An Adaptive Feature Processing Method for Pathological Section Images of Gastroenteritis

Abstract

This invention relates to the field of medical image processing technology and discloses an adaptive feature processing method for pathological slide images of gastroenteritis. The method includes constructing a multi-scale feature analysis model based on the histological features of the gastroenteritis pathological slides to extract texture and morphological features at different tissue levels in the pathological slides; then constructing an adaptive feature optimization strategy to dynamically adjust feature selection parameters based on the feature data output by the multi-scale feature analysis model; next, performing joint calculations on the multi-scale feature analysis model and the adaptive feature optimization strategy according to the selected pathological slide sample type to obtain an optimized feature dataset; finally, establishing a feature representation of the gastroenteritis pathological slide image based on the optimized feature dataset. This method can comprehensively extract pathological features, dynamically optimize feature selection parameters, improve the quality of the feature dataset and the accuracy of feature representation, adapt to different sample types, and meet the needs of pathological analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an adaptive feature processing method for pathological slide images of gastroenteritis. Background Technology

[0002] In the clinical diagnosis and pathological research of gastroenteritis, feature analysis of pathological slide images is a crucial means of understanding disease progression and assessing the severity of lesions. Traditional methods for processing pathological slide images largely rely on fixed-scale feature extraction patterns, capturing information only at a single tissue level and failing to comprehensively cover the structural features of tissues of different sizes and shapes within the slides. For example, in identifying inflammatory areas, traditional methods may focus only on local texture features at the cellular level, neglecting macroscopic morphological changes in the stroma. This results in extracted feature data that cannot fully reflect the pathological state, thus affecting the accuracy of subsequent diagnosis or research.

[0003] In existing technologies, feature selection parameters are mostly pre-set fixed values, which cannot be dynamically adjusted according to different types of pathological slide samples. Because there are significant individual differences in gastroenteritis pathological slide samples, such as variations in inflammation severity, lesion size, and tissue cell density, using fixed feature selection parameters can lead to the omission of some key features or the retention of excessive irrelevant features. This not only increases the complexity of subsequent data processing but may also reduce the accuracy of feature representation, affecting the judgment and analysis of the pathological state.

[0004] Traditional methods, in the process of feature extraction and optimization, lack effective integration and joint computation of multi-scale features, making it difficult to form feature datasets that can comprehensively and accurately reflect the information of pathological slide images. This makes the feature representations built based on these feature datasets unable to fully reflect the histological characteristics of gastroenteritis pathological slides, thus limiting their application value in clinical diagnosis, disease monitoring, and pathological mechanism research, and failing to meet the needs of high precision and high adaptability for feature processing of pathological slide images in practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive feature processing method for pathological slide images of gastroenteritis to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an adaptive feature processing method for pathological slide images of gastroenteritis, the method comprising:

[0007] A multi-scale feature analysis model was constructed based on the histological features of pathological sections of gastroenteritis. The multi-scale feature analysis model was used to extract the texture and morphological features of different tissue levels in the pathological sections.

[0008] An adaptive feature optimization strategy is constructed, which is used to dynamically adjust the feature selection parameters based on the feature data output by the multi-scale feature analysis model.

[0009] Based on the selected pathological slide sample type, the multi-scale feature analysis model and the adaptive feature optimization strategy are jointly calculated to obtain the optimized feature dataset;

[0010] Based on the optimized feature dataset, feature representations of pathological slide images of gastroenteritis are established.

[0011] Preferably, when the histological features include inflammatory areas, ulcer areas, and normal tissue areas, the construction of a multi-scale feature analysis model based on the histological features of gastroenteritis pathological sections includes:

[0012] Based on the inflammatory region, ulcer region, and normal tissue region, respectively construct an inflammation feature analysis unit, an ulcer feature analysis unit, and a normal tissue feature analysis unit;

[0013] The output data of the inflammation feature analysis unit and the input data of the ulcer feature analysis unit are subjected to feature alignment processing.

[0014] The output data of the ulcer feature analysis unit and the input data of the normal tissue feature analysis unit are subjected to feature normalization processing.

[0015] Preferably, the step of jointly operating the multi-scale feature analysis model and the adaptive feature optimization strategy based on the selected pathological slide sample type to obtain the optimized feature dataset includes:

[0016] Based on the selected pathological slide sample type, a feature selection parameter matrix is ​​generated using the adaptive feature optimization strategy;

[0017] Based on the feature selection parameter matrix, the feature data output by the inflammation feature analysis unit, the ulcer feature analysis unit and the normal tissue feature analysis unit are weighted and fused.

[0018] When the feature fusion calculation result meets the convergence condition, the optimized feature dataset is output.

