Intelligent processing method based on kidney pathological image

Abstract

The invention relates to the field of medical image processing, and provides an intelligent processing method based on a kidney pathological image, which comprises the following steps: after inputting an image to be processed, extracting morphological characteristics of each image block according to block codes; positioning a block in the layered map and storing cell nucleus distribution characteristics and weight scores; establishing an incidence matrix of morphological characteristics based on connectivity of four neighborhoods; generating a fusion processing result through multi-dimensional weighting; and finally, converting a result into a standard pathological code, and calculating confidence coefficient parameter storage of each block. According to the invention, the accuracy and efficiency of kidney pathological image processing can be improved, and intelligent processing is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and more particularly, to an intelligent processing method based on kidney pathology images. Background Art

[0002] In the field of medical image processing, traditional renal pathology processing mainly relies on pathologists' microscopic observation of renal tissue sections, identifying morphological changes in cells and tissues to determine the type and extent of the lesion. However, this method has some limitations. First, the pathologist's medical image processing may be affected by subjective factors such as experience and fatigue, resulting in inconsistent processing results. Second, renal pathology images contain a large amount of information, and it is difficult to comprehensively and accurately analyze and extract all relevant features, especially some subtle lesion features, with the naked eye alone. In addition, with the development of medical imaging technology, the amount of renal pathology image data continues to increase. Traditional processing methods have difficulty in efficiently processing and analyzing this massive amount of data, resulting in low processing efficiency and difficulty in meeting clinical needs for fast and accurate processing.

[0003] In recent years, computer-aided design (CAD) technology has been widely used in the field of medical imaging, providing a new approach to renal pathology analysis. CAD technology uses computer algorithms to analyze and process medical images, automatically extracting feature information from the images and assisting pathologists in their processing. However, existing CAD technology still has some shortcomings in renal pathology image analysis. First, existing methods primarily focus on analyzing a single type of feature, such as nuclear morphology or cellular distribution, while ignoring the complex interrelationships among multiple features in renal pathology images. The treatment of renal lesions often requires comprehensive consideration of multiple morphological features and their interrelationships, and the analysis of a single feature fails to fully capture the essence of the lesion. Second, existing CAD systems lack refined processing of image blocks during feature extraction and processing, making it impossible to accurately locate and analyze key image regions, limiting the accuracy and reliability of the results. Furthermore, existing CAD technology also presents challenges in converting processing results into standard pathology codes. The lack of effective code conversion methods makes it difficult to effectively integrate with existing pathology processing standards, limiting its application in clinical practice.

[0004] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing renal pathology processing method relies on the subjective judgment of pathologists, and there are problems of inconsistent processing results and low processing efficiency; the existing CAD technology is insufficient in the analysis of the correlation between multiple features in the analysis of renal pathology images, lacks refined processing of image blocks, and has defects in the encoding conversion of processing results, which makes it difficult to meet the clinical needs for fast, accurate and standardized renal pathology processing. Summary of the Invention

[0005] The present invention provides an intelligent processing method based on kidney pathology images, comprising: Input the scanned file of the kidney pathology image to be processed; Extracting and storing the morphological features of each image block in the scanned file according to the feature fusion parameters encoded by each image block; Locating the image block corresponding to each feature fusion parameter in the layered atlas, and saving the associated cell nucleus distribution feature and weight score value in the corresponding image block, wherein the weight score value represents the contribution of the image block in the feature fusion; Establishing a correlation matrix between the morphological features based on the connectivity of the four neighborhoods of the block; Generating a fusion processing result between the morphological features based on the correlation matrix and multi-dimensional weighting principle; The fusion processing result is converted into a standard pathology code, and the confidence parameter of each image block code of the fusion processing result is calculated and stored.

[0006] Furthermore, the processing confidence of the single image block coding is calculated by the following method: For a pathological slice, the gradient response value of each image block is calculated based on the convolution kernel size and feature map resolution parameters; For an image block of the pathological slice, obtaining characteristic sampling points of four boundaries of a region corresponding to the image block; Based on the gradient response difference between two adjacent characteristic sampling points, the characteristic intensity value of each sampling point is calculated respectively using the spatial continuity interpolation method; The maximum feature intensity value and the minimum feature intensity value of the sampling point are used as confidence parameters for encoding the image block.

[0007] Furthermore, the step of obtaining feature sampling points of four boundaries of the area corresponding to the image block includes: The midpoint coordinates of each of the four boundaries of the corresponding area of ​​the image block are used as feature sampling points, and the midpoint coordinates and the boundaries of adjacent image blocks form a continuous sampling area.

[0008] Furthermore, the pathology database calculates the four boundary coordinates of the associated image block code one by one, updates the feature fusion parameters based on the rectangular area formed by the four boundary coordinates, saves the weight score value of the associated image block code to the rectangular area, and marks the area where the weight score value exceeds the threshold as a key processing area.

[0009] Furthermore, generating the fusion processing result between the morphological features based on the correlation matrix and the multi-dimensional weighted principle refers to selecting two adjacent key areas in the feature map, one as the reference area and the other as the comparison area, using the reference area as the initial node, and generating a processing path from the reference area to the comparison area based on the following steps: T1: Load the node into the priority queue; T2: The priority queue finds unlabeled areas in the four-neighborhood connectivity areas of the node based on node feature similarity, loads the unlabeled areas into the priority queue, and selects the area with the highest weight score from the unlabeled areas as the updated node; updates the fusion parameters of the four-neighborhood connectivity areas based on the feature vector of the current node and the feature vectors stored in the four-neighborhood connectivity areas; Repeat step T2 until the four-neighborhood connectivity area of ​​the node includes the comparison area; T3: Connecting the region sequences corresponding to all nodes as the processing paths of the reference region and the comparison region.

[0010] Furthermore, the updating of the fusion parameters of the four-neighborhood connectivity regions based on the feature vector of the current node and the feature vectors stored in the four-neighborhood connectivity regions is completed by the following method, the method comprising: Performing weighted summation on the feature vector of the current node and the feature vectors stored in the four neighborhood connectivity areas of the current node to obtain fusion parameters of the neighborhood area; For the four-neighborhood connectivity area of ​​the current node that already exists in the priority queue, determine whether the fusion parameter of the neighborhood area is higher than the original fusion parameter of the area; if so, update the original fusion parameter of the area to the fusion parameter of the neighborhood area; otherwise, keep the original fusion parameter unchanged; For the four-neighborhood connectivity areas of the current node that are first loaded into the priority queue, the fusion parameters of the neighborhood areas are initialized to weighted sum results.

