements. Based on the block size of the standard image in the structural feature extraction module, starting from the top left corner of the standard image, according to the row and column indices of the response value array, the starting row number, starting column number, row span, and column span of the corresponding pixel block in the standard image are determined for each response value. For each response value, read its corresponding pixel block position, and write the response value into all pixel positions covered by the pixel block in a row-by-row and column-by-column manner, so that all pixel blocks are filled with the corresponding response value, thereby generating an intermediate saliency map. Perform maximum and minimum value calculations on all pixels in the intermediate saliency map, subtract the minimum value from the original value of each pixel and divide by the difference between the maximum and minimum values ​​to obtain the normalized pixel value, and write all the normalized pixel values ​​into the normalized saliency map according to the original pixel position; Set a fixed response threshold A comparison operation is performed on each pixel in the normalized saliency map. When the normalized pixel value is greater than or equal to... When the normalized pixel value is less than 1, the output value at that pixel position is set to 1. When the time is right, set the output value to 0 to form the final significance plot.

[0027] In this embodiment, generating the candidate lesion region map includes: Each pixel value in the saliency map is loaded into the saliency matrix according to its two-dimensional coordinates; pixels with a value of 1 in the matrix are set as foreground pixels, pixels with a value of 0 are set as background pixels, and an initial foreground pixel label matrix is ​​generated. Perform a fixed four-neighbor connected region scan on the foreground pixel labeling matrix. Iterate through each foreground pixel in each row and each column. If there are pixels that have been labeled as foreground pixels in the four adjacent positions above, below, left and right, assign the current pixel and the adjacent pixels the same connected region number. If there are no adjacent foreground pixels, assign a new number to the current pixel. After completing the traversal, generate a region index matrix containing all connected region numbers. Perform pixel counting on each of the non-zero numbered regions in the region index matrix and count the number of foreground pixels contained in each numbered region; set a pixel count threshold T, record all pixel coordinates of all numbered regions with a pixel count less than T in the elimination list, and clear the corresponding number value in the region index matrix to zero; For each region number in the reserved area, read all pixel coordinates corresponding to that number from the region index matrix, obtain the minimum row number, maximum row number, minimum column number, and maximum column number respectively, and determine a rectangular border whose coverage ranges from the minimum row number to the maximum row number and from the minimum column number to the maximum column number. Using the region number as the key, the border coordinate information and the set of pixel coordinates of each retained region are written into the lesion candidate region structure; all structures are written into the candidate region map matrix, and all pixels belonging to the candidate region are marked as 1 and other pixels are marked as 0 in binary form in the image coordinate system to generate a binary candidate lesion region map.

[0028] In this embodiment, generating the composite feature vector includes: Load the candidate lesion region map into a pixel value matrix, where each pixel position with a value of 1 represents a pixel within the candidate lesion region. Read the two-dimensional row and column coordinates of the positions with pixel values ​​of 1 one by one and store them into the lesion pixel coordinate list. The system receives a set of coupled feature data generated by the feature coupling module. This matrix corresponds one-to-one with the standard image space, and each matrix position stores a coupled feature vector. For each coordinate position in the list of lesion pixel coordinates, the corresponding coupled feature vector is read from the coupled feature data set, written into the lesion feature vector sequence in the order of reading, and the source coordinates of each vector are recorded. The lesion feature vector sequence is classified into regions according to the spatial continuity of the source coordinates. Coordinates with continuous row and column numbers are clustered and divided into independent lesion region groups. All coupled feature vectors in each group form a feature subset. For all coupled feature vectors in each feature subset, the mean is calculated along each dimension to obtain the average response value in each dimension of the subset, which is denoted as the mean vector of the subset. For each coupled feature vector in the same feature subset, calculate the squared Euclidean distance between it and the mean vector, and average all the squared Euclidean distances to obtain the variance index of the subset. Perform range calculation on the response values ​​in each dimension of the feature subset, find the maximum and minimum values ​​of each dimension, calculate the difference and arrange them in dimensional order to form the response range vector of the subset. The mean vector, variance index, and response range vector are concatenated in a predetermined order to form a composite feature vector for the lesion region. Repeat the above process for all lesion area groups, and write the composite feature vector generated for each group into the composite feature vector set.