[0019] Preferably, before performing joint calculations on the multi-scale feature analysis model and the adaptive feature optimization strategy based on the selected pathological slide sample type, the method further includes:

[0020] The standard pathological slide image is input into the inflammation feature analysis unit to calculate the feature extraction accuracy.

[0021] When the feature extraction accuracy reaches the first preset standard, the standard pathological slide image is input into the ulcer feature analysis unit to calculate the feature matching accuracy;

[0022] When the feature matching accuracy reaches the second preset standard, the standard pathological slide image is input into the normal tissue feature analysis unit to calculate the feature stability index.

[0023] Preferably, the step of establishing feature representations of gastroenteritis pathological slide images based on the optimized feature dataset includes:

[0024] When the optimized feature dataset satisfies the first and second judgment criteria, the feature representation is established.

[0025] The first criterion is that the discriminative power of the inflammatory region features in the optimized feature dataset is greater than a preset discriminative power threshold;

[0026] The second criterion is that the dispersion of the ulcer region features in the optimized feature dataset is less than a preset dispersion threshold.

[0027] Preferably, the method further includes:

[0028] When the feature representation fails to meet the expected quality requirements, the feature selection parameters in the adaptive feature optimization strategy are adjusted.

[0029] Based on the adjusted feature selection parameters, the joint operation of the multi-scale feature analysis model and the adaptive feature optimization strategy is re-executed until a feature representation that meets the quality requirements is obtained.

[0030] Preferably, the unit for constructing inflammatory feature analysis includes:

[0031] Based on the distribution characteristics of inflammatory cells, the inflammatory area is divided into multiple feature analysis blocks;

[0032] The feature analysis blocks are connected through a feature association network, and feature analysis priorities are set;

[0033] The inflammation feature analysis unit is constructed based on the feature analysis blocks and feature association network.

[0034] Preferably, the ulcer feature analysis unit includes:

[0035] Based on the morphological characteristics of the ulcer edge, the ulcer region is divided into multiple feature calculation blocks;

[0036] The feature calculation blocks are topologically connected according to their spatial distribution;

[0037] Based on the feature calculation blocks and topological connection relationships, the ulcer feature analysis unit is constructed.

[0038] Preferably, the construction of the normal tissue feature analysis unit includes:

[0039] The structural features of normal tissue are used as the direction of feature analysis. A feature sampling path is set based on the feature analysis direction, and the normal tissue feature analysis unit is constructed along the feature sampling path.

[0040] Preferably, the adaptive feature optimization strategy includes:

[0041] Calculate the importance score of each region feature output by the multi-scale feature analysis model;

[0042] A feature-optimized weight matrix is ​​generated based on the importance score;

[0043] The adaptive feature optimization strategy is constructed by dynamically optimizing the features of each region using the feature optimization weight matrix.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] This adaptive feature processing method for gastroenteritis pathological slide images constructs a multi-scale feature analysis model based on the histological features of gastroenteritis pathological slides. This model effectively extracts texture and morphological features at different tissue levels within the pathological slides. Compared to traditional fixed-scale feature extraction methods, the multi-scale feature analysis model covers information at different scales, from the cellular level to the stroma level, comprehensively capturing various key features in the pathological slides. This avoids feature omissions caused by single-scale limitations, ensuring that the extracted feature data more completely reflects the histological characteristics of gastroenteritis pathological slides, providing richer and more comprehensive foundational information for subsequent feature characterization.

[0046] In the feature optimization stage, the constructed adaptive feature optimization strategy dynamically adjusts the feature selection parameters based on the feature data output by the multi-scale feature analysis model. This dynamic adjustment mechanism can adaptively adjust for different types of pathological slide samples, fully considering individual differences in gastroenteritis pathological slide samples, such as the degree of inflammation, lesion size, and tissue cell density. It effectively screens out key features closely related to the pathological state while reducing interference from irrelevant features. This not only reduces the complexity of subsequent data processing but also improves the effectiveness and specificity of feature data, avoiding feature selection bias caused by fixed parameters and ensuring that feature data more accurately reflects the actual situation of the pathological slides.

[0047] By jointly operating a multi-scale feature analysis model and an adaptive feature optimization strategy based on the selected pathological slide sample type, the features extracted from multiple scales can be organically combined with dynamically optimized parameters to form an optimized feature dataset. This joint operation method fully leverages the advantages of the multi-scale feature analysis model in feature coverage and the adaptive feature optimization strategy in parameter adaptability, achieving a synergistic effect between feature extraction and optimization. This results in an optimized feature dataset that is both comprehensive and possesses high accuracy and effectiveness, more accurately reflecting the characteristic information of gastroenteritis pathological slides.

