A hybrid intelligent-based intelligent diagnosis and treatment auxiliary method for blood diseases

Abstract

The application discloses a kind of blood disease intelligent diagnosis and treatment auxiliary methods based on hybrid intelligence, comprising the following steps: obtaining cell morphology image and pathological whole section image and pre-processing, obtain the image to be analyzed;Cell image is segmented, boundary correction and morphological feature is extracted;Cell is gathered into morphological cluster and abstract is generated;According to expert cluster level decision correction cell evidence score;Pathological image is cut, segmented, polished and constructed instance package;According to the long axis of bone marrow input ABMIL, fusion cell evidence generates classification, confidence, image evidence and suggestion.The application realizes blood disease intelligent diagnosis and treatment based on cell morphology and pathological whole section, and high-precision classification and visual evidence generation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to an intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence. Background Technology

[0002] The diagnosis of hematological diseases typically requires a comprehensive assessment combining peripheral blood smears, bone marrow smears, bone marrow cell morphology images, and whole-section pathological images from bone marrow biopsies. Existing intelligent assisted diagnostic methods often employ techniques such as image preprocessing, cell segmentation, morphological feature extraction, pathological region segmentation, and weakly supervised classification to identify cell nuclei, cytoplasm, granular structures, pathological region distribution, and abnormal infiltration. Based on the model's output, disease classification results or supplementary suggestions are generated.

[0003] Current methods in cell morphology analysis typically treat individual cells, relying primarily on cell categories, abnormality scores, or morphological indicators output by the model for judgment. They lack mechanisms for unified organization and verification of morphologically similar and diagnostically relevant cell populations. When the model misclassifies a particular cell type, encounters artifacts, or exhibits classification uncertainty, physicians often have to manually verify each cell individually. The verification results are difficult to apply effectively to similar cell populations and are hard to translate into corrective evidence for subsequent disease classification. Furthermore, existing technologies lack an effective hybrid intelligent synergy mechanism between expert experience, cell morphology evidence, and weakly supervised pathological findings. This results in insufficient fusion of cellular and pathological evidence, and the interpretability of classification confidence, image evidence, and examination recommendations still needs improvement.

[0004] Therefore, how to provide a hybrid intelligence-based intelligent diagnostic and treatment method for blood diseases is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence. This invention makes full use of cell morphology image analysis, pathological whole-slice image segmentation, and the ABMIL weakly supervised aggregation algorithm. It describes in detail the implementation process of cell feature extraction, pathological region instance package construction, and long axis serialization aggregation, and has the advantages of accurate disease classification, clear image basis, and interpretable diagnostic and treatment suggestions.

[0006] According to an embodiment of the present invention, a method for intelligent diagnosis and treatment of blood diseases based on hybrid intelligence includes the following steps:

[0007] Acquire cell morphology images and pathological whole slide images of the subject, perform preprocessing, and generate cell images and pathological whole slide images to be analyzed;

[0008] Cell morphology analysis is performed on the cell image to be analyzed, staining decomposition is performed, cell center point, ray direction and initial endpoint are obtained based on cell instance segmentation model, cell boundary is corrected, and a set of cell morphology features is generated.

[0009] Based on the set of cell morphological features, abnormal scores, predicted categories, and their contribution to disease classification, cells with similar morphology and related diagnostic significance are aggregated into cell morphology clusters. These cell morphology clusters are then used as human-machine collaborative review objects to generate cluster-level summaries.

[0010] Receive cluster-level decision information from experts on cell morphology clusters, convert the cluster-level decision information into hybrid intelligent correction instructions, perform overall correction on cell morphology clusters and related cell groups, and generate a cluster-level corrected cell morphology evidence score.

[0011] Image block segmentation, pathological region segmentation, and boundary refinement are performed on the whole-slice images of the pathological tissue to be analyzed.

[0012] Based on the refined pathological region segmentation map and the spatial coordinates of image blocks, the pathological region identifier is determined, a set of pathological region instance packages is constructed, image features are extracted, and associated spatial coordinates are determined.

[0013] The pathological region instance package is serialized according to the long axis of the bone marrow biopsy tissue and then input into the ABMIL model. The ABMIL output results are fused with the cluster-level corrected cell morphology evidence score to generate blood disease classification results, classification confidence, image evidence and examination suggestions.

[0014] Optionally, the preprocessing specifically includes:

[0015] Cell morphology images are one or more of the following: peripheral blood smear images, bone marrow smear images, and high-power field images of bone marrow cells; pathological whole slide images are one or more of the following: bone marrow biopsy HE-stained whole slide images, immunohistochemical stained whole slide images, reticular fiber stained whole slide images, and lymph node pathological whole slide images.

[0016] The cell morphology image is converted in format, pixel size is unified and staining is normalized. The sharpness value, background pixel ratio and cell connected region area of ​​each field of view are calculated. Fields of view with sharpness value less than the preset sharpness threshold, fields of view with background pixel ratio greater than the preset background threshold, and cell connected region area greater than the preset overlap threshold are deleted to generate cell images to be analyzed.

[0017] The image pyramid hierarchy of the whole pathological slide image is read, the tissue mask is extracted, the blank background outside the tissue mask is removed, the pixel blocks corresponding to slide folds, bubbles, staining deposits and out-of-focus areas are identified and removed from the image pyramid, the original spatial coordinates of the unremoved pixel blocks are retained, and the whole pathological slide image to be analyzed is generated.

[0018] Optionally, performing cell morphology analysis on the cell image to be analyzed specifically includes:

[0019] Stain decomposition is performed on the cell images to be analyzed to generate nuclear staining channels, cytoplasmic staining channels, and granular structure channels;

[0020] Input the nuclear staining channel, cytoplasmic staining channel, and granular structure channel into the StarDist model, and output the cell instance center point, ray direction, and initial ray endpoint;

[0021] A one-dimensional color scale profile is extracted along the ray direction from the center point of the cell instance to the outer side of the initial ray endpoint;

[0022] Input the one-dimensional color scale profile into the endpoint discriminator to generate endpoint category discrimination results;

[0023] When the endpoint category is determined to be a non-real cell boundary, search for real cell boundary candidate points along the same ray direction and replace the initial ray endpoint with the real cell boundary candidate points.

[0024] Cell contour data and cell nucleus contour data are generated based on the replaced ray endpoints. Cell morphology indicators are calculated based on the cell contour data and cell nucleus contour data to generate a set of cell morphology features.