[0029] In this embodiment, the generated diagnostic result dataset includes: Each composite feature vector in the composite feature vector set is loaded sequentially according to its index, and each composite feature vector is assigned a unique number to establish a mapping table between the number and the original lesion area. For each composite feature vector, numerical features are extracted in order of vector dimension to form an input vector, which is then input into the trained classification model. Forward computation is performed, and a probability distribution vector containing probability values ​​of several categories is output. In the probability distribution vector, obtain the target category index with the largest value, and write the category identifier corresponding to the index into the classification identifier information field; at the same time, use the maximum value as the confidence value of the corresponding lesion area and write it into the confidence information field. Based on the number of the composite feature vector, the coordinate information of the lesion region associated with the number is found in the original candidate lesion region map, and the row and column number, width and height parameters of the upper left corner of the border are extracted to generate image location indexing information. The classification identification information, image location indexing information, and confidence information are written into the result structure in a unified format, and all result structures are arranged in numerical order to form a diagnostic result dataset.

[0030] Example To verify the feasibility and effectiveness of the technical solution of this invention in actual clinical screening scenarios, the image recognition-based hypertensive retinopathy detection system proposed in this invention was deployed in the intelligent lesion screening platform of an ophthalmology center in a hospital. This hospital undertakes the task of retinal screening for hypertensive patients at the Chronic Disease Prevention and Control Center, collecting over 100,000 fundus color images annually. The existing manual interpretation process is time-consuming, highly subjective, and prone to missing early lesions under conditions of uneven lighting, vascular reflection, or slight blurring. To improve screening efficiency and accuracy, the hospital embedded the system of this invention into its existing image archiving and communication system starting in May 2024. This automated algorithm enables the extraction of structural features from fundus images, entropy response calculation, feature coupling mapping, saliency generation, and lesion region identification.

[0031] In this scenario, the system first receives raw images from a fundus camera, with a uniform resolution of 2048×1536 pixels. After the image preprocessing module starts, it performs three core technical operations: format normalization, noise suppression, and local contrast enhancement. Format normalization converts the image into a unified RGB channel structure using an adaptive grayscale mapping algorithm to ensure channel consistency in subsequent structural moment calculations. Noise suppression employs an improved bilateral filtering model, which suppresses high-frequency noise from imaging equipment or fundus reflections while maintaining the clarity of blood vessel edges. The local contrast enhancement uses a histogram-based illumination balancing strategy, resulting in smoother brightness gradients in the optic disc, macula, and surrounding areas, ensuring overall dynamic range stability.

[0032] After preprocessing, the image is input into the structural feature extraction module. This module divides the standard image into 32×32 non-overlapping rectangular structural blocks, each corresponding to 64×48 pixels. The system performs grayscale sequence generation within each block, reading the grayscale values ​​of adjacent pixels and recording them as a sequence of "+", "-", and "0" signs, forming a sign numerical grid through sign substitution. Based on the improved Krautchouk moment recursive formula, the module calculates multi-order moment coefficients for each structural block, where the horizontal and vertical basis function tables record the structural response values ​​of pixels distributed within their local ranges. The calculated local moment coefficients are accumulated to form a first-order moment value sequence, which is then concatenated in the scanning order to form the overall structural feature vector. Experiments show that this feature extraction method can still stably output feature sets in low-contrast images, with an average feature vector length of approximately 3200 dimensions, providing sufficient data support for subsequent response calculations.