[0048] The image feature representation of gastroenteritis pathological slides, established based on the optimized feature dataset, can more comprehensively and accurately reflect the histological characteristics and pathological state of the slides. This feature representation can better adapt to the needs of different scenarios such as clinical diagnosis, disease monitoring, and pathological mechanism research, helping relevant personnel to more clearly identify lesion features in pathological slides, more accurately determine the degree of inflammation and lesion status, provide more reliable image feature support for the diagnosis and research of gastroenteritis, help improve the understanding and analysis of the pathological state of gastroenteritis, and promote the development of related fields.

[0049] This method does not rely on complex preset conditions or fixed processing modes, and has strong adaptability and flexibility. It can handle pathological slide samples of gastroenteritis from different sources and of different types, reducing the problem of unstable processing results caused by sample differences. It expands the scope of application of the method, making it more practical and operable in actual applications, and can better meet the diverse needs of the medical field for the feature processing of pathological slide images. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the adaptive feature processing method for pathological slide images of gastroenteritis described in this invention.

[0051] Figure 2 A schematic diagram illustrating the working principle of a multi-scale feature analysis model for histological features containing specific regions.

[0052] Figure 3 This diagram illustrates the working principle of the joint operation of the multi-scale feature analysis model and the adaptive feature optimization strategy. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides an adaptive feature processing method for pathological slide images of gastroenteritis, comprising feature data of key areas such as inflamed areas, ulcerated areas, and normal tissue areas. A multi-scale feature analysis model employs a hierarchical processing mechanism. First, the pathological slide images are preprocessed, including image enhancement and region segmentation, to identify different tissue regions. Subsequently, specialized feature extraction units are designed for each region, capable of operating at multiple scales, such as cellular, tissue, and overall image levels, thereby capturing feature information from microscopic to macroscopic levels. An adaptive feature optimization strategy is constructed, which dynamically adjusts feature selection parameters based on the feature data output by the multi-scale feature analysis model. The optimization process relies on feature importance assessment and weight allocation to ensure higher discriminative power and stability of the feature data in subsequent processing. Depending on the selected pathological slide sample type, the multi-scale feature analysis model and the adaptive feature optimization strategy are jointly computed. This joint computation involves feature data fusion and filtering, and iterative calculations ensure that the output data meets preset convergence conditions. Finally, a feature representation of the gastroenteritis pathological slide image is established based on the optimized feature dataset. This representation accurately reflects the pathological state and provides support for clinical diagnosis. The entire approach emphasizes adaptability and multi-scale analysis to accommodate the variability of different samples.

[0055] Example 1: See Figure 2 In constructing a multi-scale feature analysis model, specific feature analysis units need to be established for each of the specific histological features presented in pathological sections of gastroenteritis, namely, inflammatory areas, ulcer areas, and normal tissue areas, to achieve accurate feature capture. The construction of the inflammation feature analysis unit begins with the accurate identification and segmentation of inflammatory areas in pathological section images. These areas typically exhibit typical morphologies such as extensive inflammatory cell infiltration, tissue edema, and vascular congestion. Region growing algorithms in digital image processing technology or semantic segmentation models based on deep learning can effectively separate inflammatory areas from complex tissue backgrounds. After successfully segmenting the target area, this unit will perform multi-scale feature extraction operations. At the cellular level, it may analyze the density, types (such as neutrophils and lymphocytes), spatial distribution patterns, and morphological parameters such as nucleocytoplasmic ratio of inflammatory cells. At the tissue level, it focuses on the overall shape of the inflammatory foci, the irregularity of its boundaries, and its relative positional relationship with surrounding tissues. These local texture features and global morphological features obtained from different scales are integrated into a comprehensive feature vector to comprehensively describe the biological state of the inflammatory area.

[0056] The construction logic of the ulcer feature analysis unit is similar to that of the inflammation unit, but its focus is on tissue defect areas. Ulcers, in pathological sections, manifest as discontinuity of epithelial tissue, exposure of the lamina propria, and even deeper necrosis. The construction of this unit first relies on accurate edge detection algorithms to delineate the ulcer boundary. Morphological features such as ulcer depth, area, geometry (whether it is creeping), and the composition of granulation tissue at the base are the focus of the analysis. The feature extraction process also follows a multi-scale principle. At the microscale, it may analyze the regeneration of epithelial cells at the ulcer edge and the density of newly formed capillaries; at the macroscale, it quantifies the overall contour features of the ulcer and its distribution pattern on the mucosal surface. The data output by the ulcer feature analysis unit is a set of quantitative indicators that reflect the dynamic process of tissue destruction and repair.

[0057] The normal tissue feature analysis unit focuses on mucosal tissue unaffected by lesions. This tissue represents the standard physiological structure of the organ, characterized by orderly cell arrangement, intact glandular structure, and normal stroma ratio. This unit aims to establish a comparable baseline. Its feature extraction strategy emphasizes measuring the regularity and consistency of structures, such as the consistency of glandular size, shape, and orientation, as well as cell nucleus size and chromatin homogeneity. By analyzing the characteristics of these normal areas, a reference can be provided for the degree of deviation in lesion areas.