[0011] Furthermore, the fusion processing result is converted into a standard pathology code by the following method: Starting from the starting point of the processing path, the path is framed at intervals of a fixed window width; The feature coding operation is performed on each frame image area in sequence, and after coding redundancy removal processing and hierarchical aggregation processing, the standard pathology coding sequence is generated.

[0012] Furthermore, the feature encoding operation is performed on each frame image area in sequence by the following steps, which include: If the two endpoints of the image frame are in the same coding unit, the unit is used as the target unit. If the target unit does not exist in the feature coding set, the target unit is added to the feature coding set. If the two endpoints of the image frame are in different coding units, the two different coding units are added to the feature coding set, and a first candidate region is generated based on the coordinate ranges of the two units, low-response units are removed from the first candidate region, and units that meet the processing criteria are selected from the remaining units in the first candidate region and added to the feature coding set; The target unit or the first candidate area is expanded based on the fixed window width to generate a second candidate area, low-response units are eliminated from the second candidate area, and units that meet the processing criteria are selected from the remaining units of the second candidate area and added to the feature coding set.

[0013] Furthermore, the fixed window width is a histological feature scale, and a coding unit size matching the histological feature scale or any one of the two most adjacent level coding unit sizes is selected as a basic coding level of the feature coding set.

[0014] Furthermore, the hierarchical aggregation process includes the following steps: TA1: Arrange the coding units of the feature coding set in spatial order; TA2: Divide every four adjacent units in the arranged coding unit sequence into the same aggregation group; TA3: Determine whether the coding units in each aggregation group belong to the same parent unit and are spatially continuous. If so, merge the coding units in the same aggregation group into the parent unit to obtain an updated feature coding set, and return to step TA1; When the four adjacent coding units of no aggregation group can be merged, the process is terminated.

[0015] The above embodiments of the present invention have at least the following beneficial effects: 1. Through the dynamic calculation mechanism of feature fusion parameters and weighted scores, the contribution of different image blocks in processing can be accurately quantified. Combined with multi-level morphological feature analysis, the recognition accuracy of tiny lesions in renal pathology images can be significantly improved, solving the problems of missed diagnoses and misdiagnoses caused by visual fatigue or experience differences in traditional manual reading.

[0016] 2. The association matrix constructed based on four-neighborhood connectivity combined with a dynamic priority queue algorithm can intelligently establish the topological relationship of pathological features across blocks, realize automated association analysis of key processing areas, and overcome the shortcomings of traditional methods in insufficient analysis of the association characteristics of diffuse lesions and lesion edge transition zones.

[0017] 3. The coding conversion technology that combines frame processing with hierarchical aggregation can adaptively generate pathological coding sequences that meet clinical standards, ensure the integrity of the coding structure through redundancy elimination and spatial continuity verification, and effectively solve the sharing barriers of multi-center medical data caused by differences in coding systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which: Figure 1 A flowchart of an intelligent processing method based on kidney pathology images provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0019] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0020] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0021] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0022] Reference below Figure 1 , Figure 1 FIG1 is a flow chart of an intelligent processing method based on kidney pathology images provided by an embodiment of the present invention. Figure 1 As shown, an intelligent processing method based on kidney pathology images includes: S1. Input the scanned file of the kidney pathology image to be processed; S2. extracting and storing the morphological features of each image block in the scanned file according to the feature fusion parameters encoded by each image block; S3. Locating the image block corresponding to each feature fusion parameter in the layered atlas, and saving the associated cell nucleus distribution feature and weight score value in the corresponding image block, wherein the weight score value represents the contribution of the image block to the feature fusion; S4. establishing a correlation matrix among the morphological features based on the connectivity of the four neighborhoods of the block; S5. generating a fusion processing result between the morphological features based on the correlation matrix and the multi-dimensional weighting principle; S6. Convert the fusion processing result into a standard pathology code, and calculate the confidence parameter of each image block code of the fusion processing result for storage.

[0023] It should be noted that the present invention proposes an intelligent processing method based on renal pathology images. Its core lies in implementing intelligent processing of renal pathology images through a pathology database. The pathology database stores renal tissue slice images, lesion annotation data, and corresponding image block codes. These image block codes are identifiers used to distinguish different regions in the image. Each code is assigned feature fusion parameters, which are a set of parameters used to comprehensively analyze multiple morphological features within the image block. By inputting a scanned file of the renal pathology image to be processed, the method extracts and stores the morphological features of each image block in the scanned file based on the feature fusion parameters of each image block code. Morphological features herein refer to characteristics such as the shape, size, and distribution of cells and tissues in the image, which can reflect the pathological state of renal tissue. The image block corresponding to each feature fusion parameter is located in a hierarchical atlas. The hierarchical atlas is a map that divides the image into different layers for better analysis of image features. Corresponding image blocks store associated cell nucleus distribution features and weighted scores. The weighted scores indicate the contribution of each image block to the feature fusion, that is, the importance of the block in the processing process. Based on the four-neighborhood connectivity of the block, a correlation matrix is ​​established between morphological features. Four-neighborhood connectivity refers to the connection relationship between an image block and its four adjacent blocks above, below, left, and right. By establishing a correlation matrix through this relationship, the relationship between image blocks can be analyzed. Based on the correlation matrix and the multi-dimensional weighted principle, the fusion processing results between morphological features are generated. The multi-dimensional weighted principle refers to considering the weights of multiple dimensions when analyzing features and processing multiple factors comprehensively. The fusion processing results are converted into standard pathology codes, and the confidence parameters of the codes of each image block of the fusion processing results are calculated and stored. Standard pathology codes convert the processing results into a coding form that meets pathology standards, and the confidence parameters are parameters used to measure the reliability of the processing results.