[0025] Optionally, the aggregation of morphologically similar and diagnostically relevant cells into cell morphology clusters specifically includes:

[0026] Read the cell morphology feature set, anomaly score, predicted category and contribution to disease classification for each cell, and input the cell morphology feature set and anomaly score into the representation encoder to generate cell morphology representation vector;

[0027] All cell morphology representation vectors are input into the clustering model to form initial cell clusters according to the proximity distribution of cells in the feature space;

[0028] Based on the concentration of predicted categories and the range of variation in disease classification contribution within the initial cell cluster, cell regions with inconsistent diagnostic significance within the same initial cell cluster are identified, and these cell regions are reassigned to adjacent initial cell clusters or independently form new initial cell clusters.

[0029] The cluster center position and intra-cluster morphological dispersion are calculated for the adjusted initial cell clusters. Initial cell clusters with intra-cluster morphological dispersion not greater than a preset dispersion threshold are identified as cell morphology clusters.

[0030] Within a cell morphology cluster, cells near the cluster center with anomaly scores higher than the average anomaly score within the cluster are selected as representative cells. Cluster-level summaries are generated based on representative cells, intra-cluster morphological dispersion, dominant prediction category, and disease support strength.

[0031] Optionally, the overall correction of cell morphology clusters and associated cell populations specifically includes:

[0032] Cluster-level adjudication information refers to the category confirmation, category modification, artifact exclusion, key diagnostic confirmation, or re-clustering instructions made by experts for the entire cell morphology cluster;

[0033] Cluster-level decision information is encoded into hybrid intelligent correction instructions, which are parameterized instructions consisting of expert decision direction, target cell type, correction range, and correction intensity.

[0034] Based on the cell morphology representation vector, the cells that are closest to the morphological clusters of the adjudicated cells and are related to the predicted category are retrieved from the cells that have not been adjudicated by experts, and the associated cell population is generated.

[0035] According to the hybrid intelligent correction instructions, the abnormal scores, predicted category probabilities, disease classification contributions, and artifact exclusion status of the adjudicated cell morphology clusters and related cell populations are updated synchronously.

[0036] Based on the updated anomaly scores, predicted category probabilities, and disease classification contributions, the support value of each cell morphology cluster for each blood disease category is recalculated, and the support values ​​are normalized to generate a cluster-level corrected cell morphology evidence score.

[0037] Optionally, the boundary refinement specifically includes:

[0038] Extract adjacent pixel pairs of different categories from the initial pathological region segmentation map to form the boundary lines of adjacent pathological regions;

[0039] The image content of the whole slice image to be analyzed and the category label in the initial pathological region segmentation map are extracted along the normal direction of the boundary line of the adjacent pathological region to generate the boundary ring;

[0040] Set boundary sampling points according to the direction of the boundary lines of adjacent pathological regions, and sample the boundary ring along the normal direction of the boundary sampling points to generate a rectangular image and a rectangular category map.

[0041] The first boundary candidate point is determined based on the staining change value and cell density change value in the rectangular image, and the second boundary candidate point is determined based on the continuous interval of the category in the rectangular category map.

[0042] The first boundary candidate point is matched with the second boundary candidate point. If the matching does not meet the preset conditions, the category continuous interval is searched again along the corresponding normal sampling line to determine the corrected boundary point.

[0043] The corrected boundary points are connected to form a corrected boundary line. The pixel category within the boundary ring is updated based on the corrected boundary line to generate a refined pathological region segmentation map.

[0044] Optionally, the construction of the pathological region instance package specifically includes:

[0045] Read the pathological region identifier, pathological region category ratio, and spatial coordinates of each image block, and establish spatial adjacency relationships for image blocks with the same pathological region identifier and a spatial distance not greater than a preset adjacency distance;

[0046] Based on spatial adjacency, image blocks under the same pathological region are divided into several connected image block groups;

[0047] Calculate the average proportion of pathological region categories and spatial coverage of image blocks within each connected image block group. Determine the connected image block group with an average proportion of pathological region categories not less than a preset purity threshold as a region instance pack. Determine the connected image block group with combined coding and located between two region instance packs as a boundary instance pack.

[0048] Image features are extracted from each image block within the region instance package and the boundary instance package. The image features, spatial coordinates, pathological region identifiers, and pathological region category proportions are bound together according to the image block number to generate a set of pathological region instance packages.

[0049] Optionally, the fusion of the ABMIL output and the cluster-level corrected cell morphology evidence score specifically includes:

[0050] Extract the bone marrow biopsy tissue region from the whole slice image to be analyzed in pathology, and determine the long axis of the bone marrow biopsy tissue based on the principal direction of the bone marrow biopsy tissue region;

[0051] Read the image features and spatial coordinates of image blocks within the pathological region instance package set, project the center point of the image block onto the long axis of the bone marrow biopsy tissue to obtain the long axis coordinates, and generate a long axis image block sequence according to the long axis coordinates within the corresponding region instance package or boundary instance package.

[0052] The sequence of long-axis image patches and their long-axis coordinates are input into the ABMIL model for weakly supervised aggregation, generating package-level classification results and image patch contribution of instance packages;

[0053] The package-level classification results of each instance package are fused with the cluster-level corrected cell morphology evidence scores to generate blood disease classification results and classification confidence.

[0054] The basis image block is determined based on the contribution of the image block, and the image basis and examination suggestions are generated by combining the spatial coordinates of the basis image block, the blood disease classification results and classification confidence. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of an intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence proposed in this invention;

[0057] Figure 2 This is a schematic diagram of the cell morphology analysis process for a hybrid intelligence-based intelligent diagnostic and treatment auxiliary method for blood diseases proposed in this invention.

[0058] Figure 3 This is a schematic diagram of the instance package construction and long-axis ABMIL aggregation process for an intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] Hybrid intelligence refers to a collaborative approach that combines human intelligence with artificial intelligence. It doesn't simply rely on machines to automatically complete tasks, but rather fully leverages the advantages of AI in data processing, pattern recognition, and rapid computation, while retaining the dominant role of humans in experiential judgment, value selection, innovative thinking, and complex decision-making. Through human-machine collaboration, hybrid intelligence can improve the efficiency, accuracy, and adaptability of systems, enabling them to possess stronger analytical and decision-making capabilities when facing complex, uncertain, or scenarios requiring comprehensive judgment.

[0061] The hybrid intelligence in this invention refers to a collaborative diagnostic approach that combines the image recognition, feature extraction, cluster analysis, and classification prediction capabilities of an artificial intelligence model with the professional review and adjudication results of medical experts. In this invention, the system first automatically analyzes cell morphology images and whole-slice pathological images, generating cell morphology clusters, anomaly scores, predicted categories, disease classification contributions, and pathological image analysis results. Then, it receives expert decisions regarding cell morphology clusters, such as category confirmation, category modification, artifact exclusion, or key diagnostic confirmation, and converts these decisions into hybrid intelligence correction instructions to correct the evidence scores of relevant cell populations. Thus, expert experience can be transformed into quantitative evidence that can participate in model fusion calculations, jointly generating hematological disease classification results, confidence levels, image evidence, and examination suggestions with the pathological classification results output by the ABMIL model, thereby improving the accuracy and interpretability of diagnostic and treatment assistance results.