[0058] After constructing the three independent feature analysis units, their data streams need to be effectively integrated and coordinated. Feature alignment is a crucial step in connecting the inflammation feature analysis unit and the ulcer feature analysis unit. Since the inflammatory and ulcer regions may be spatially adjacent or overlapping, and features extracted from different units may exist in different scale spaces or coordinate systems, spatial registration is essential. This process may involve feature point-based matching techniques, such as finding corresponding marker points in the feature maps output by the two units, and then mapping the feature data output by the inflammation unit to a spatial frame consistent with the input data of the ulcer unit using a spatial transformation model (such as affine transformation or elastic registration). This ensures that features from different pathological changes can be correlated and analyzed within the same spatial context.

[0059] The next step is feature normalization, which applies a normalization process to the output of the ulcer feature analysis unit and the input of the normal tissue feature analysis unit. The goal of feature normalization is to eliminate systematic errors introduced by differences in staining, slice thickness, or image acquisition conditions, making data from ulcer regions (which typically have a large dynamic range of feature values) and normal tissue regions (where feature values ​​are relatively concentrated) comparable. A common method is Z-score normalization, which subtracts the mean from the data of each feature dimension output by the ulcer unit and then divides by its standard deviation, transforming it into a distribution with a mean of zero and a standard deviation of one; or min-max normalization, which linearly scales the feature values ​​to the [0,1] interval. Through this normalization, features from different sources are placed within a uniform numerical range, reducing the impact of non-biological variations on subsequent analysis.

[0060] The final multi-scale feature analysis model is a composite computational framework integrating three specific analysis units—inflammation, ulceration, and normal tissue—along with feature alignment and normalization preprocessing. This model can process pathological slide images in parallel or sequentially, extracting standardized and comparable texture and morphological feature datasets from different tissue levels (cells, tissues, and whole-body regions) and different pathological change regions (inflammation, ulceration, and normal tissue). This dataset comprehensively characterizes the microstructural information of gastroenteritis pathological slides, providing high-quality, multi-dimensional input data for subsequent adaptive feature optimization and the final feature representation.

[0061] Example 2: See Figure 3 The initiation of joint operations relies on the accurate determination of the pathological slide sample type. Sample types can be categorized according to clinical diagnosis into different phases, such as acute active phase, chronic persistent phase, and healing phase. The pathological features corresponding to each type differ significantly in proportion and degree. Based on this determination, an adaptive feature optimization strategy is activated to generate a dynamic feature selection parameter matrix. The generation logic of this matrix is ​​to match the focus of feature selection with the pathological characteristics of the sample. For example, for acute active phase samples, the matrix parameters are configured to assign higher weights to inflammation-related features (such as neutrophil infiltration density and tissue edema), while for chronic or predominantly ulcerated samples, the parameters are tilted towards dimensions such as the morphological features of the ulcer edge and the maturity of granulation tissue. This parameter matrix is ​​essentially a weighting blueprint that guides subsequent discriminative selection from the vast pool of original features.

[0062] Guided by the feature selection parameter matrix, the system begins a weighted fusion calculation of the output feature data from the inflammation feature analysis unit, the ulcer feature analysis unit, and the normal tissue feature analysis unit. This calculation is not a simple summation and averaging, but rather a linear or non-linear combination of the feature vectors contributed by each feature analysis unit based on the weighting coefficients specified in the parameter matrix. The inflammation unit may provide dynamic range data on cell infiltration, the ulcer unit describes the geometric morphological parameters of tissue defects, and the normal tissue unit provides baseline values ​​for structural integrity. The weighted fusion process needs to integrate this heterogeneous information into a unified, high-dimensional feature vector. The entire fusion calculation is an iterative optimization process. The system continuously monitors whether certain statistical properties of the output vector (e.g., the rate of change of its variance or the distance between feature vectors) reach the preset convergence condition. The convergence condition means that the fused feature set has stabilized and no longer changes significantly with the increase of the number of iterations. Once this condition is met, this stable, optimized feature dataset is finally output for subsequent use in building feature representations.

[0063] To ensure the reliability of the joint operation and the quality of the output feature dataset, the three core feature analysis units must undergo independent performance verification before initiating the operation. This is a progressive calibration process. The verification process begins by inputting known, accurately labeled standard pathological slide images (usually confirmed by consensus among pathology experts) into the inflammation feature analysis unit. After performing its feature extraction task, the unit's performance is evaluated by calculating the feature extraction accuracy. The accuracy calculation involves comparing the features automatically extracted by the unit with the gold standard annotation, using comprehensive indicators such as the F1 score to quantify its ability to correctly identify inflammatory region features. Only when the accuracy reaches a preset first performance standard threshold, indicating that the inflammation feature analysis unit is working reliably, will the system proceed to the next verification stage.