[0024] Specifically, inputting a scanned file of a renal pathology image to be processed refers to inputting a digital image file generated by a scanning device of a renal tissue slice into the system. Image block codes are pre-set codes used to identify different regions in the image, with each code corresponding to a specific region in the image. Feature fusion parameters are a set of parameters pre-set based on the feature type and importance of the image block and are used to adjust the weights of different features during the feature fusion process. A hierarchical atlas is a map that divides an image into different layers, such as cells and tissues, for more detailed analysis of image features. Cell nucleus distribution features refer to features such as the location, number, and distribution density of cell nuclei in an image, which are important for pathology processing. The weight score value is a pre-set value based on the importance and processing value of the features in the image block and is used to measure the contribution of each image block during the feature fusion process. The correlation matrix is ​​a matrix established by analyzing the four-neighborhood connectivity between image blocks and is used to represent the relationships between image blocks. The multidimensional weighting principle considers the weights of multiple dimensions, such as nuclear distribution and cell size, when generating fusion processing results. Each dimension has a corresponding weight value, which is pre-set based on pathology knowledge and experience. Standard pathology coding converts the processing results into a coding format that conforms to pathology standards, facilitating integration with existing pathology processing standards. Confidence parameters are calculated based on the reliability of the processing results and are used to measure the credibility of the processing results.

[0025] Preferably, the feature fusion parameters can be adjusted based on the importance and processing value of different features in the image block, such as the nuclear distribution characteristics and cell size characteristics. For example, for image blocks where the nuclear distribution characteristics are more important, the weight of the nuclear distribution characteristics can be increased. When establishing the association matrix, the similarity between adjacent blocks can be calculated by analyzing the four-neighborhood connectivity between the image blocks. The similarity can be calculated based on the nuclear distribution characteristics, cell size characteristics, etc. Blocks with high similarity also have high association in the association matrix. When generating the fusion processing results, the features of the image blocks can be comprehensively analyzed through an algorithm based on the association matrix and multi-dimensional weighting principles. The algorithm can adopt a machine learning algorithm, with input parameters including the image block's feature vector and weight score value, and the output is the fusion processing result. When calculating the confidence parameter, it can be calculated based on the feature strength of the image block and the consistency of the processing results. For example, image blocks with high feature strength and good consistency of processing results will have higher confidence parameters.

[0026] In some embodiments, the processing confidence of single image block coding is calculated by the following method: For a pathological slice, the gradient response value of each image block is calculated based on the convolution kernel size and feature map resolution parameters; For an image block of the pathological slice, obtaining characteristic sampling points of four boundaries of a region corresponding to the image block; Based on the gradient response difference between two adjacent characteristic sampling points, the characteristic intensity value of each sampling point is calculated respectively using the spatial continuity interpolation method; The maximum feature intensity value and the minimum feature intensity value of the sampling point are used as confidence parameters for encoding the image block.

[0027] It should be noted that the method for calculating the processing confidence of single image block coding in the present invention is implemented by analyzing the features of the pathology slice image block. Specifically, the gradient response value of each image block is first calculated based on the convolution kernel size and feature map resolution parameter. Here, the convolution kernel size refers to the size of the convolution kernel used to extract features in image processing, and the feature map resolution parameter refers to the resolution of the generated feature map. These two parameters together determine the calculation accuracy of the gradient response value. Next, for an image block in the pathology slice, feature sampling points are obtained at the four boundaries of the corresponding area of ​​the image block. Feature sampling points are points selected on the image block boundary for feature analysis. Then, based on the gradient response difference between two adjacent feature sampling points, the feature intensity value of each sampling point is calculated using the spatial continuity interpolation method. The spatial continuity interpolation method is a method for inferring the feature value of an unknown point based on the feature value of a known point, which is used to ensure a smooth transition of feature intensity values. Finally, the maximum and minimum feature intensity values ​​of the sampling points are used as confidence parameters for the image block coding. The confidence parameters are used to measure the reliability of the image block coding processing result.

[0028] Specifically, the convolution kernel size is set based on the image's feature complexity and computational resources. For example, in image regions with densely populated cell nuclei, a smaller convolution kernel size can be used to extract finer features. In regions with sparsely populated cells, the convolution kernel size can be appropriately increased to improve computational efficiency. The feature map resolution parameter refers to the clarity of the generated feature map. High-resolution feature maps provide richer detail, but this increases computational complexity, so a balance needs to be struck based on actual needs. The four boundaries of an image block refer to the top, bottom, left, and right edges. Feature sampling points can be selected based on boundary length and feature distribution uniformity. For example, sampling points can be evenly spaced. The gradient response value reflects the rate of change of an image in a certain direction. By calculating the gradient response value, the edge and texture characteristics of an image block can be understood. The implementation of the spatial continuity interpolation method must consider the distance between sampling points and the gradient variation trend to ensure the accuracy of the interpolation results. The maximum and minimum feature intensity values ​​are obtained by comparing the feature intensity values ​​of all sampling points. These two values ​​reflect the strength of the image block's features and serve as confidence parameters to measure the reliability of the processing results.

[0029] Preferably, when calculating the gradient response value, a multi-scale convolution kernel can be used, that is, convolution kernels of different sizes are used simultaneously to extract features at different levels, so that the feature information of the image block can be captured more comprehensively. When selecting feature sampling points, the number and position of the sampling points can be dynamically adjusted according to the feature distribution of the image block, for example, the number of sampling points can be increased in areas where the features change dramatically. When calculating the feature intensity value, the texture features and cell nucleus distribution features of the image can be combined to comprehensively consider the impact of different features on the processing results in a weighted manner, and the weights can be set according to pathological knowledge and experience. For example, if the cell nucleus distribution feature is more critical to the processing, a higher weight can be given. When determining the confidence parameter, in addition to considering the maximum and minimum feature intensity values, statistics such as the standard deviation of the feature intensity value can also be introduced to more comprehensively evaluate the reliability of the processing results.

[0030] In some embodiments, obtaining the feature sampling points of the four boundaries of the area corresponding to the image block includes: using the midpoint coordinates of each of the four boundaries of the area corresponding to the image block as the feature sampling point, and the midpoint coordinates and the boundaries of adjacent image blocks form a continuous sampling area.