[0062] refer to Figures 1-3 A method for intelligent diagnosis and treatment assistance of hematological diseases based on hybrid intelligence includes the following steps:

[0063] Acquire cell morphology images and pathological whole slide images of the subjects, perform preprocessing, and generate cell images and pathological whole slide images to be analyzed;

[0064] Cell morphology analysis is performed on the cell image to be analyzed, staining decomposition is performed, cell center point, ray direction and initial endpoint are obtained based on cell instance segmentation model, cell boundary is corrected, and a set of cell morphology features is generated.

[0065] Based on the set of cell morphological features, abnormal scores, predicted categories, and their contribution to disease classification, cells with similar morphology and related diagnostic significance are aggregated into cell morphology clusters. These cell morphology clusters are then used as human-machine collaborative review objects to generate cluster-level summaries.

[0066] Receive cluster-level decision information from experts on cell morphology clusters, convert the cluster-level decision information into hybrid intelligent correction instructions, perform overall correction on cell morphology clusters and related cell groups, and generate a cluster-level corrected cell morphology evidence score.

[0067] Image block segmentation, pathological region segmentation, and boundary refinement are performed on the whole-slice images to be analyzed in pathology.

[0068] Based on the refined pathological region segmentation map and the spatial coordinates of image blocks, the pathological region identifier is determined, a set of pathological region instance packages is constructed, and image features and associated spatial coordinates are extracted.

[0069] The pathological region instance package is serialized according to the long axis of the bone marrow biopsy tissue and then input into the ABMIL model. The ABMIL output results are fused with the cluster-level corrected cell morphology evidence score to generate blood disease classification results, classification confidence, image evidence and examination suggestions.

[0070] In this embodiment, the preprocessing specifically includes:

[0071] Cell morphology images are one or more of the following: peripheral blood smear images, bone marrow smear images, and high-power field images of bone marrow cells; pathological whole slide images are one or more of the following: bone marrow biopsy HE-stained whole slide images, immunohistochemical stained whole slide images, reticular fiber stained whole slide images, and lymph node pathological whole slide images.

[0072] The cell morphology images are subjected to format conversion, pixel size unification, and staining normalization. The sharpness value, background pixel ratio, and cell connected region area are calculated for each field of view. Fields of view with sharpness values ​​less than a preset sharpness threshold, background pixel ratios greater than a preset background threshold, and cell connected region areas greater than a preset overlap threshold are deleted to generate cell images for analysis. The sharpness value is obtained by calculating the Laplacian response variance after performing grayscale processing on each field of view image. The cell connected region area is obtained by performing connected component analysis on the binary image of the cell region, and the pixel area of ​​each connected component is counted. The largest connected component area or the average connected component area is used as the cell connected region area of ​​the corresponding field of view.

[0073] The process involves reading the image pyramid hierarchy of the whole pathological slide image, extracting the tissue mask, removing the blank background outside the tissue mask, identifying pixel blocks corresponding to slide folds, bubbles, staining deposits, and out-of-focus areas, and removing them from the image pyramid while retaining the original spatial coordinates of the remaining pixel blocks to generate the whole pathological slide image to be analyzed. The tissue mask is generated at a low magnification image pyramid hierarchy. The whole pathological slide image is converted to the HSV color space, and tissue areas are distinguished from blank background areas based on saturation and brightness channels. Then, hole filling and small area removal are performed on the obtained tissue areas. Slide fold areas are obtained by detecting local brightness abrupt changes, texture direction abrupt changes, and tissue edge overlap features. Bubble areas are obtained by detecting near-circular high-brightness boundaries and internal low-texture areas. Staining deposit areas are obtained by detecting local high-saturation dark patches. Out-of-focus areas are obtained by calculating the Laplacian response variance of the image blocks.

[0074] In this embodiment, performing cell morphology analysis on the cell image to be analyzed specifically includes:

[0075] The cell image to be analyzed is subjected to staining decomposition to generate nuclear staining channels, cytoplasmic staining channels and granular structure channels. The staining decomposition uses the color deconvolution method to convert the cell image to be analyzed from the RGB color space to the optical density space, and separates the nuclear staining channels and cytoplasmic staining channels according to the preset staining matrix. The granular structure channels are detected by performing LoG spot detection on the cytoplasmic staining channels.

[0076] The nuclear staining channel, cytoplasmic staining channel, and granular structure channel are input into the StarDist model, and the cell instance center point, ray direction, and initial ray endpoint are output. The one-dimensional color scale profile starts from the cell instance center point and samples pixel by pixel along the ray direction to 30 pixels outside the initial ray endpoint. The sampling content includes the gray value of the nuclear staining channel, the gray value of the cytoplasmic staining channel, the response value of the granular structure channel, and the gray value difference between adjacent sampling points. When the sampling length is inconsistent, it is unified by linear interpolation.

[0077] A one-dimensional color scale profile is extracted along the ray direction from the center point of the cell instance to the outer side of the initial ray endpoint; the endpoint discriminator is a multilayer perceptron classifier; the endpoint category discrimination results include the real cell boundary, nucleolar boundary, granule boundary, staining precipitation boundary and background transition boundary; the input of the endpoint discriminator is the one-dimensional color scale profile, and the output is the endpoint category probability vector. The category corresponding to the highest probability in the endpoint category probability vector is taken as the endpoint category discrimination result.

[0078] Input the one-dimensional color scale profile into the endpoint discriminator to generate endpoint category discrimination results;

[0079] When the endpoint category is determined to be a non-real cell boundary, search for real cell boundary candidate points along the same ray direction and replace the initial ray endpoint with the real cell boundary candidate points.

[0080] Cell contour data and cell nucleus contour data are generated based on the replaced ray endpoints. Cell morphology indicators are calculated based on the cell contour data and cell nucleus contour data to generate a set of cell morphology features.

[0081] In this embodiment, grouping cells with similar morphology and related diagnostic significance into cell morphology clusters specifically includes:

[0082] The algorithm reads the cell morphology feature set, anomaly score, predicted category, and contribution to disease classification for each cell. The cell morphology feature set and anomaly score are then input into a representation encoder to generate a cell morphology representation vector. Each numerical feature in the cell morphology feature set is standardized before being input into the representation encoder by subtracting the mean of the training samples and dividing by the standard deviation of the training samples. The anomaly score ranges from 0 to 1, with higher values ​​indicating a greater deviation from normal cell morphology. The contribution to disease classification ranges from 0 to 1 and is obtained by attention weight normalization. The representation encoder is a multilayer perceptron encoder, and its inputs are the standardized cell morphology feature set, anomaly score, and cell morphology representation vector.