[0064] Standard pathological slide images are input into the ulcer feature analysis unit. At this stage, feature matching accuracy needs to be calculated. Matching accuracy focuses on the degree of spatial and feature agreement between the ulcer features extracted by the unit and the actual ulcer area in the image. It assesses the accuracy of feature localization and description, not just correct classification. After the matching accuracy meets the second preset standard, the verification process proceeds to the final step: testing the normal tissue feature analysis unit. After inputting the standard image into the unit, a feature stability index needs to be calculated. This index is typically obtained by performing multiple sampling analyses at different time points or on the same region, and then calculating the coefficient of variation or intraclass correlation coefficient of the obtained feature values. It aims to measure the consistency and repeatability of the unit's output results. Only when these three unit performance indicators—inflammatory feature extraction accuracy, ulcer feature matching accuracy, and normal tissue feature stability—each meet their respective standards, does the system determine that the multi-scale feature analysis model is ready to proceed to the joint computation stage with the adaptive feature optimization strategy. This joint computational process, which includes pre-validation, ensures that the final optimized feature dataset is built upon the premise that each feature analysis unit has passed rigorous performance testing. The dynamic generation of the feature selection parameter matrix allows the processing flow to adapt to samples with different clinical manifestations, while weighted fusion and convergence judgment guarantee the stability and validity of the output data. The entire implementation process embodies a systematic and rigorous methodology that emphasizes pre-validation.

[0065] Example 3: Establishing feature representations is a conditional decision-making process, not a simple transformation of the optimized feature dataset. This system sets the first criterion as the discriminative power of inflammatory region features must be greater than a preset discriminative power threshold. This discriminative power quantifies the degree to which inflammatory feature vectors are separable from other types of feature vectors (especially normal tissue features) in the feature space. Its calculation depends on measuring the margin between different categories of feature clusters. One specific implementation is to quantify this by calculating the ratio of the average intra-class distance of inflammatory features to the average inter-class distance between inflammatory and normal features. The calculation formula can be expressed as:

[0066]

[0067] Where: symbol This represents the calculated discrimination index; a higher value indicates that the inflammatory feature is more easily distinguished from other features. (Symbol) The Euclidean distance, representing the global center of the feature set of inflamed areas and the global center of the feature set of normal tissue areas, reflects the degree of spatial separation between different categories of features. (Symbol) This represents the average distance from all feature points within the feature set of the inflammatory region to the center of its own set, reflecting the density of clustering of similar features. (Symbol) It is a very small positive constant, its purpose being to prevent the denominator from being zero and to ensure the mathematical stability of the formula. The preset discrimination threshold is a threshold value determined based on a large number of prior experiments or clinical needs; it is only applied when the calculated value is zero. The first criterion is considered satisfied only when the value exceeds this threshold.

[0068] The system employs a second criterion, requiring that the dispersion of ulcer region features be less than a preset dispersion threshold. Dispersion focuses on the central tendency of the ulcer feature vector's distribution; higher dispersion may indicate instability or inconsistency in the ulcer feature description. It is typically calculated using the standard deviation or coefficient of variation of the ulcer feature vector. The dispersion threshold defines an acceptable upper limit for the internal variation of ulcer features; values ​​below this threshold indicate good consistency and reliability of the extracted ulcer features. The system only triggers the feature representation creation process when the optimized feature dataset simultaneously meets both criteria: the inflammatory features possess sufficient discriminative power, and the ulcer features exhibit necessary stability. This feature representation typically exists as a structured data object containing key features and their metadata, which can be used to train classification models or directly assist pathologists in diagnostic analysis.

[0069] When the initially established feature representation is deemed to fail to meet expected quality requirements after evaluation (e.g., through cross-validation or performance on independent test sets), the system initiates a feedback adjustment loop. Quality requirements may involve the feature representation failing to meet preset standards in terms of accuracy, recall, or clinical interpretability in classification tasks. In this case, the root cause is considered to be an inadequate configuration of feature selection parameters in the adaptive feature optimization strategy. The adjustment process first targets these feature selection parameters. Adjustment strategies can be multifaceted, such as changing the method of feature importance scoring, switching from a variance-based method to a model-based method like random forest feature importance; or modifying the weight update rule, introducing a momentum term to smooth weight changes and prevent oscillations. The adjustment is often based on the specific pattern of feature representation failure. If inflammation discrimination is insufficient, the initial weights of inflammation-related features may be increased; if ulcer dispersion is too large, the consistency constraints on ulcer features may be strengthened.