[0031] It should be noted that the method for obtaining feature sampling points of the four boundaries of the corresponding area of ​​the image block mentioned in the present invention is to more accurately analyze the feature distribution of the image block. Specifically, the method ensures that the sampling points can be evenly distributed on the boundary of the image block by taking the midpoint coordinates of each of the four boundaries of the corresponding area of ​​the image block as the feature sampling point. The midpoint coordinate refers to the midpoint position of the boundary segment. By selecting the midpoint as the sampling point, the feature changes on the boundary can be effectively reflected. In addition, the midpoint coordinates form a continuous sampling area with the boundaries of adjacent image blocks, which means that the sampling points can not only cover the boundaries of the current image block, but also form a continuous sampling area with the boundary sampling points of the adjacent blocks, thereby better capturing the continuity and correlation of the boundary features, and providing more comprehensive data support for subsequent feature analysis and processing.

[0032] Specifically, the four boundaries of the image block refer to the upper boundary, lower boundary, left boundary and right boundary of the image block, which separate the image block from other areas. The midpoint coordinates are obtained by calculating the midpoint position of each boundary line segment. For example, for a horizontal boundary, the midpoint coordinates can be determined by calculating the average value of the starting point and the end point of the boundary. An adjacent image block refers to a block that is adjacent to the current image block in spatial position, and its boundary is connected to the boundary of the current image block. A continuous sampling area refers to a continuous sampling area composed of the boundary sampling points of the current image block and the boundary sampling points of the adjacent image blocks. This continuity ensures that there will be no faults or omissions when analyzing features, making the feature analysis more complete and accurate. In this way, the feature changes on the image block boundary can be better captured, providing a more reliable basis for subsequent processing.

[0033] Preferably, to further enhance the representativeness of the feature sampling points, additional sampling points can be added in addition to the midpoint coordinates. For example, multiple sampling points can be selected at equal intervals along each boundary. In addition to the midpoint, these can also include points near the boundary's starting and ending points, as well as points in the middle of the boundary. This allows for a more comprehensive reflection of the characteristic variations along the boundary. When calculating the midpoint coordinates, more precise mathematical methods can be employed, such as taking into account the slope and length of the boundary segment, to ensure the accuracy of the midpoint coordinates.

[0034] Furthermore, when forming continuous sampling regions, the boundary sampling points of adjacent image blocks can be calibrated to ensure a smoother transition between sampling points and avoid feature analysis errors caused by inaccurate sampling point positions. These optimization measures can further improve the quality of feature sampling points, providing more accurate data support for intelligent processing of renal pathology images.

[0035] In some embodiments, the pathology database calculates the four boundary coordinates of the associated image block code one by one, updates the feature fusion parameters based on the rectangular area formed by the four boundary coordinates, saves the weight score value of the associated image block code to the rectangular area, and marks the area where the weight score value exceeds the threshold as a key processing area.

[0036] It should be noted that the pathology database in the present invention calculates the four boundary coordinates of the associated image block encoding, updates the feature fusion parameters based on the rectangular area formed by these boundary coordinates, and saves the weight score value to the rectangular area. This process aims to identify areas that are critical to processing through precise boundary positioning and parameter updating. Among them, the boundary coordinates refer to the coordinate positions of the starting point and the end point of the four boundaries of the image block. These coordinates are used to define the precise range of the image block. The rectangular area is the area surrounded by these four boundary coordinates, which is used to limit the update range of the feature fusion parameters. The weight score value is an indicator to measure the importance of the image block in the processing process. Saving it to the rectangular area is to quickly identify the key area in subsequent processing. In addition, when the weight score value exceeds the preset threshold, the area is marked as a key processing area. This operation helps to highlight those areas that have a greater impact on the processing results, thereby improving the accuracy and efficiency of the processing.

[0037] Specifically, the associated image block code refers to the code associated with a specific image block in the pathology database, which is used to uniquely identify the block. The four boundary coordinates refer to the coordinates of the upper boundary, lower boundary, left boundary, and right boundary of the image block. These coordinates can be automatically extracted by the image analysis algorithm or manually annotated by a pathologist. The definition of the rectangular area is based on these four boundary coordinates, and a minimum rectangular area containing the image block is determined by calculation. Feature fusion parameters refer to the parameters used in the feature fusion process to adjust and optimize the combination of features. These parameters can be dynamically updated based on the distribution of features within the rectangular area. The weight score value is a pre-set value based on the feature importance and processing value of the image block. It is saved in the rectangular area so that these weight values ​​can be quickly accessed and used in subsequent processing. The threshold is a pre-set value used to determine whether the weight score value is high enough to mark the area as a critical processing area. This threshold can be adjusted based on pathology experience and processing requirements.

[0038] Preferably, when calculating the four boundary coordinates, an image segmentation algorithm, such as an edge detection algorithm, can be used to accurately extract the boundaries of the image blocks. For the update of the rectangular area, the feature fusion parameters can be adjusted according to the density of the cell nucleus distribution, the uniformity of the cell size and other features within the rectangular area, for example, the weight of the area with dense cell nucleus distribution can be increased. When saving the weight score value, a data structure such as a hash table can be used to quickly store and retrieve these values. For the setting of the threshold, a statistical analysis can be performed based on different types of kidney pathology images in the pathology database to determine a reasonable threshold range. For example, by analyzing the distribution of weight score values ​​in a large number of pathology images, a threshold that can effectively distinguish between key areas and non-key areas can be selected.

[0039] Further, in order to further improve the identification accuracy of the key processing area, after marking the key processing area, secondary analysis can be performed on the area, such as feature extraction and classification by a deep learning model, to verify whether it is truly important for processing.

[0040] In some embodiments, the generating of the fusion processing result between the morphological features based on the association matrix and the multi-dimensional weighting principle refers to selecting two adjacent key regions in the feature map, one as a reference region and the other as a comparison region, taking the reference region as an initial node, and generating a processing path from the reference region to the comparison region based on the following steps: T1: loading the node into a priority queue; T2: the priority queue finds an unmarked region in the four-neighbor connectivity region of the node based on the node feature similarity, loads the unmarked region into the priority queue, and selects the region with the highest weight score from the unmarked region as an updated node; and updates the fusion parameters of the four-neighbor connectivity region based on the feature vector of the current node and the feature vector stored in the four-neighbor connectivity region; repeating step T2 until the four-neighbor connectivity region of the node contains the comparison region; T3: connecting all node corresponding region sequences as the processing path of the reference region and the comparison region.