[0083] All cell morphology representation vectors are input into a clustering model to form initial cell clusters based on the proximity distribution of cells in the feature space. The clustering model adopts a density-based clustering model, using the Euclidean distance or cosine distance between two cell morphology representation vectors as a measure of proximity. When using Euclidean distance, the neighborhood radius is set to 0.6, and the minimum number of cluster samples is set to 10. Cells that do not meet the clustering conditions are temporarily designated as outliers. Outliers do not directly form cell morphology clusters and are redistributed or retained as outliers to be reviewed during subsequent cluster boundary adjustments based on their distance from adjacent initial cell clusters.

[0084] Based on the concentration of predicted categories within initial cell clusters and the range of variation in disease classification contribution, cell regions with inconsistent diagnostic significance within the same initial cell cluster are identified. These cell regions are then reassigned to adjacent initial cell clusters or form new initial cell clusters independently. The concentration of predicted categories is defined as the proportion of the most numerous predicted categories within an initial cell cluster. When this proportion is less than 0.65, the initial cell cluster is considered to have mixed categories. The range of variation in disease classification contribution is defined as the difference between the maximum and minimum disease classification contribution within the cluster. When this difference is greater than 0.4, the initial cell cluster is considered to have cell regions with excessively large differences in contribution. For cell regions with mixed categories or excessively large differences in contribution, the system calculates the distance between the cell region and the center of adjacent initial cell clusters. When the minimum distance is not greater than 1.2 times the average intra-cluster distance of the corresponding adjacent initial cell clusters, the cell region is reassigned to the corresponding adjacent initial cell cluster. When the reassignment condition is not met, the cell region forms a new initial cell cluster independently.

[0085] The cluster center position and intra-cluster morphological dispersion are calculated for the adjusted initial cell clusters. Initial cell clusters with intra-cluster morphological dispersion not greater than a preset dispersion threshold are identified as cell morphology clusters. The cluster center position is the mean vector of all cell morphological representation vectors within the adjusted initial cell cluster. The intra-cluster morphological dispersion is the average distance between each cell morphological representation vector within the cluster and the cluster center position, and this average value is normalized. The preset dispersion threshold is set to 0.35. When the intra-cluster morphological dispersion is not greater than 0.35, it indicates that the cell morphology consistency within the initial cell cluster meets the requirements, and the initial cell cluster is identified as a cell morphology cluster. When the intra-cluster morphological dispersion is greater than 0.35, the initial cell cluster is re-clustered or labeled as a mixed cluster.

[0086] Within a cell morphology cluster, cells near the cluster center with anomaly scores higher than the cluster's average anomaly score are selected as representative cells. Cluster-level summaries are generated based on representative cells, intra-cluster morphological dispersion, dominant predicted category, and disease support strength. Cells near the cluster center are those within the top 30% of the cluster in terms of distance from the cluster center. Cells with anomaly scores higher than the cluster's average anomaly score are those with anomaly scores greater than the average anomaly score of all cells within the cell morphology cluster. When the number of cells meeting both of the above conditions is greater than 5, the top 5 cells are selected as representative cells in descending order of anomaly score. The dominant predicted category is the predicted category with the highest percentage within the cell morphology cluster. Disease support strength is obtained by summing the disease classification contributions of each cell within the cell morphology cluster and dividing by the number of cells within the cluster.

[0087] In this embodiment, the overall correction of cell morphology clusters and related cell groups specifically includes:

[0088] Cluster-level adjudication information refers to the expert's instructions on category confirmation, category modification, artifact exclusion, diagnostic key confirmation, or re-clustering for the entire cell morphology cluster. Category confirmation means that the expert confirms that the dominant predicted category of the cell morphology cluster is consistent with the actual morphological category. Category modification means that the expert changes the dominant predicted category of the cell morphology cluster to another preset cell category. Artifact exclusion means that the expert confirms that the cell morphology cluster is mainly composed of stained deposits, broken cells, overlapping cells, out-of-focus cells, or non-cellular structures. Diagnostic key confirmation means that the expert confirms that the cells in the cell morphology cluster have a positive supporting role in the classification of blood diseases. Re-clustering means that the expert confirms that there are two or more cell subpopulations with significant morphological differences or diagnostic significance differences within the cell morphology cluster, requiring re-execution of intra-cluster clustering.

[0089] Cluster-level adjudication information is encoded into hybrid intelligent correction instructions, which are parameterized instructions consisting of expert adjudication direction, target cell category, correction range, and correction intensity. The expert adjudication direction indicates the direction of the expert adjudication's effect on the model's original judgment: category confirmation and key diagnostic confirmation are enhancement directions, category modification is a transfer direction, artifact removal is a suppression direction, and re-clustering is a reconstruction direction. The target cell category is the cell category confirmed or modified by the expert. The correction range includes the adjudicated cell morphology cluster, its adjacent cell morphology clusters, and associated cell populations retrieved based on cell morphology representation vectors. The correction intensity ranges from 0 to 1, with a correction intensity of 0.4 for category confirmation, 0.6 for category modification, 0.9 for artifact removal, 0.75 for key diagnostic confirmation, and 1 for re-clustering. The correction intensity represents the weighted correction magnitude of the expert adjudication result on the model's original output. Let the original model value be X, the target value corresponding to the expert adjudication be T, and the correction intensity be α, then the corrected value X' can be expressed as: X' = X + α × (TX). Where X can be the predicted category probability, anomaly score, disease classification contribution, or disease support value, and T is the target value determined based on the expert decision direction. The correction strength for re-clustering is set to 1, which is used to ensure that the evidence from the original cluster is completely replaced by the evidence from the sub-clusters generated after re-clustering.

[0090] The reason for adopting cluster-level adjudication is that blood cell morphology images often contain a large number of cells with similar morphology and the same or similar diagnostic significance. If each cell is reviewed individually, it is not only inefficient, but also difficult for experts to effectively transmit their judgments on a certain type of cell to the same cell population. By aggregating cells with similar morphology and related diagnostic significance into cell morphology clusters, experts can confirm, modify, eliminate artifacts, or confirm key diagnostic points for the entire cluster. This can transform expert experience into unified correction instructions, and make overall corrections to the abnormal scores, category probabilities, and disease contributions of the cluster and related cell populations, thereby forming more stable cell morphology evidence.

[0091] Based on cell morphology representation vectors, cells that are closest to and predicted to be related to the morphological clusters of adjudicated cells are retrieved from the unadjudicated cells, generating associated cell groups. The cluster center of the adjudicated cell morphology cluster is the mean vector of all cell morphology representation vectors within that cluster. The distance between unadjudicated cells and the adjudicated cell morphology cluster is the Euclidean distance between the cell morphology representation vector and the cluster center. When this distance is not greater than a preset association distance threshold, and the predicted class of the cell is the same as the target cell class of the adjudicated cell morphology cluster, or belongs to an adjacent class in a preset cell class association table, the cell is included in the associated cell group. The preset association distance threshold can be set to 1.5 times the average intra-cluster distance of the adjudicated cell morphology clusters, and the association strength of each cell in the associated cell group is obtained by normalizing the inverse of its distance to the cluster center.