[0033] The system then proceeds to the entropy response construction module. This module takes the set of structural features as input, calculates the difference between adjacent components for each feature vector, and performs sign mapping on the direction of the difference to generate a directional numerical sequence. The module statistically analyzes the occurrence ratios of the three components (1, 0, and -1) over a fixed-length interval, forming an interval weight vector. After accumulating these weight vectors to form a weight matrix, a Rényi entropy function with order parameter α=1.5 is introduced to perform exponentiation and logarithmic calculations, yielding the entropy response value corresponding to each structural feature. Unlike traditional Shannon entropy, the Rényi entropy function is more sensitive to confidence weights when dealing with local texture perturbations, thereby improving the response amplitude in small lesion areas. Calculations show that the system's output entropy response vector set exhibits a significant hierarchical distribution of information, with the entropy value in local abnormal regions being approximately 15% to 30% higher than the neighborhood average.

[0034] After completing the entropy response calculation, the system enters the feature coupling module. This module creates an M×K mapping matrix based on index differences to represent the relative positional relationship between structural feature components and entropy response components. Each element in the matrix represents the absolute index difference between the two components, which is normalized to obtain a column-normalized position sequence. The system performs an item-by-item multiplication operation in each column, using the entropy response component as the multiplier to obtain a coupling weight sequence. After all column operations are completed, the rows are summed to form a coupled feature vector, achieving full-space mapping and fusion of structural features and entropy response. This coupling method effectively avoids the defect of a single feature dimension affecting the overall detection in traditional algorithms, enabling the system to maintain high stability under complex image conditions.

[0035] Feature coupling data is input into the saliency mapping module. This module first constructs a two-dimensional coupled feature array based on the structural feature block size, calculates the average response value for each array element, and writes it into the response value array. Subsequently, the module maps the response values ​​to original image pixel blocks according to the array row and column indices, generating an intermediate saliency map. The saliency map is then processed by min-max normalization to form a normalized saliency map, with pixel values ​​limited to the range of 0–1. The system sets a fixed threshold of 0.65 and performs binarization on each pixel to generate the final saliency map. This saliency map visually presents a concentrated brightness in the lesion area and significant suppression of background noise, providing an accurate spatial distribution reference for subsequent candidate region extraction.

[0036] The candidate region extraction module receives the saliency map and loads it into a two-dimensional matrix. The system marks pixels with a value of 1 as foreground pixels, executes a four-neighbor connected component scan algorithm to identify spatially continuous regions, and counts the number of pixels in each region. The module sets a pixel threshold T=120 to eliminate isolated noise regions, retaining only stable spatial structures. For the retained regions, the system calculates the minimum and maximum row and column numbers, determines the rectangular bounding box, and generates a candidate region map. Each pixel with a value of 1 in the candidate region map corresponds to a suspected lesion region, and its bounding box information and pixel coordinates are stored in the candidate region data structure.

[0037] The composite feature construction module extracts coupled feature vectors corresponding to the coordinate positions from the candidate region map. The module aggregates features within the same lesion region according to spatial continuity to form a feature subset. It then calculates the mean vector by performing mean calculation on all features within the subset and averages the squared Euclidean distance between each feature and the mean vector to obtain a variance index. Subsequently, the module calculates the difference between the maximum and minimum values ​​in each dimension to obtain a response range vector. Finally, the mean vector, variance index, and response range vector are concatenated in a preset order to form a composite feature vector. This composite feature vector contains multidimensional information about the lesion region in terms of response intensity, morphological distribution, and structural changes, and is the core technical foundation for the diagnostic accuracy of this invention.

[0038] The diagnosis result generation module inputs the composite feature vector into a pre-trained classification model. The classification model employs a dual-channel convolutional structure, with the input channels corresponding to the composite feature vector and the region's gray-level statistical features, respectively. The model outputs a vector containing the probability distribution of each category. The system selects the category with the highest probability value as the final classification result and records the confidence score and the corresponding region coordinates. The resulting dataset is overlaid on the original image in the system interface with a rectangular border and color coding for doctors to verify.