[0041] It should be noted that the process of generating the fusion processing result between the morphological features based on the association matrix and the multi-dimensional weighting principle in the present application is realized by constructing a processing path. Specifically, this method selects two adjacent key regions in the feature map, one as a reference region and the other as a comparison region. The reference region is the starting point of the processing path, and the comparison region is the terminal point. The generation of the processing path is based on the four-neighbor connectivity of the node, the priority queue is used to manage the node, and the path is gradually expanded according to the node feature similarity until the comparison region is reached. The priority queue is a data structure used to process nodes in a specific priority order. The node feature similarity refers to the similarity of the current node and the surrounding nodes in terms of features, which is used to determine the expansion direction of the path. In this way, a processing path from the reference region to the comparison region can be generated, providing an effective path analysis method for intelligent processing of kidney pathological images.

[0042] Specifically, a feature map refers to a feature distribution map extracted from renal pathology images, encompassing morphological features such as cell nucleus distribution and cell size. Key regions are areas within the feature map that are important for processing and typically have high weight scores. The selection of baseline and comparison regions can be determined based on pathological knowledge and processing requirements. For example, regions with abnormal cell nucleus distribution can be selected as baseline regions, while regions with potentially related lesions can be selected as comparison regions. Nodes in the priority queue are image regions considered during path generation. Each node contains its feature vector and weight score. Four-neighborhood connectivity refers to the connection between a node and its four adjacent nodes above, below, and to the left and right. This relationship is used to determine the path extension direction. Node feature similarity can be calculated by calculating the similarity between node feature vectors, for example, using methods such as cosine similarity or Euclidean distance. During path generation, the region with the highest weight score is selected from the priority queue as a new node, and the fusion parameters of its four-neighborhood connectivity regions are updated until the comparison region is reached. Fusion parameters are parameters used to adjust feature fusion during path generation and can be dynamically adjusted based on node feature similarity.

[0043] Preferably, when constructing a processing path, a deep learning model can be used to extract key areas in the feature map. For example, a convolutional neural network CNN can be used to extract features from renal pathology images, and the input parameters include image pixel values ​​and pre-trained model weights. By training the model to identify features such as cell nucleus distribution and cell size, the model can output feature maps of key areas. When calculating the similarity of node features, a more complex similarity measurement method can be used, such as a comprehensive similarity measurement that combines multiple features, to improve the accuracy of path generation. For the management of priority queues, the nodes can be comprehensively sorted according to their weight score values ​​and feature similarity to ensure that the nodes selected each time are optimal. When updating the fusion parameters of the four-neighborhood connectivity area, a learning rate parameter can be introduced to control the update speed of the fusion parameters to avoid inaccurate path generation due to too fast or too slow updates.

[0044] Furthermore, in order to optimize the generation of the processing path, path smoothing processing may be performed after the path generation is completed, for example, by optimizing the path through an interpolation method to improve the continuity and interpretability of the processing path.

[0045] In some embodiments, the updating of the fusion parameters of the four-neighborhood connectivity regions based on the feature vector of the current node and the feature vectors stored in the four-neighborhood connectivity regions is accomplished by the following method, the method comprising: Performing weighted summation on the feature vector of the current node and the feature vectors stored in the four neighborhood connectivity areas of the current node to obtain fusion parameters of the neighborhood area; For the four-neighborhood connectivity area of ​​the current node that already exists in the priority queue, determine whether the fusion parameter of the neighborhood area is higher than the original fusion parameter of the area; if so, update the original fusion parameter of the area to the fusion parameter of the neighborhood area; otherwise, keep the original fusion parameter unchanged; For the four-neighborhood connectivity areas of the current node that are first loaded into the priority queue, the fusion parameters of the neighborhood areas are initialized to weighted sum results.

[0046] It should be noted that the method for updating the fusion parameters of the four-neighborhood connectivity area mentioned in the present invention is to dynamically adjust the feature fusion relationship between image blocks in the process of generating the processing path. Specifically, the method obtains new fusion parameters by weighted summing the feature vector of the current node and the feature vector stored in the four-neighborhood connectivity area. This process not only takes into account the feature information of the current node, but also combines the feature information of its neighborhood, thereby more comprehensively reflecting the correlation between image blocks. For the neighborhood area already in the priority queue, if the new fusion parameter is higher than the original fusion parameter, the fusion parameter of the area is updated; otherwise, the original fusion parameter remains unchanged. For the neighborhood area loaded into the priority queue for the first time, the weighted summation result is directly initialized as the fusion parameter. This method can ensure that the update of the fusion parameter is based on the latest feature information and can dynamically reflect the changes in the relationship between image blocks.

[0047] Specifically, the feature vector of the current node refers to the feature representation of the image block currently being processed. It contains morphological information about the block, such as cell nucleus distribution and cell size. The four-neighborhood connectivity regions refer to the four regions spatially adjacent to the current node, connected to the current node in the image from top to bottom and left to right. The weighted summation of feature vectors linearly combines the feature vector of the current node with the feature vectors of the neighboring regions according to certain weights. The weights can be set based on the importance of the features and the relative position of the neighborhood. A priority queue is a data structure used to store and manage nodes to be processed. The priority of a node can be determined based on its fusion parameters or other features. The fusion parameters of a neighborhood region refer to the feature fusion results of the neighborhood regions already calculated in the priority queue. They reflect the degree of correlation between the region and other regions. By comparing the new fusion parameters with the original fusion parameters, a decision can be made as to whether to update the fusion parameters. This process ensures the dynamic and accurate nature of the fusion parameters.

[0048] Preferably, an adaptive weight adjustment strategy can be adopted when performing the weighted summation of feature vectors. For example, the weight is dynamically adjusted based on the spatial distance and feature similarity between the current node and the neighboring area. If the two areas are close in space and have high feature similarity, a larger weight is given; otherwise, a smaller weight is given. When updating the fusion parameters, a threshold can be introduced to determine whether an update is needed. This threshold can be determined based on experience or through a data-driven method. For example, if the difference between the new fusion parameter and the original fusion parameter is greater than a certain threshold, it is considered that the new fusion parameter has sufficient information and is worth updating.