[0070] Based on the adjusted feature selection parameters, the system needs to re-execute the joint computation process of the entire multi-scale feature analysis model and adaptive feature optimization strategy. This means starting from the raw or preprocessed pathological slide image data, performing feature extraction again through three feature analysis units: inflammation, ulcer, and normal tissue. Then, the new parameter matrix is ​​used for weighted fusion and optimization of the features, and the convergence condition is checked again to obtain a new, optimized feature dataset. Subsequently, the system uses this new data to attempt to build feature representations again and reapplies the first and second decision criteria for verification. This "evaluation-adjustment-recomputation" cycle continues until the generated feature representations are verified to meet the predetermined quality requirements. A maximum number of iterations may be set during the cycle to prevent infinite loops. The entire process embodies a self-optimization mechanism based on closed-loop feedback, gradually approaching the optimal feature representation scheme through iterative refinement, thereby ensuring the reliability and practicality of the final output results. This design allows the method to adapt to pathological slide data of different sources and qualities, exhibiting strong robustness and adaptability.

[0071] Example 4: This example details the construction of the inflammation feature analysis unit and the ulcer feature analysis unit. Its core lies in transforming the macroscopic pathological region into a computable, structured network of feature blocks. Using a hypothetical proctitis pathological slide sample as an example, the image clearly shows an inflammatory infiltration area and an ulcer area with a certain morphology. The construction of the inflammation feature analysis unit begins with the refined division of the inflammatory region. This process is not a simple uniform segmentation, but rather based on the biological nature of the distribution characteristics of inflammatory cells. In a given sample, a pathologist may observe that inflammatory cells are not uniformly diffusely distributed, but rather exhibit various patterns such as focal aggregation and perivascular cuff-like infiltration. Based on this, the image processing algorithm first identifies local high-density cell aggregation points within the inflammatory region and uses these as seed points or centers to divide the entire inflammatory region into multiple feature analysis blocks. One block may correspond to a "perivascular cuff" region centered on a small venule and densely surrounded by lymphocytes; another block may cover a "diffuse infiltration" region where neutrophils are diffusely distributed in the glandular stroma. Each block carries unique local texture and morphological information. The division can be based on the spatial variation rate of cell density and cell type ratio, achieved through clustering algorithms (such as k-means based on spatial coordinates and pixel intensity) or adaptive grid division based on prior knowledge. See Table 1 for parameter descriptions of some feature analysis blocks after dividing the inflammatory region of this sample.

[0072] Table 1: Analysis Blocks of Inflammatory Region Characteristics

[0073]

[0074] These segmented feature analysis blocks are connected by a feature association network, which aims to establish spatial and pathophysiological connections between blocks. The connections are not simply physical adjacencies, but rather based on the strength of the association according to pathological significance. For example, two blocks spatially separated by normal glands but with similar high-density lymphocyte infiltration may have a stronger association than two physically adjacent blocks with completely different cell types. The association network can be modeled as a graph structure, where nodes represent the feature analysis blocks, and edges represent the associations between blocks. The weight of each edge is determined by a comprehensive calculation based on factors such as spatial distance between blocks, similarity of cell composition, and correlation of texture features. For example, blocks B1 (high-density lymphocytes) and B4 (low-density lymphocytes) may be assigned a moderately strong association edge because they have the same cell type, while B1 and B2 (neutrophils) may have a lower association weight due to different infiltrating cell types and greater spatial distance.

[0075] While constructing the network, feature analysis priorities need to be set. These priorities determine the order in which these blocks are processed and the allocation of computational resources. The priority setting is directly related to the pathological activity of the blocks; generally, blocks with higher inflammatory cell density and stronger activity (e.g., neutrophil infiltration indicating acute activity) are assigned higher priorities. In this example, blocks B3 (mixed infiltration, high density) and B1 (lymphocytes, high density) may have the highest priority because they represent the most significant centers of inflammatory activity, and the system will prioritize extracting and analyzing features from these blocks. Block B4 (low density) has a relatively lower priority. Based on these feature analysis blocks, the connections between them formed through the feature association network, and the set analysis priorities, an inflammatory feature analysis unit is constructed. The workflow of this unit is as follows: starting with high-priority blocks, its local features are extracted, and using the association network information, features from adjacent or similar blocks are considered for supplementation or correction, ultimately outputting a structured description that reflects the overall and local characteristics of the inflammatory region.

[0076] The construction logic of the ulcer feature analysis unit is similar, but its benchmark is the morphological characteristics of the ulcer edge. The analysis of the ulcer region focuses on the geometric shape of its destruction and signs of repair. Using the morphological characteristics of the ulcer edge as the benchmark means that when dividing the blocks, special attention will be paid to the irregularity of the edge, the slope, and the boundary between the ulcer base and the normal mucosa. In the ulcer region of the example sample, the edge may present different shapes such as steep, punched, or gentle, sloping. Therefore, when dividing the feature calculation blocks, the entire ulcer region (including the edge and the base) will be divided. The edge area may be divided into multiple small blocks according to the curvature change, and the base may be divided according to the tissue type (such as necrotic tissue, granulation tissue). Each block will be used to calculate morphological parameters, such as calculating the rate of change of its tangential direction for the edge block and calculating the texture uniformity for the base block.