[0092] Following the hybrid intelligent correction instructions, the abnormal scores, predicted category probabilities, disease classification contributions, and artifact exclusion status of the adjudicated cell morphology clusters and associated cell populations are updated synchronously. When the expert judgment direction is enhancement, the abnormal scores and disease classification contributions of cells in the adjudicated cell morphology clusters and associated cell populations are increased according to the correction intensity, and the predicted category probability corresponding to the target cell category is increased. When the expert judgment direction is transfer, the original predicted category probability is reduced, and the reduced portion is transferred to the target cell category probability according to the correction intensity. The disease classification contribution is recalculated based on the target cell category. When the expert judgment direction is suppression, the artifact exclusion status of cells within the adjudicated cell morphology clusters is set to excluded, and their abnormal scores and disease classification contributions are reduced to 0. The associated cell populations are reduced synchronously according to the association intensity. When the expert judgment direction is reconstruction, the cells within the adjudicated cell morphology clusters are re-input into the clustering model to generate new sub-clusters before subsequent correction.

[0093] Based on the updated anomaly scores, predicted category probabilities, and disease classification contributions, the support value of each cell morphology cluster for each blood disease category is recalculated. The support values ​​are then normalized to generate a cluster-corrected cell morphology evidence score. The support value for each cell morphology cluster for a blood disease category is calculated from the cells within that cluster that are not excluded by artifacts. For the same blood disease category, the support value for that cell morphology cluster for that blood disease category is obtained by multiplying the cell's updated anomaly score, the corresponding cell category probability, the disease classification contribution, and the preset association weight between the cell category and the blood disease category, and then summing the results. The preset association weight ranges from 0 to 1, representing the degree of support different cell categories for different blood disease categories. The support values ​​for each blood disease category are then Softmax normalized to obtain the cluster-corrected cell morphology evidence score.

[0094] In this embodiment, the pathological region segmentation specifically includes:

[0095] The image pyramid hierarchy of the whole pathological slide image to be analyzed is read, and the target magnification level is selected as the slicing level. The image pyramid hierarchy includes several magnification levels generated from the whole pathological slide image according to different downsampling factors. The target magnification level is determined based on the scanning resolution and the accuracy of pathological region recognition. When there is no image pyramid hierarchy corresponding to the target magnification, the image pyramid hierarchy closest to the target magnification is selected, and the image pyramid hierarchy is resampled to the target magnification.

[0096] At the slicing level, sliding window segmentation is performed on the whole slide image to be analyzed in the pathology section according to the preset image block size and preset step size to generate multiple candidate image blocks;

[0097] The tissue region proportion of each candidate image block is calculated based on the tissue mask of the whole slice image to be analyzed in pathology. Candidate image blocks with a tissue region proportion not less than a preset tissue threshold are identified as image blocks. The tissue region proportion is the ratio of the number of pixels belonging to the tissue mask in the candidate image block to the total number of pixels in the candidate image block. The preset tissue threshold is 0.5. When the tissue region proportion of a candidate image block is less than the preset tissue threshold, the candidate image block is marked as a blank candidate image block and is not input into the DeepLabv3+ model.

[0098] Record the top-left x and y coordinates of each image patch in the whole slice image to be analyzed in the pathology; the width and height of the image patch; and the image pyramid level.

[0099] Training image patches with pathological region annotations are input into the DeepLabv3+ network to perform parameter training, resulting in the DeepLabv3+ model. The pathological region annotations are annotation maps corresponding to the pixel positions of the training image patches, and each pixel in the annotation map has a pathological region category label. The pathological region category labels include at least two categories from hematopoietic region, fat region, fibrotic region, necrotic region, abnormal infiltration region, trabecular bone region, immunohistochemical positive region, and background region. During training, the training image patches are used as input, and the corresponding annotation maps are used as supervision labels to perform iterative training on the DeepLabv3+ network.

[0100] After performing size unification and pixel value normalization on the image patches, they are input into the DeepLabv3+ model to generate a region category probability map for each image patch.

[0101] The region category probability map is mapped back to the pathological whole slice image to be analyzed according to the spatial coordinates of each image patch, generating an initial pathological region segmentation map.

[0102] In this embodiment, the boundary refinement specifically includes:

[0103] Extract adjacent pixel pairs of different categories from the initial pathological region segmentation map, and connect the midpoints of the pixel pairs to form the boundary lines of adjacent pathological regions.

[0104] The boundary ring is generated by expanding a preset width to both sides along the normal direction of the boundary line of the adjacent pathological region, extracting the corresponding image content from the whole slice image of the pathological region to be analyzed and the corresponding category label from the initial pathological region segmentation image; the preset width is 32 pixels; the tangent direction is calculated for each boundary point on the boundary line of the adjacent pathological region, and then the normal direction is calculated based on the tangent direction; the boundary ring is generated by expanding a preset width to both sides along the normal direction to obtain a strip-shaped region around the boundary line of the adjacent pathological region.

[0105] Boundary sampling points are set sequentially according to the direction of the boundary lines of adjacent pathological regions. The boundary ring is sampled along the normal direction of each boundary sampling point, and then expanded to generate a rectangular image and a rectangular category map. The boundary sampling points are set in order of the arc length of the boundary lines of adjacent pathological regions. Pixel by pixel is sampled within the boundary ring along the normal direction of each boundary sampling point. The sampled image pixel values ​​are arranged in the order of the boundary sampling points to form a rectangular image, and the sampled category labels are arranged in the same order to form a rectangular category map.

[0106] The staining change value and cell density change value are calculated on each normal sampling line of the rectangular image to obtain the first boundary candidate point; the staining change value is obtained by calculating the color difference between adjacent sampling points on the normal sampling line; the cell density change value is obtained by statistically analyzing the change in the number of cell nucleus pixels per unit length on the normal sampling line; the position where the weighted sum of the staining change value and the cell density change value reaches a local extremum is determined as the first boundary candidate point;

[0107] Based on the rectangular category map, the continuous category intervals on each normal sampling line are counted. The continuous category intervals with a length less than a preset length threshold are deleted to obtain the second boundary candidate points. The continuous category interval is the sampling point segment with the same category label on the same normal sampling line. After deleting the continuous category intervals with a length less than the preset length threshold, the intersection position between the remaining continuous category intervals is determined as the second boundary candidate point.