[0049] Furthermore, when a neighborhood region is first loaded into the priority queue, in addition to initializing the fusion parameters, other relevant information about the region, such as the source of the feature vector and the calculation time, can be recorded for tracing and analysis in subsequent processing. These optimization measures can further improve the accuracy and efficiency of fusion parameter updates, thereby better supporting the intelligent processing of renal pathology images.

[0050] In some embodiments, the fusion processing result is converted into a standard pathology code by the following method: Starting from the starting point of the processing path, the path is framed at intervals of a fixed window width; The feature coding operation is performed on each frame image area in sequence, and after coding redundancy removal processing and hierarchical aggregation processing, the standard pathology coding sequence is generated.

[0051] It should be noted that the process of converting the fusion processing results into standard pathology codes in the present invention is to enable the processing results to be effectively connected with existing pathology standards. Specifically, the process starts from the starting point of the processing path and performs frame processing on the path at intervals of a fixed window width. The fixed window width here refers to the width of the image frame divided on the path, which determines the size of each frame of the image. Feature coding operations are performed on each image frame area in sequence. Feature coding operations refer to converting the feature information in the image frame into a coding form. After coding de-redundancy processing and hierarchical aggregation processing, a standard pathology coding sequence is generated. Coding de-redundancy processing is to remove repeated or redundant coding information, while hierarchical aggregation processing is to merge the coding information according to a certain hierarchical structure to generate a more concise coding sequence. The standard pathology coding sequence finally generated can accurately reflect the processing results of the renal pathology image and meet the pathology standards, which is convenient for pathologists to conduct further analysis and judgment.

[0052] Specifically, the fixed window width is selected based on the histological feature scale, which refers to the typical size range of cells and tissues in renal pathology images. A coding unit size that matches the histological feature scale can be selected, or one of the two most adjacent hierarchical coding unit sizes can be selected as the basic coding level. The coding unit size refers to the basic unit size used to divide the image during the encoding process. Coding de-redundancy processing refers to the removal of repeated coding information through an algorithm. For example, if adjacent coding units have the same feature information, they can be merged into one coding unit. Hierarchical aggregation processing refers to merging coding units according to a certain hierarchical structure. For example, coding units with similar features are merged into a higher-level coding unit. The standard pathology coding sequence refers to converting the fusion processing results into a coding form that meets the pathology standards. This coding form can be recognized and processed by the pathology system.

[0053] Preferably, when performing feature encoding operations, a deep learning model can be used to extract features of each image region. For example, a convolutional neural network (CNN) can be used to extract features from each image frame, and the input parameters include the pixel values ​​of the image frame and the pre-trained model weights. The model can output feature vectors for each image frame, which will be used to generate the code. In the process of coding redundancy removal, a clustering algorithm, such as the K-means algorithm, can be used to cluster similar coding units together to remove redundant information. In the hierarchical aggregation process, a hierarchical structure can be defined, for example, the coding units can be layered according to cell type, tissue type, etc., and then aggregated in each layer. In addition, in order to improve the accuracy and efficiency of the coding, a verification step can be performed after generating the standard pathology coding sequence to ensure the correctness of the generated coding sequence by comparing it with the known pathology coding.

[0054] In some embodiments, the feature encoding operation is performed on each image region in sequence by the following steps, the steps including: If the two endpoints of the image frame are in the same coding unit, the unit is used as the target unit. If the target unit does not exist in the feature coding set, the target unit is added to the feature coding set. If the two endpoints of the image frame are in different coding units, the two different coding units are added to the feature coding set, and a first candidate region is generated based on the coordinate ranges of the two units, low-response units are removed from the first candidate region, and units that meet the processing criteria are selected from the remaining units in the first candidate region and added to the feature coding set; The target unit or the first candidate area is expanded based on the fixed window width to generate a second candidate area, low-response units are eliminated from the second candidate area, and units that meet the processing criteria are selected from the remaining units of the second candidate area and added to the feature coding set.

[0055] It should be noted that the process of performing feature coding operations on each frame image area in sequence in the present invention is to convert the feature information in the image frame into a standard pathological code. Specifically, the process first determines whether the two endpoints of the image frame are in the same coding unit. If they are in the same coding unit, the unit is used as the target unit; if they are not in the same coding unit, the two different coding units are added to the feature coding set, and a first candidate area is generated based on the coordinate range of the two units. Next, the low-response units are removed from the first candidate area, and the units that meet the processing standards are selected from the remaining units to be added to the feature coding set. Finally, the target unit or the first candidate area is expanded based on the fixed window width to generate a second candidate area, and the low-response units are removed from the second candidate area, and the units that meet the processing standards are selected to be added to the feature coding set. This process ensures the accuracy and completeness of the feature coding through gradual screening and expansion.

[0056] Specifically, an image frame refers to an image region divided according to a fixed window width on the processing path, and each image frame contains information about a portion of the processing path. Endpoints refer to the starting and ending points of an image frame, and their positions determine the range of the image frame. A coding unit refers to the basic unit used to represent image features during the feature encoding process, and each coding unit corresponds to a specific feature or region in the image. A target unit refers to a coding unit that is the target unit when the two endpoints of an image frame are in the same coding unit. A first candidate region refers to a region generated based on the coordinate range of two coding units when the two endpoints of an image frame are in different coding units, and is used to further screen units that meet the processing criteria. A low-response unit refers to a unit with a low response value in feature analysis, and these units may not contain sufficient processing information. A second candidate region refers to a region generated by expanding the target unit or the first candidate region based on a fixed window width, and is used to further expand the search range to ensure that important features are not missed.

[0057] Preferably, when generating the first candidate region, an image segmentation algorithm can be used to accurately delineate the boundaries of the coding units, such as by determining the unit outlines based on an edge detection algorithm. When eliminating low-response units, a response threshold can be set; only units with response values ​​above this threshold are considered valid. When generating the second candidate region, the fixed window width can be dynamically adjusted, for example, based on the distribution density of features in the image frame, to better adapt to feature variations in different regions.

[0058] Furthermore, to improve the accuracy of feature encoding, machine learning models can be introduced to assist in the selection of cells that meet the processing criteria. For example, a support vector machine (SVM) model can be used. The input parameters include the cell's feature vector and the known processing criteria, and the model outputs whether the cell meets the processing criteria. These optimization measures can further improve the efficiency and accuracy of feature encoding operations, thereby better supporting the intelligent processing of renal pathology images.