[0077] These feature calculation blocks are then topologically connected according to their spatial distribution. The topological connections focus on the relative positions and connectivity between blocks, rather than precise geometric distances. For example, a block located at the very edge of the ulcer (labeled U1) will be connected to its adjacent block (U2) facing the ulcer center. U2 will then be connected to a block (U3) closer to the bottom of the ulcer, forming a topological chain from the edge to the center. Simultaneously, all edge blocks (U1, U4, U5...) may also be interconnected, forming a topological loop describing the ulcer's perimeter. This topological structure clearly reflects the geometric layout of the ulcer region and is crucial for understanding the ulcer's three-dimensional morphology. Finally, based on these feature calculation blocks and their topological connections, an ulcer feature analysis unit is constructed. After calculating the local morphological features (such as area, perimeter, and shape factor) of each block, this unit integrates these local features using topological relationships. For example, by analyzing the shape factor sequence of each block on the edge topological ring, the overall irregularity of the ulcer can be determined; by analyzing the changes in tissue texture on the topological chain from the edge to the center, the healing stage of the ulcer can be assessed. The output of this unit is a set of ulcer morphological features containing spatial structural information.

[0078] Example 5: When constructing a normal tissue feature analysis unit, the fundamental starting point is to use the structural features of normal tissue as a clear direction for feature analysis. The structural features of normal tissue at the digestive tract mucosa level exhibit high regularity; for example, the tubular structures of intestinal glands (crypts) are neatly arranged, cell nuclei are located on the basal side, and goblet cells are evenly distributed. These inherent, healthy morphological patterns provide clear guidance for feature extraction. Based on this feature analysis direction, a logically coherent feature sampling path needs to be established. This path does not wander randomly across the image but follows the anatomical structure of normal tissue. For example, in a typical normal area of ​​colonic mucosa, the feature sampling path can be set along the long axis of a complete intestinal gland, from the bottom of the gland duct to its opening, with a series of sampling points set at fixed intervals. The direction of the path ensures that the sampling process can capture the complete morphological changes of the gland duct unit. Another path setting method may be perpendicular to the mucosal surface, extending from the muscularis mucosae to the luminal surface, using this sampling path to characterize the integrity of the tissue layers. Along this preset feature sampling path, the normal tissue feature analysis unit is systematically constructed. This unit applies a series of feature extraction operations at each sampling point. These operations may include calculating the texture indicators of the local image (such as entropy and contrast), measuring the consistency of the cell nucleus alignment direction, and quantifying the smoothness of the glandular lumen edge. All feature values ​​obtained along the path are recorded sequentially, ultimately forming a feature sequence that can characterize the spatial continuity and structural regularity of normal tissue.

[0079] The construction of an adaptive feature optimization strategy is a dynamic, data-driven process. Its first step is to calculate the importance scores of each region's features output by the multi-scale feature analysis model. The importance score aims to quantify the contribution of each feature to the final differentiation of different pathological states (such as inflammation activity and ulcer depth). The calculation process typically relies on a machine learning model embedded in the feature usage phase. For example, after obtaining a dataset of unoptimized features and their corresponding pathological grading labels, a random forest classifier can be trained. This model automatically evaluates the importance of each feature based on the information gain (or decrease in Gini impurity) brought about by feature splitting at decision tree nodes, outputting an importance score vector. Cell density features in inflammatory regions may receive high scores, while certain texture features may receive lower scores. This scoring process is adaptive because it depends on the statistical regularities inherent in the specific input sample set.

[0080] Based on the calculated importance score, the system generates a feature optimization weight matrix. This matrix is ​​a diagonal matrix or weight vector, where each element corresponds to the optimization weight of a specific feature. The weight assignment rules are directly linked to the importance score; features with high importance scores are assigned larger weights, meaning they will play a more dominant role in the subsequent optimization and fusion stages; conversely, features with low scores have smaller weights, and their influence is suppressed. Non-linear mapping functions may be introduced when generating the weight matrix. For example, a softmax function may be used to transform the importance score into a normalized weight distribution to ensure that the sum of all weights is 1, or a threshold may be set to directly set the weights of features below the threshold to zero to achieve feature selection.