[0108] The first boundary candidate point and the second boundary candidate point are matched in position. If the matching distance is not greater than the preset matching threshold, the corrected boundary point is determined. If the matching distance is greater than the preset matching threshold, the category continuous interval is searched again on the corresponding normal sampling line to determine the corrected boundary point. If the matching distance is not greater than the preset matching threshold, the midpoint of the line connecting the first boundary candidate point and the second boundary candidate point is taken as the corrected boundary point.

[0109] Connect all corrected boundary points according to the direction of adjacent pathological region boundary lines to form corrected boundary lines. Update the pixel categories within the boundary ring according to the corrected boundary lines to generate a refined pathological region segmentation map. Connect all corrected boundary points in the order of boundary sampling points to form corrected boundary lines. Perform smoothing processing on the corrected boundary lines using spline interpolation. Divide the boundary ring into two regions according to the corrected boundary lines. Reassign pixel values ​​within the boundary ring according to the original category labels on both sides of the corrected boundary lines.

[0110] In this embodiment, determining the pathological region identifier based on the refined pathological region segmentation map and the spatial coordinates of the image blocks specifically includes:

[0111] Based on the spatial coordinates of each image patch, locate the pixel region corresponding to the image patch in the refined pathological region segmentation map;

[0112] The number of pixels corresponding to each pathological region category within a pixel region is counted, and the ratio of the number of pixels corresponding to each pathological region category to the total number of pixels in the pixel region is taken as the proportion of the corresponding pathological region category.

[0113] The pathological regions were sorted from largest to smallest according to their proportion, and the pathological region category with the highest proportion was determined as the primary pathological region category.

[0114] When the proportion of the pathological region category corresponding to the main pathological region category is not less than the preset main category threshold, the main pathological region category is encoded as the pathological region identifier of the image block; the preset main category threshold is set to 0.7.

[0115] When the proportion of the pathological region category corresponding to the main pathological region category is less than the preset main category threshold, the top two pathological region categories are combined and encoded into the pathological region identifier of the image block according to the proportion order. The combined code is formed by connecting the first and second pathological region category codes in a fixed order, with a separator in the middle. The combined code is used to indicate that the image block crosses the boundary of two pathological regions.

[0116] Establish a connection between pathological region identifiers, pathological region category proportions, and spatial coordinates and corresponding image blocks.

[0117] In this embodiment, the construction of the pathological region instance package set specifically includes:

[0118] Read the pathological region identifier, pathological region category ratio, and spatial coordinates of each image block, and establish spatial adjacency relationships for image blocks with the same pathological region identifier and a spatial distance not greater than a preset adjacency distance;

[0119] Based on spatial adjacency, image blocks under the same pathological region are divided into several connected image block groups;

[0120] Calculate the average proportion of pathological region categories and spatial coverage of image blocks within each connected image block group. Connected image block groups with an average proportion of pathological region categories not less than a preset purity threshold are defined as region instance packages. Connected image block groups with combined codes and located between two region instance packages are defined as boundary instance packages. Spatial coverage is the area covered by the bounding rectangle of all image blocks within a connected image block group in the pathological whole-slice image to be analyzed. The preset purity threshold is 0.8. When the spatial coverage of a connected image block group with combined codes touches the spatial coverage of both region instance packages, the connected image block group with combined codes is determined to be located between the two region instance packages.

[0121] Image features are extracted from each image block within the region instance package and the boundary instance package. The image features, spatial coordinates, pathological region identifiers, and pathological region category proportions are bound together according to the image block number to generate a set of pathological region instance packages.

[0122] In this embodiment, the fusion of the ABMIL output result and the cluster-level corrected cell morphology evidence score specifically includes:

[0123] The bone marrow biopsy tissue region is extracted from the whole-slice image of the pathological examination. The principal direction of the bone marrow biopsy tissue region is calculated, and the straight line passing through the center of the bone marrow biopsy tissue region and parallel to the principal direction is determined as the long axis of the bone marrow biopsy tissue. The bone marrow biopsy tissue region is determined by the tissue mask in the whole-slice image of the pathological examination. The center of the bone marrow biopsy tissue region is the centroid of all pixel coordinates within the tissue mask. The principal direction is obtained by performing principal component analysis on all pixel coordinates within the tissue mask, and the direction of the first principal component is taken as the principal direction.

[0124] Read the image features and spatial coordinates of image blocks within the pathological region instance set, project the center point of each image block onto the long axis of the bone marrow biopsy tissue, and obtain the long axis coordinates of each image block;

[0125] Within each region instance package and each boundary instance package, image blocks are sorted according to their major axis coordinates to generate a sequence of major axis image blocks for the corresponding instance package. When two image blocks have the same major axis coordinates, they are sorted in ascending order according to the vertical distance from the center point of the image block to the major axis of the bone marrow biopsy tissue. When the number of image blocks in a region instance package or boundary instance package is less than the preset minimum number of instances, the region instance package or boundary instance package is merged with the instance package that is spatially closest and has the same pathological region identifier.

[0126] The ABMIL model inputs the long-axis image patch sequence, long-axis coordinates, and long-axis distances between adjacent image patches. The ABMIL model performs weakly supervised aggregation on each instance package, generating a package-level classification result and image patch contribution for each instance package. The long-axis distance between adjacent image patches is the difference in long-axis coordinates between two adjacent image patches in the long-axis image patch sequence. The ABMIL model receives the image features, long-axis coordinate encoding, and long-axis distance encoding for each image patch, and calculates the attention value corresponding to each image patch. The normalized attention value for each image patch is then used as the image patch contribution. Based on the image patch contribution, the image features within the same instance package are aggregated to generate the package-level classification result for the corresponding instance package.

[0127] The package-level classification results of each instance package are fused with the cell morphology evidence score to generate a blood disease classification result and classification confidence. The package-level classification results of each instance package are arranged in the order of pathological region identifiers to form a pathological classification vector. The pathological classification vector is concatenated with the cell morphology evidence score and input into a fully connected classifier to output the probability value corresponding to each blood disease category. The blood disease category with the highest probability value is taken as the blood disease classification result, and the probability value is taken as the classification confidence.

[0128] Image blocks with a contribution value not less than a preset contribution threshold are identified as basis image blocks. Image basis is generated based on the spatial coordinates of the basis image blocks. Examination suggestions are generated based on the blood disease classification results, classification confidence, and image basis. The preset contribution threshold is 0.3. The spatial coordinates of the basis image blocks include the horizontal and vertical coordinates of the upper left corner of the basis image block, the width and height of the image block, and the image pyramid level. The image basis consists of the position annotation map of the basis image block in the whole slice image to be analyzed in the pathology and the pathological region identifier corresponding to the basis image block. The examination suggestions are generated from a preset disease category and examination item correspondence table.