[0059] In some embodiments, the fixed window width is a histological feature scale, and a coding unit size matching the histological feature scale or any one of the two nearest level coding unit sizes is selected as the basic coding level of the feature coding set.

[0060] It should be noted that the fixed window width selection mentioned in the present invention is determined based on the histological feature scale in order to better match the cell and tissue structure in the renal pathology image. Specifically, a coding unit size that matches the histological feature scale can be selected, or one of the two most adjacent level coding unit sizes can be selected as the basic coding level of the feature coding set. This selection method can ensure that when performing feature encoding, the size of the coding unit matches the actual cell and tissue structure in the renal pathology image, thereby improving the accuracy and efficiency of the encoding. In this way, the characteristic information of the renal pathology image can be better reflected, providing a more accurate basis for subsequent processing.

[0061] Specifically, the histological feature scale refers to the typical size range of cells and tissues in renal pathology images, such as the diameter of the cell nucleus, the size of the cell, etc. The coding unit size refers to the basic unit size used to divide the image during the feature encoding process, which determines the range of the image area that each coding unit can cover. The basic coding level refers to the lowest-level coding unit used to represent image features in the feature coding set, which directly affects the granularity and accuracy of the encoding. Selecting a coding unit size that matches the histological feature scale means that the size of the coding unit can adapt to the actual size of cells and tissues in the image, thereby more accurately capturing the feature information of the image. The two most adjacent level coding unit sizes refer to the two coding unit sizes that are closest to the histological feature scale in the coding hierarchy structure. Selecting one of them as the basic coding level can improve the coding efficiency while ensuring the coding accuracy.

[0062] Preferably, when selecting a fixed window width, dynamic adjustments can be made based on the specific characteristics of renal pathology images. For example, in areas with densely populated cell nuclei, a smaller coding unit size can be selected as the fixed window width to more precisely capture the characteristics of the cell nuclei; whereas in areas with sparsely populated cells, a larger coding unit size can be selected to improve encoding efficiency.

[0063] Furthermore, when determining the base coding level, an adaptive algorithm can be introduced to automatically select the most appropriate coding unit size based on the feature complexity of different image regions. For example, an image analysis algorithm can be used to assess the feature complexity of each region and then select a matching coding unit size based on the assessment results. In practical applications, the fixed window width and base coding level can be fine-tuned based on the pathologist's experience and processing needs to further improve coding accuracy and adaptability.

[0064] In some embodiments, the hierarchical aggregation process includes the following steps: TA1: Arrange the coding units of the feature coding set in spatial order; TA2: Divide every four adjacent units in the arranged coding unit sequence into the same aggregation group; TA3: Determine whether the coding units in each aggregation group belong to the same parent unit and are spatially continuous. If so, merge the coding units in the same aggregation group into the parent unit to obtain an updated feature coding set, and return to step TA1; When the four adjacent coding units of no aggregation group can be merged, the process is terminated.

[0065] It should be noted that the hierarchical aggregation process mentioned in the present invention is used to merge the coding units in the feature coding set to generate a more concise and hierarchical standard pathology coding sequence. Specifically, the process first arranges the coding units of the feature coding set in spatial order, and then divides every 4 adjacent units in the arranged coding unit sequence into the same aggregation group. Then, it is judged whether the coding units in each aggregation group belong to the same parent unit and the spatial distribution is continuous. If so, each coding unit of the same aggregation group is merged into the parent unit to obtain the updated feature coding set. This process is repeated until it can no longer be merged. The purpose of the hierarchical aggregation process is to reduce the redundancy of coding by merging coding units with similar features and spatial continuity, while retaining important processing information, so that the final generated standard pathology coding sequence is more concise and efficient.

[0066] Specifically, a feature coding set refers to the set of all coding units, which are feature information extracted from an image frame. A coding unit is the basic unit of feature coding, and each coding unit represents a specific region or feature in the image. Spatial order refers to the ordering of coding units based on their position in the image, typically arranged from left to right and top to bottom. An aggregation group is a temporary set of adjacent coding units combined during the aggregation process. A parent unit is a higher-level coding unit in a hierarchical structure that contains multiple child coding units. Spatially continuous distribution means that the positions of coding units in the image are continuous and without gaps. By determining whether coding units belong to the same parent unit and are spatially continuous, it is possible to determine whether to merge them into a single parent unit. This process helps reduce coding redundancy while preserving important processing information.

[0067] Preferably, when performing hierarchical aggregation, more aggregation rules can be introduced to improve the efficiency and accuracy of aggregation. For example, in addition to determining whether the coding units belong to the same parent unit and have continuous spatial distribution, the feature similarity of the coding units can also be considered. If the feature similarity of two coding units is above a certain threshold, they can be merged into one parent unit even if their spatial distribution is not completely continuous.

[0068] Furthermore, when merging coding units, a weighted average approach can be used to calculate the eigenvalues ​​of the parent unit, with weights set based on the confidence parameters of the coding units. For example, coding units with higher confidence parameters can be given greater weight when calculating the eigenvalues ​​of the parent unit. In practical applications, the aggregation rules can be adjusted and optimized based on the pathologist's experience and processing needs to further improve the effectiveness of the aggregation process.

[0069] The above embodiments of the present invention have the following beneficial effects: 1. Through the dynamic calculation mechanism of feature fusion parameters and weighted scores, the contribution of different image blocks in processing can be accurately quantified. Combined with multi-level morphological feature analysis, the recognition accuracy of tiny lesions in renal pathology images can be significantly improved, solving the problems of missed diagnoses and misdiagnoses caused by visual fatigue or experience differences in traditional manual reading. 2. The association matrix constructed based on four-neighborhood connectivity combined with a dynamic priority queue algorithm can intelligently establish the topological relationship of pathological features across blocks, realize automated association analysis of key processing areas, and overcome the shortcomings of traditional methods in insufficient analysis of the association characteristics of diffuse lesions and lesion edge transition zones. 3. The coding conversion technology that combines frame processing with hierarchical aggregation can adaptively generate pathological coding sequences that meet clinical standards, ensure the integrity of the coding structure through redundancy elimination and spatial continuity verification, and effectively solve the sharing barriers of multi-center medical data caused by differences in coding systems.