[0081] By generating a feature optimization weight matrix, the system dynamically optimizes the features of each region. This dynamic optimization is reflected in the fact that the weight matrix is ​​not fixed but continuously adjusted according to different input samples or the model's self-update during training. During each feature fusion or model training iteration, the system uses the current weight matrix to weight the original features, highlighting important features and reducing the interference of redundant or noisy features. This complete closed loop—from calculating importance scores to generating the weight matrix and then applying the weights for feature optimization—constitutes the aforementioned adaptive feature optimization strategy. This strategy enables the entire feature processing system to intelligently focus on the feature information most relevant to the current diagnostic task, thereby improving the quality and discriminative ability of the final feature representation.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive feature processing method for pathological slide images of gastroenteritis, characterized in that, Includes the following steps: A multi-scale feature analysis model is constructed based on the histological features of pathological sections of gastroenteritis. The multi-scale feature analysis model is used to extract texture and morphological features of different tissue levels in the pathological sections. The histological features include the histological features of the inflamed area, the ulcerated area, and the normal tissue area. Based on the histological features of the inflamed area, the ulcerated area, and the normal tissue area, respectively, an inflammation feature analysis unit, an ulcer feature analysis unit, and a normal tissue feature analysis unit are constructed. An adaptive feature optimization strategy is constructed, which is used to dynamically adjust the feature selection parameters based on the feature data output by the multi-scale feature analysis model. Based on the selected pathological slide sample type, the multi-scale feature analysis model and the adaptive feature optimization strategy are jointly processed to obtain an optimized feature dataset, including: Based on the selected pathological slide sample type, a feature selection parameter matrix is ​​generated using the adaptive feature optimization strategy; Based on the feature selection parameter matrix, the feature data output by the inflammation feature analysis unit, the ulcer feature analysis unit and the normal tissue feature analysis unit are weighted and fused. When the feature fusion calculation result meets the convergence condition, the optimized feature dataset is output; Based on the optimized feature dataset, feature representations of pathological slide images of gastroenteritis are established.

2. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 1, characterized in that, When the histological features include those of inflamed areas, ulcerated areas, and normal tissue areas, the construction of a multi-scale feature analysis model based on the histological features of gastroenteritis pathological sections includes: The output data of the inflammation feature analysis unit and the input data of the ulcer feature analysis unit are subjected to feature alignment processing. The output data of the ulcer feature analysis unit and the input data of the normal tissue feature analysis unit are subjected to feature normalization processing.

3. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 2, characterized in that, Before performing joint calculations on the multi-scale feature analysis model and the adaptive feature optimization strategy based on the selected pathological slide sample type, the method further includes: The standard pathological slide image is input into the inflammation feature analysis unit to calculate the feature extraction accuracy. When the feature extraction accuracy reaches the first preset standard, the standard pathological slide image is input into the ulcer feature analysis unit to calculate the feature matching accuracy; When the feature matching accuracy reaches the second preset standard, the standard pathological slide image is input into the normal tissue feature analysis unit to calculate the feature stability index.

4. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 1, characterized in that, The step of establishing feature representations for pathological slide images of gastroenteritis based on the optimized feature dataset includes: When the optimized feature dataset satisfies the first and second judgment criteria, the feature representation is established. The first criterion is that the discriminative power of the inflammatory region features in the optimized feature dataset is greater than a preset discriminative power threshold; The second criterion is that the dispersion of the ulcer region features in the optimized feature dataset is less than a preset dispersion threshold.

5. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 1, characterized in that, The method further includes: When the feature representation fails to meet the expected quality requirements, the feature selection parameters in the adaptive feature optimization strategy are adjusted. Based on the adjusted feature selection parameters, the joint operation of the multi-scale feature analysis model and the adaptive feature optimization strategy is re-executed until a feature representation that meets the quality requirements is obtained.

6. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 2, characterized in that, The unit for constructing inflammatory feature analysis includes: Based on the distribution characteristics of inflammatory cells, the inflammatory area is divided into multiple feature analysis blocks; The feature analysis blocks are connected through a feature association network, and feature analysis priorities are set; The inflammation feature analysis unit is constructed based on the feature analysis blocks and feature association network.

7. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 2, characterized in that, The ulcer feature analysis unit includes: Based on the morphological characteristics of the ulcer edge, the ulcer region is divided into multiple feature calculation blocks; The feature calculation blocks are topologically connected according to their spatial distribution; Based on the feature calculation blocks and topological connection relationships, the ulcer feature analysis unit is constructed.

8. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 2, characterized in that, The construction of the normal tissue feature analysis unit includes: The structural features of normal tissue are used as the direction of feature analysis. A feature sampling path is set based on the feature analysis direction, and the normal tissue feature analysis unit is constructed along the feature sampling path.

9. The adaptive feature processing method for pathological slide images of gastroenteritis according to claim 1, characterized in that, The adaptive feature optimization strategy includes: Calculate the importance score of each region feature output by the multi-scale feature analysis model; A feature-optimized weight matrix is ​​generated based on the importance score; The adaptive feature optimization strategy is constructed by dynamically optimizing the features of each region using the feature optimization weight matrix.

Images