[0039] In practical hospital applications, the average processing time from image acquisition to diagnostic output is 12.6 seconds, with image preprocessing taking approximately 2.3 seconds, structural feature extraction and entropy response calculation taking 5.8 seconds, and subsequent modules taking less than 5 seconds. Compared to the average 70 seconds per image for manual image interpretation, this invention significantly improves screening efficiency. From June to August 2024, the hospital selected 2000 hypertensive patients for a comparative experiment, with 1000 cases undergoing traditional manual interpretation and the other 1000 cases being automatically detected by this system. The experiment used lesion detection rate, false positive rate, localization accuracy, early lesion recognition rate, and physician consistency score as the main evaluation indicators. The comparison results are shown in the table below.

[0040] Table 1 Comparison of detection results between traditional manual interpretation and the system of this invention. ; As shown in Table 1, the system of this invention significantly outperforms traditional manual interpretation methods in terms of detection speed, recognition accuracy, and positioning precision. Especially under low-light or slightly blurred image conditions, the system maintains a detection stability of over 90%, demonstrating its robustness in complex environments. In early lesion detection, the system effectively amplifies minute gray-scale differences through a structure-response coupling mechanism, enabling the identification of microbleeds and mild hard exudative areas that are difficult to discern manually. During review, doctors found that approximately 92% of the high-confidence lesions annotated by the system were consistent with actual pathology, proving the algorithm's high clinical reliability.

[0041] According to statistics from the hospital's Clinical Information Center, in the three months after the system went live, the average number of images doctors reviewed per day increased from 120 to 310, significantly reducing their workload. In one batch screening, the system completed the analysis of approximately 18,000 fundus images within 6 hours, with all results synchronously transmitted back to the server. Doctors only needed to confirm suspicious lesions on the terminal interface, greatly improving the efficiency of medical resource utilization.

[0042] From a technical perspective, this invention addresses key issues in existing technologies through the following features. First, the improved Krawtchouk moment structure analysis model achieves discrete orthogonal feature extraction at the image patch level, preserving structural edge features in low signal-to-noise ratio environments. Second, the introduction of the Rényi entropy response function replaces the traditional statistical entropy model, resulting in a smoother distribution of response values ​​under local gray-level perturbations and enhancing the stability of lesion features. Third, the feature-entropy coupling matrix constructed based on index differences achieves spatial fusion of structural information and information content, making feature representation more globally consistent. Finally, through a combination of saliency mapping and candidate region selection, the system effectively reduces false detection regions and improves the confidence of the final classification.

[0043] In summary, this embodiment demonstrates that the present invention achieves fully intelligent processing of hypertensive retinopathy from image preprocessing to automatic diagnosis through the coordinated operation of multiple algorithm modules, including structural feature extraction, entropy response construction, feature coupling mapping, saliency generation, and candidate region analysis. This system can complete high-precision lesion detection in a short time, significantly improving the early lesion detection rate, reducing the workload of doctors, and possesses significant clinical application value and promotion potential.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A hypertensive retinopathy detection system based on image recognition, characterized in that, include: The image preprocessing module receives fundus images and performs format conversion, noise suppression, and image enhancement on the fundus images to generate standard images. The structural feature extraction module performs structural analysis on standard images based on improved Krautchouk moments to generate a set of structural features. The entropy response construction module calculates the response value of each structural feature in the structural feature set based on the Rényi entropy function, and generates a set of entropy response vectors. The feature coupling module maps the set of structural features to the set of entropy response vectors to generate coupled feature data. The saliency mapping module maps coupled feature data to a standard image space and generates a saliency map based on a preset response threshold. The candidate region extraction module judges the spatial continuity and distribution stability of the coupled feature data in the saliency map and generates a candidate lesion region map. The composite feature construction module extracts coupled feature data from the candidate lesion area map and generates a composite feature vector; The diagnostic result generation module generates classification identification information, image location indexing information, and confidence information based on composite feature vectors.

2. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, The generated structural feature set includes: The standard image is divided into sections based on a fixed number of rows and columns. For each non-overlapping rectangular region, the gray values ​​of adjacent pixels are read row by row, the increase or decrease relationship of the gray values ​​of adjacent pixels is compared, and recorded as "+", "-" or "0" in turn. The recorded results are arranged in row order to form a symbol sequence grid. Perform a sign substitution operation on the sign sequence grid, replacing the sign "+" with a positive number, the sign "-" with a negative number, and the sign "0" with a zero value; Set a position index for each row of the symbolic numerical grid. ,in An integer incrementing from zero, representing the horizontal pixel position within the current structural block; Let the width of the current structure block be... , representing the total number of pixels contained in the row; let the projection order be . , indicating the current expansion level of the Krawtchouk basis functions; Will Assign the value to the constant 1, and The initial value is defined according to the formula. ,in This is a probability parameter, with a value between 0 and 1; For all orders In the location index The Krawtchouk basis function values ​​are calculated using a recursive formula, and each calculation result is written into the corresponding column of the horizontal basis function table in ascending order of order. Write all recursive results into the horizontal basis function table, and then... OK, List organization; Each column in each structural block is processed in the same way as the horizontal basis function table is generated, and the vertical pixel index is set to . The column width is Repeat the recursive calculation process and sort the corresponding results according to... OK, Columns are written into the vertical basis function table; For each pixel position Read the pixel value at the location from the symbolic numerical grid, and read the basis function values ​​of the corresponding order from the horizontal and vertical basis function tables respectively. Multiply the three values ​​as the moment product value of the current projection point and record them in the moment coefficient accumulation table of the corresponding order. After traversing all pixel positions, the product of moments of all orders in each structural block is accumulated and used as the local Krawtchouk moment coefficients corresponding to the structural block. The local moment coefficients of all structural blocks are written into the contiguous storage area in the order of the structural blocks to form a first-level moment value sequence. The first-level moment value sequence is further concatenated into an overall moment feature vector according to the scanning order, and divided into N moment value segments of fixed length. The difference comparison of corresponding items is performed on each pair of adjacent moment value segments in turn. If the value of the previous segment is greater than that of the next segment, it is recorded as "1", less than it is recorded as "-1", and equal to it is recorded as "0", generating a difference sign sequence. The local Krautchouk moment coefficient sequence, the global moment feature vector, and the difference sign sequence are combined sequentially to form a set of structural features.

3. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, The set of generated entropy response vectors includes: Read each structural feature vector sequentially from the set of structural features, and represent the current structural feature vector as a vector containing m real components. And establish an index number for the structural feature vector in memory; For each component in the current structural feature vector, calculate the adjacent difference sequentially, for indices from 1 to... Calculate the difference at each position Set a fixed difference threshold : when At position i, record the symbol "+"; when At position i, record the symbol "-"; when Less than or equal to At position i, record the symbol "0"; Arrange all symbols in index order to obtain the incremental direction symbol sequence; Perform numerical mapping on the incremental direction symbol sequence, replacing the symbol "+" at each position with the value 1, the symbol "-" with the value -1, and the symbol "0" with the value 0, generating a direction value sequence, and storing it in a one-dimensional array in the original index order; Set interval length parameter For positive integers, starting from the first element of the direction value sequence, proceed in order with each... Each element is used to divide adjacent intervals. For each complete interval, ... Iterate through each element and count the number of times the value 1 appears. The number of times the value 0 appears is recorded. The number of times the value -1 appears is recorded. and will Write the results into a ternary array in a fixed order as the statistical results for that interval; Calculate the total value for the ternary statistical results for each interval. Perform a weight calculation for each component: Divide by Obtain the proportion ,Will Divide by Obtain the proportion ,Will Divide by Obtain the proportion Generate a weight vector And write them into the weight vector array in interval order; The cumulative weight vector array is processed by keeping the weight vector of the first interval unchanged. Starting from the second interval, the weight vector of each interval is added to the weight vector of the immediately preceding interval according to the components to calculate the cumulative weight vector. The cumulative weight vector is then written into the cumulative weight array in interval order. Set the order parameter of the Rényi entropy function, apply the Rényi entropy function to each cumulative weight vector in the cumulative weight array, perform exponentiation, summation and logarithmic operations, and calculate the corresponding entropy response value; Write the entropy response value of each interval into a one-dimensional array according to the interval index to obtain the entropy response vector corresponding to the current structural feature vector; Repeat the above entropy response vector generation operation for all structural feature vectors in the structural feature set. Store the entropy response vector corresponding to each structural feature vector into the entropy response vector storage area in the order of the structural feature vectors in the structural feature set to form an entropy response vector set.

4. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, The generation of coupling feature data includes: Read the structural feature vectors sequentially from the set of structural features according to their index numbers, and represent each structural feature vector as a sequence of real number components. Read the entropy response vectors with the same index number from the set of entropy response vectors. Each entropy response vector contains a fixed number of entropy response values. Set the number of components in the current structural feature vector to . The number of components in the entropy response vector is set to ,according to and The numerical relationship constructs a size of Position index mapping matrix; Perform a linear transformation operation on each column of the position index mapping matrix, read the corresponding values ​​sequentially from the first row to the last row, divide each read value by the sum of all values ​​in that column to obtain the column normalization result, write the normalization result into the column vector to form the column normalized position sequence; For each column of normalized position sequence, perform an element-by-element multiplication operation with the corresponding component of the entropy response vector, using the entropy response component of the corresponding column as the multiplier to obtain the product sequence; Read all product sequences in column order, accumulate all column values ​​in each row, and write the accumulated result to the corresponding position of the coupled feature vector to obtain the complete coupled feature vector of the current structure feature vector; Repeat the above steps for generating complete coupled feature vectors for all structural feature vectors in the structural feature set; All generated coupled feature vectors are written sequentially into the coupled feature data storage area according to the original index order of the structural feature set, forming a coupled feature data set.

5. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, Generating a saliency map includes: The set of coupled feature data is organized into a two-dimensional coupled feature array according to the number of rows and columns used when constructing the feature set, where each element in the array corresponds to a coupled feature vector. For each coupled feature vector in the two-dimensional coupled feature array, starting from the first component of the vector, all components are read sequentially, and a summation operation is performed on all components. The summation value is divided by the number of components to obtain the response value corresponding to the vector. All response values ​​are written into the response value array according to the row and column indices of the array elements. Based on the block size of the standard image, starting from the top left corner of the standard image, and according to the row and column indices of the response value array, determine the starting row number, starting column number, row span, and column span of the corresponding pixel block in the standard image for each response value. For each response value, read the corresponding pixel block position, and write the response value into all pixel positions covered by the pixel block in a row-by-column manner, so that all pixel blocks are filled with the corresponding response value, and generate an intermediate saliency map. Perform maximum and minimum value calculations on all pixels in the intermediate saliency map to obtain normalized pixel values, and write all normalized pixel values ​​into the normalized saliency map according to their original pixel positions; Set a fixed response threshold A comparison operation is performed on each pixel in the normalized saliency map. When the normalized pixel value is greater than or equal to... When the normalized pixel value is less than 1, the output value at that pixel position is set to 1. When the time is right, set the output value to 0 to form the final significance plot.

6. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, The generation of candidate lesion region maps includes: Each pixel value in the saliency map is loaded into the saliency matrix according to its two-dimensional coordinates; Pixels with a value of 1 in the matrix are set as foreground pixels, and pixels with a value of 0 are set as background pixels, and an initial foreground pixel marker matrix is ​​generated; Perform a fixed four-neighbor connected region scan on the foreground pixel label matrix, and traverse each foreground pixel in each row and each column in turn. If there are pixels that have been marked as foreground pixels in its four adjacent positions above, below, left, and right, then assign the current pixel and its adjacent pixels the same connectivity number. If there are no adjacent foreground pixels, then assign a new number to the current pixel; After completing the traversal, a region index matrix containing the numbers of all connected regions is generated. Perform pixel counting on each of the non-zero numbered regions in the region index matrix and count the number of foreground pixels contained in each numbered region; set a pixel count threshold T, record all pixel coordinates of all numbered regions with a pixel count less than T in the elimination list, and clear the corresponding number value in the region index matrix to zero; For each region number in the reserved area, read all pixel coordinates corresponding to that number from the region index matrix, obtain the minimum row number, maximum row number, minimum column number, and maximum column number respectively, and determine a rectangular border whose coverage ranges from the minimum row number to the maximum row number and from the minimum column number to the maximum column number. Using the region number as the key, the border coordinate information and the set of pixel coordinates of each reserved region are written into the lesion candidate region structure; All structures are uniformly written into the candidate region map matrix, and all pixels belonging to the candidate region are marked as 1 and other pixels as 0 in the image coordinate system in binary form to generate the candidate lesion region map.

7. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, Generating composite feature vectors includes: Load the candidate lesion region map into a pixel value matrix, where each pixel position with a value of 1 represents a pixel within the candidate lesion region. Read the two-dimensional row and column coordinates of the positions with pixel values ​​of 1 one by one and store them into the lesion pixel coordinate list. Receive the set of coupled feature data, map the set of coupled feature data to the standard image space one by one, and store a coupled feature vector at each matrix position; For each coordinate position in the list of lesion pixel coordinates, the corresponding coupled feature vector is read from the coupled feature data set, written into the lesion feature vector sequence in the order of reading, and the source coordinates of each vector are recorded. The lesion feature vector sequence is classified into regions according to the spatial continuity of the source coordinates. Coordinates with continuous row and column numbers are clustered and divided into independent lesion region groups. All coupled feature vectors in each group form a feature subset. For all coupled feature vectors in each feature subset, the mean is calculated along each dimension to obtain the average response value in each dimension of the subset, which is denoted as the mean vector of the subset. For each coupled feature vector in the same feature subset, calculate the squared Euclidean distance between it and the mean vector, and average all the squared Euclidean distances to obtain the variance index of the subset. Perform range calculation on the response values ​​in each dimension of the feature subset, find the maximum and minimum values ​​of each dimension, calculate the difference and arrange them in dimensional order to form the response range vector of the subset. The mean vector, variance index, and response range vector are concatenated in a predetermined order to form a composite feature vector for the lesion region.

8. The image recognition-based hypertensive retinopathy detection system as described in claim 1, characterized in that, The dataset used to generate diagnostic results includes: Each composite feature vector in the composite feature vector set is loaded sequentially according to its index, and each composite feature vector is assigned a unique number to establish a mapping table between the number and the original lesion area. For each composite feature vector, numerical features are extracted in order of vector dimension to form an input vector, which is then input into the trained classification model. Forward computation is performed, and a probability distribution vector containing class probability values ​​is output. In the probability distribution vector, obtain the target category index with the largest value, and write the category identifier corresponding to the index into the classification identifier information field; at the same time, use the maximum value as the confidence value of the corresponding lesion area and write it into the confidence information field. Based on the number of the composite feature vector, the coordinate information of the lesion region associated with the number is found in the original candidate lesion region map, and the row and column number, width and height parameters of the upper left corner of the border are extracted to generate image location indexing information. The classification identification information, image location indexing information, and confidence information are written into the result structure in a unified format, and all result structures are arranged in numerical order to form a diagnostic result dataset.

Images