[0070] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, server, mobile phone, or tablet.

[0071] The above descriptions merely illustrate some preferred embodiments of the present invention and the underlying technical principles. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present invention is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. An intelligent processing method based on renal pathology images, characterized in that: The pathology database stores kidney tissue slice images, lesion annotation data, and corresponding image block codes. Each image block code is set with feature fusion parameters. The pathology database is intelligently processed by the following method: Input the scanned file of the kidney pathology image to be processed; Extracting and storing the morphological features of each image block in the scanned file according to the feature fusion parameters encoded by each image block; Locating the image block corresponding to each feature fusion parameter in the layered atlas, and saving the associated cell nucleus distribution feature and weight score value in the corresponding image block, wherein the weight score value represents the contribution of the image block in the feature fusion; Establishing a correlation matrix between the morphological features based on the connectivity of the four neighborhoods of the block; Generating a fusion processing result between the morphological features based on the correlation matrix and multi-dimensional weighting principle; The fusion processing result is converted into a standard pathology code, and the confidence parameter of each image block code of the fusion processing result is calculated and stored.

2. The intelligent processing method based on renal pathology images according to claim 1, characterized in that: The processing confidence of single image block coding is calculated by the following method: For a pathological slice, the gradient response value of each image block is calculated based on the convolution kernel size and feature map resolution parameters; For an image block of the pathological slice, obtaining characteristic sampling points of four boundaries of a region corresponding to the image block; Based on the gradient response difference between two adjacent characteristic sampling points, the characteristic intensity value of each sampling point is calculated respectively using the spatial continuity interpolation method; The maximum feature intensity value and the minimum feature intensity value of the sampling point are used as confidence parameters for encoding the image block.

3. The intelligent processing method based on renal pathology images according to claim 2, characterized in that: The acquiring of feature sampling points of the four boundaries of the region corresponding to the image block includes: using the midpoint coordinates of each of the four boundaries of the region corresponding to the image block as feature sampling points, wherein the midpoint coordinates and the boundaries of adjacent image blocks form a continuous sampling region.

4. The intelligent processing method based on renal pathology images according to claim 3, characterized in that: The pathology database calculates the four boundary coordinates of the associated image block code one by one, updates the feature fusion parameters based on the rectangular area formed by the four boundary coordinates, saves the weight score value of the associated image block code to the rectangular area, and marks the area where the weight score value exceeds the threshold as a key processing area.

5. The intelligent processing method based on renal pathology images according to claim 1, characterized in that: Generating the fusion processing result between the morphological features based on the correlation matrix and the multi-dimensional weighted principle refers to selecting two adjacent key areas in the feature map, one as the reference area and the other as the comparison area, using the reference area as the initial node, and generating a processing path from the reference area to the comparison area based on the following steps: T1: Load the node into the priority queue; T2: The priority queue finds unlabeled areas in the four-neighborhood connectivity areas of the node based on node feature similarity, loads the unlabeled areas into the priority queue, and selects the area with the highest weight score from the unlabeled areas as the updated node; updates the fusion parameters of the four-neighborhood connectivity areas based on the feature vector of the current node and the feature vectors stored in the four-neighborhood connectivity areas; Repeat step T2 until the four-neighborhood connectivity area of ​​the node includes the comparison area; T3: Connecting the region sequences corresponding to all nodes as the processing paths of the reference region and the comparison region.

6. The intelligent processing method based on renal pathology images according to claim 5, characterized in that: The updating of the fusion parameters of the four-neighborhood connectivity regions based on the feature vector of the current node and the feature vectors stored in the four-neighborhood connectivity regions is completed by the following method, the method comprising: Performing weighted summation on the feature vector of the current node and the feature vectors stored in the four neighborhood connectivity areas of the current node to obtain fusion parameters of the neighborhood area; For the four-neighborhood connectivity area of ​​the current node that already exists in the priority queue, determine whether the fusion parameter of the neighborhood area is higher than the original fusion parameter of the area; if so, update the original fusion parameter of the area to the fusion parameter of the neighborhood area; otherwise, keep the original fusion parameter unchanged; For the four-neighborhood connectivity areas of the current node that are first loaded into the priority queue, the fusion parameters of the neighborhood areas are initialized to weighted sum results.

7. The intelligent processing method based on renal pathology images according to claim 1, characterized in that: The fusion processing result is converted into a standard pathology code by the following method: Starting from the starting point of the processing path, the path is framed at intervals of a fixed window width; The feature coding operation is performed on each frame image area in sequence, and after coding redundancy removal processing and hierarchical aggregation processing, the standard pathology coding sequence is generated.

8. The intelligent processing method based on renal pathology images according to claim 7, characterized in that: The feature encoding operation is performed on each frame image area in sequence by the following steps, the steps comprising: If the two endpoints of the image frame are in the same coding unit, the unit is used as the target unit. If the target unit does not exist in the feature coding set, the target unit is added to the feature coding set. If the two endpoints of the image frame are in different coding units, the two different coding units are added to the feature coding set, and a first candidate region is generated based on the coordinate ranges of the two units, low-response units are removed from the first candidate region, and units that meet the processing criteria are selected from the remaining units in the first candidate region and added to the feature coding set; The target unit or the first candidate area is expanded based on the fixed window width to generate a second candidate area, low-response units are eliminated from the second candidate area, and units that meet the processing criteria are selected from the remaining units of the second candidate area and added to the feature coding set.

9. The intelligent processing method based on renal pathology images according to claim 7, characterized in that: The fixed window width is a histological characteristic scale, and a coding unit size matching the histological characteristic scale or any one of the two most adjacent level coding unit sizes is selected as a basic coding level of the feature coding set.

10. The intelligent processing method based on kidney pathology images according to claim 7, characterized in that: The hierarchical aggregation process includes the following steps: TA1: Arrange the coding units of the feature coding set in spatial order; TA2: Divide every four adjacent units in the arranged coding unit sequence into the same aggregation group; TA3: Determine whether the coding units in each aggregation group belong to the same parent unit and are spatially continuous. If so, merge the coding units in the same aggregation group into the parent unit to obtain an updated feature coding set, and return to step TA1; When the four adjacent coding units of no aggregation group can be merged, the process is terminated.

Images