[0129] Example 1: To verify the feasibility of this invention in practice, it was applied to the analysis of whole-section images of blood smears and bone marrow biopsies. In this scenario, the diagnosis of hematological diseases relies on the comprehensive judgment of whole-section images of peripheral blood smears, bone marrow smears, and bone marrow biopsies with HE staining, immunohistochemistry, and reticular fiber staining. Traditional methods typically use only a single image source for analysis, making it difficult to simultaneously reflect single-cell morphological characteristics and tissue spatial distribution characteristics. Furthermore, their accuracy is limited when dealing with areas with blurred boundaries, abnormal infiltration, abundant granules, or uneven staining, thus affecting the accuracy of disease classification and pathological evidence.

[0130] In this embodiment, peripheral blood smear images, bone marrow smear images, and whole bone marrow biopsy slide images are acquired. All images undergo uniform preprocessing, including format conversion, pixel size standardization, staining normalization, background removal, and removal of out-of-focus or overlapping areas, generating cell images and whole pathological slide images for analysis. Cell morphology analysis is performed on the cell images. The StarDist model is used for preliminary segmentation, and a one-dimensional color gradient profile is extracted along the ray direction. An endpoint discriminator is used to perform secondary localization of structures that may be misidentified as boundaries, such as nucleoli, cytoplasmic granules, and staining deposits, generating real cell and nuclear contours. Based on this, cell morphological characteristics, including cell size, nucleocytoplasmic ratio, contour regularity, and granule distribution, are calculated.

[0131] For whole-section pathological images, sliding window segmentation is performed according to the target magnification level to select image blocks with a tissue region proportion greater than a threshold. These image blocks are then input into the DeepLabv3+ model to generate an initial pathological region segmentation map. Next, the boundaries of adjacent pathological regions are extracted and extended along the normal direction to form boundary rings. Boundary refinement is performed through a multi-level candidate point judgment and re-search mechanism to obtain a refined pathological region segmentation map. Subsequently, the images are sorted according to the proportion of pathological region categories in each image block, pathological region identifiers are configured, and spatially adjacent image blocks are divided into region instance bags and boundary instance bags. At the same time, image features are extracted and correlated with spatial coordinates to form a set of pathological region instance bags.

[0132] The principal orientation is extracted from bone marrow biopsy tissue to generate the tissue long axis. Image blocks within the instance package are projected onto the long axis of the bone marrow biopsy tissue to obtain the long axis coordinates, which are then sorted to form a long axis sequence. The long axis sequence, long axis coordinates, and distances between adjacent image blocks are input into the ABMIL model. Weakly supervised aggregation is performed on each instance package to generate package-level classification results and image block weights. The package-level classification results are fused with cell morphology features to generate blood disease classification results and classification confidence scores. Visualization images are determined based on image block weights exceeding a preset threshold. Examination suggestions are generated by combining disease classification results and confidence scores, achieving comprehensive utilization of single-cell information and tissue distribution information, thereby improving the accuracy and interpretability of disease diagnosis.

[0133] Table 1: Comparison of Indicators of Intelligent Diagnosis and Treatment Methods for Hematological Diseases

[0134] Disease classification accuracy (%) 82.3 88.7 94.6 F1 score 0.81 0.87 0.93 Boundary refinement average IoU 0.71 0.82 0.89 Mean single-cell nucleus profile deviation 2.8 1.9 1.2 Image patch information utilization rate (%) 85 92 96 Detection rate of abnormal bone marrow infiltration areas (%) 76 84 91

[0135] As shown in the table, the traditional unimodal method (Method A) achieved a disease classification accuracy of 82.3% and an F1 score of 0.81, with an average IoU of only 0.71 for boundary refinement and a significant single-cell nucleus contour deviation of 2.8 pixels. Furthermore, the image patch utilization and abnormal infiltration region detection rates were both below 80%, indicating deficiencies in boundary recognition and spatial information integration. The multimodal method, but without bone marrow long-axis sequencer (Method B), improved across all metrics, achieving an accuracy of 88.7% and an IoU of 0.82. This demonstrates that utilizing multimodal information can partially improve segmentation and classification accuracy, but it still fails to fully leverage the advantages of long-axis spatial relationships and instance aggregation. The method of this invention, through instance package construction, long axis serialization aggregation, and fusion of cell morphological features, achieves a disease classification accuracy of 94.6%, an F1 score of 0.93, an average IoU of 0.89 after boundary refinement, a single-cell nuclear contour deviation of 1.2 pixels, an image patch utilization rate of 96%, and a bone marrow abnormal infiltration area detection rate of 91%. The overall effect is superior to the existing technology, fully demonstrating the accuracy and interpretability advantages of this invention in intelligent diagnosis and treatment of hematological diseases.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent diagnosis and treatment assistance of blood diseases based on hybrid intelligence, characterized in that, Includes the following steps: Acquire cell morphology images and pathological whole slide images of the subject, perform preprocessing, and generate cell images and pathological whole slide images to be analyzed; Cell morphology analysis is performed on the cell image to be analyzed, staining decomposition is performed, cell center point, ray direction and initial endpoint are obtained based on cell instance segmentation model, cell boundary is corrected, and a set of cell morphology features is generated. Based on the set of cell morphological features, abnormal scores, predicted categories, and their contribution to disease classification, cells with similar morphology and related diagnostic significance are aggregated into cell morphology clusters. These cell morphology clusters are then used as human-machine collaborative review objects to generate cluster-level summaries. Receive cluster-level decision information from experts on cell morphology clusters, convert the cluster-level decision information into hybrid intelligent correction instructions, perform overall correction on cell morphology clusters and related cell groups, and generate a cluster-level corrected cell morphology evidence score. Image block segmentation, pathological region segmentation, and boundary refinement are performed on the whole-slice images of the pathological tissue to be analyzed. Based on the refined pathological region segmentation map and the spatial coordinates of image blocks, the pathological region identifier is determined, a set of pathological region instance packages is constructed, image features are extracted, and associated spatial coordinates are determined. The pathological region instance package is serialized according to the long axis of the bone marrow biopsy tissue and then input into the ABMIL model. The ABMIL output results are fused with the cluster-level corrected cell morphology evidence score to generate blood disease classification results, classification confidence, image evidence and examination suggestions.

2. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 1, characterized in that, The preprocessing specifically includes: Cell morphology images are one or more of the following: peripheral blood smear images, bone marrow smear images, and high-power field images of bone marrow cells; pathological whole slide images are one or more of the following: bone marrow biopsy HE-stained whole slide images, immunohistochemical stained whole slide images, reticular fiber stained whole slide images, and lymph node pathological whole slide images. The cell morphology image is converted in format, pixel size is unified and staining is normalized. The sharpness value, background pixel ratio and cell connected region area of ​​each field of view are calculated. Fields of view with sharpness value less than the preset sharpness threshold, fields of view with background pixel ratio greater than the preset background threshold, and cell connected region area greater than the preset overlap threshold are deleted to generate cell images to be analyzed. The image pyramid hierarchy of the whole pathological slide image is read, the tissue mask is extracted, the blank background outside the tissue mask is removed, the pixel blocks corresponding to slide folds, bubbles, staining deposits and out-of-focus areas are identified and removed from the image pyramid, the original spatial coordinates of the unremoved pixel blocks are retained, and the whole pathological slide image to be analyzed is generated.

3. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 2, characterized in that, The specific steps of performing cell morphology analysis on the cell image to be analyzed include: Stain decomposition is performed on the cell images to be analyzed to generate nuclear staining channels, cytoplasmic staining channels, and granular structure channels; Input the nuclear staining channel, cytoplasmic staining channel, and granular structure channel into the StarDist model, and output the cell instance center point, ray direction, and initial ray endpoint; A one-dimensional color scale profile is extracted along the ray direction from the center point of the cell instance to the outer side of the initial ray endpoint; Input the one-dimensional color scale profile into the endpoint discriminator to generate endpoint category discrimination results; When the endpoint category is determined to be a non-real cell boundary, search for real cell boundary candidate points along the same ray direction and replace the initial ray endpoint with the real cell boundary candidate points. Cell contour data and cell nucleus contour data are generated based on the replaced ray endpoints. Cell morphology indicators are calculated based on the cell contour data and cell nucleus contour data to generate a set of cell morphology features.

4. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 3, characterized in that, The process of aggregating morphologically similar and diagnostically relevant cells into morphological clusters specifically includes: Read the cell morphology feature set, anomaly score, predicted category and contribution to disease classification for each cell, and input the cell morphology feature set and anomaly score into the representation encoder to generate cell morphology representation vector; All cell morphology representation vectors are input into the clustering model to form initial cell clusters according to the proximity distribution of cells in the feature space; Based on the concentration of predicted categories and the range of variation in disease classification contribution within the initial cell cluster, cell regions with inconsistent diagnostic significance within the same initial cell cluster are identified, and these cell regions are reassigned to adjacent initial cell clusters or independently form new initial cell clusters. The cluster center position and intra-cluster morphological dispersion are calculated for the adjusted initial cell clusters. Initial cell clusters with intra-cluster morphological dispersion not greater than a preset dispersion threshold are identified as cell morphology clusters. Within a cell morphology cluster, cells near the cluster center with anomaly scores higher than the cluster average are selected as representative cells. Cluster-level summaries are generated based on representative cells, intra-cluster morphological dispersion, dominant prediction category, and disease support strength.

5. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 4, characterized in that, The overall correction of cell morphology clusters and related cell populations specifically includes: Cluster-level adjudication information refers to the category confirmation, category modification, artifact exclusion, key diagnostic confirmation, or re-clustering instructions made by experts for the entire cell morphology cluster; Cluster-level decision information is encoded into hybrid intelligent correction instructions, which are parameterized instructions consisting of expert decision direction, target cell type, correction range, and correction intensity. Based on the cell morphology representation vector, the cells that are closest to the morphological clusters of the adjudicated cells and are related to the predicted category are retrieved from the cells that have not been adjudicated by experts, and the associated cell population is generated. According to the hybrid intelligent correction instructions, the abnormal scores, predicted category probabilities, disease classification contributions, and artifact exclusion status of the adjudicated cell morphology clusters and related cell populations are updated synchronously. Based on the updated anomaly scores, predicted category probabilities, and disease classification contributions, the support value of each cell morphology cluster for each blood disease category is recalculated, and the support values ​​are normalized to generate a cluster-level corrected cell morphology evidence score.

6. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 5, characterized in that, The boundary refinement specifically includes: Extract adjacent pixel pairs of different categories from the initial pathological region segmentation map to form the boundary lines of adjacent pathological regions; The image content of the whole slice image to be analyzed and the category label in the initial pathological region segmentation map are extracted along the normal direction of the boundary line of the adjacent pathological region to generate the boundary ring; Set boundary sampling points according to the direction of the boundary lines of adjacent pathological regions, and sample the boundary ring along the normal direction of the boundary sampling points to generate a rectangular image and a rectangular category map. The first boundary candidate point is determined based on the staining change value and cell density change value in the rectangular image, and the second boundary candidate point is determined based on the continuous interval of the category in the rectangular category map. The first boundary candidate point is matched with the second boundary candidate point. If the matching does not meet the preset conditions, the category continuous interval is searched again along the corresponding normal sampling line to determine the corrected boundary point. The corrected boundary points are connected to form a corrected boundary line. The pixel category within the boundary ring is updated based on the corrected boundary line to generate a refined pathological region segmentation map.

7. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 6, characterized in that, The specific collection of pathological region instance packages includes: Read the pathological region identifier, pathological region category ratio, and spatial coordinates of each image block, and establish spatial adjacency relationships for image blocks with the same pathological region identifier and a spatial distance not greater than a preset adjacency distance; Based on spatial adjacency, image blocks under the same pathological region are divided into several connected image block groups; Calculate the average proportion of pathological region categories and spatial coverage of image blocks within each connected image block group. Determine the connected image block group with an average proportion of pathological region categories not less than a preset purity threshold as a region instance pack. Determine the connected image block group with combined coding and located between two region instance packs as a boundary instance pack. Image features are extracted from each image block within the region instance package and the boundary instance package. The image features, spatial coordinates, pathological region identifiers, and pathological region category proportions are bound together according to the image block number to generate a set of pathological region instance packages.

8. The intelligent diagnostic and treatment assistance method for blood diseases based on hybrid intelligence according to claim 7, characterized in that, The fused ABMIL output results and the cluster-level corrected cell morphology evidence score specifically include: Extract the bone marrow biopsy tissue region from the whole slice image to be analyzed in pathology, and determine the long axis of the bone marrow biopsy tissue based on the principal direction of the bone marrow biopsy tissue region; Read the image features and spatial coordinates of image blocks within the pathological region instance package set, project the center point of the image block onto the long axis of the bone marrow biopsy tissue to obtain the long axis coordinates, and generate a long axis image block sequence according to the long axis coordinates within the corresponding region instance package or boundary instance package. The sequence of long-axis image patches and their long-axis coordinates are input into the ABMIL model for weakly supervised aggregation, generating package-level classification results and image patch contribution of instance packages; The package-level classification results of each instance package are fused with the cluster-level corrected cell morphology evidence scores to generate blood disease classification results and classification confidence. The basis image block is determined based on the contribution of the image block, and the image basis and examination suggestions are generated by combining the spatial coordinates of the basis image block, the blood disease classification results and classification confidence.

Images