A breast cancer chemotherapy prediction method based on pathology and spatial transcriptome anchor library

Abstract

The present application relates to the technical field of digital pathology analysis, and discloses a breast cancer chemotherapy prediction method based on pathology and spatial transcriptome anchor library, which acquires full-image, spatial transcriptome data, baseline clinical information and efficacy label in the training stage, maps transcriptome points to analysis areas corresponding to images, respectively extracts niche pathological morphology vector and molecular mechanism vector; constructs a pathological response anchor library based on morphology and molecular prototype, efficacy consistency and co-occurrence intensity; in the clinical reasoning stage, only needs to input patient's routine H&E staining full-image and clinical information, extracts regional pathological morphology vector, matches with anchor library morphology centroid, generates hit intensity, and then outputs response probability, drug resistance risk, scheme ranking, case evidence and mechanism prompt. The present application limits high-cost transcriptome detection to the library construction period, and does not need molecular detection as input in the reasoning period, so that low-cost and interpretable chemotherapy response evaluation can be realized.

Description

Technical Field

[0001] This invention relates to the field of digital pathology analysis technology, specifically to a method for predicting breast cancer chemotherapy based on pathology and spatial transcriptome anchor libraries. Background Technology

[0002] Breast cancer patients often face the choice of different chemotherapy drugs or combination regimens during neoadjuvant chemotherapy, adjuvant chemotherapy, and advanced systemic therapy. Current clinical decision-making mainly relies on molecular subtyping, histological grading, imaging findings, previous treatment history, and physician experience. Even within the same molecular subtyping, different patients exhibit significant differences in sensitivity to the same treatment regimen. Some patients show low sensitivity or potential resistance before treatment begins, while others gradually develop acquired resistance during treatment. Therefore, identifying patients who may be sensitive or insensitive to a specific treatment regimen before or early in the course of treatment is crucial for improving treatment efficiency and reducing ineffective treatment.

[0003] Digital pathology images, capable of recording cellular morphology, histological conformation, stromal reaction, immune infiltration, and spatial heterogeneity of tumor tissue, have become crucial inputs for artificial intelligence analysis in pathology. Existing prediction methods based on digital pathology images mostly employ end-to-end classification models, typically directly outputting response probabilities or risk scores. While these methods possess some predictive capability, their interpretations often remain at the level of attention heatmaps or local image patches, failing to explain the potential molecular mechanisms underlying image regions or explain why a particular morphological pattern is associated with specific drug sensitivity or resistance.

[0004] Spatial transcriptomics technology can provide local gene expression status while preserving tissue location information, making it highly valuable for identifying drug resistance microenvironments, cellular composition, pathway activity, and local cell interactions. However, spatial transcriptomics testing is costly, has a long preparation cycle, and suffers from significant platform differences. Furthermore, a single site on a conventional site-based platform often covers multiple cells and cannot be simply equated to a single cell or pathological structure. Requiring spatial transcriptomics testing for every patient before clinical decision-making would significantly increase time and economic costs, hindering widespread deployment in real-world clinical procedures.

[0005] Therefore, a novel data processing method is needed in this field. This method should be able to establish a stable correspondence between pathological morphological patterns and local molecular mechanisms and therapeutic directions using a small amount of high-value digital pathology whole-slice images paired with spatial transcriptome data during the training phase; and during the clinical inference phase, it should not require spatial transcriptome, immunohistochemistry, liquid biopsy, or panoramic molecular detection as necessary inputs, but rather output chemotherapy response probabilities, drug resistance risk stratification, and verifiable potential mechanism hints under specific treatment scenarios based solely on conventional H&E digital pathology whole-slice images and baseline clinical information.

[0006] Existing technologies include individual techniques such as multimodal fusion, image classification, spatial omics analysis, and case retrieval. However, a complete technical solution is lacking, which involves establishing a pathological and spatial transcriptomic response anchor library around a specific treatment scenario and using anchor hit strength to predict single-modal pathology and provide mechanism hints during the inference stage. Simply superimposing multiple model concepts makes it difficult to form clear and implementable steps, and also makes it difficult to clearly define the objects of protection at the patent level. To address the above shortcomings, this invention proposes a data processing flow centered on niche registration, prototype binding, anchor library construction, anchor matching, and probability calculation. Summary of the Invention

[0007] The purpose of this invention is to provide a method for predicting breast cancer chemotherapy based on pathology and spatial transcriptome anchor libraries. During the training phase, the system uses spatial transcriptome data to establish an interpretable correspondence between pathology and molecular mechanisms. During the clinical reasoning phase, it can complete the auxiliary assessment by relying only on conventional H&E digital pathology full slide images and baseline clinical information, thereby reducing detection costs and improving the interpretability of results and the feasibility of clinical deployment.

[0008] To achieve the above objectives, the present invention adopts the following technical solution. The system first constructs a training case set that conforms to a preset treatment scenario. The treatment scenario is jointly defined by drug category, combination regimen, treatment line, molecular subtyping, sampling time point, and efficacy labeling rules to avoid model deviation caused by the mixing of different drug backgrounds, different labeling windows, and different sampling times.

[0009] The system reads digital pathology full-slice images, spatial transcriptome location data, baseline clinical information, and efficacy outcome labels from training cases. Based on the coordinate transformation relationship between the spatial transcriptome reference image and the digital pathology full-slice image, each spatial transcriptome location is mapped to a corresponding analysis region in the digital pathology full-slice image. For each corresponding analysis region, the system calculates tissue coverage, boundary clutter, and registration confidence, filtering out or downweighting low-quality regions.

[0010] The system extracts cellular morphological features, tissue structure features, microenvironmental features, and spatial topological features from the corresponding analysis region to generate a niche pathological morphology vector. Simultaneously, the system calculates the activity of drug resistance-related pathways, cell composition ratios, and local cell interaction strengths from the spatial transcriptome site expression matrix to generate a niche molecular mechanism vector.

[0011] The system extracts prototypes from niche pathological morphology vectors and niche molecular mechanism vectors under the same treatment scenario encoding, forming morphological prototypes and molecular prototypes. The system further calculates the co-occurrence strength between morphological and molecular prototypes, and, considering the consistency of therapeutic direction and scenario consistency, binds morphological and molecular prototypes that meet the conditions as pathological response anchors. Multiple pathological response anchors constitute a pathological response anchor library.

[0012] During the clinical inference phase, target patients only need to provide H&E-stained digital pathology images and baseline clinical information. The system performs the same preprocessing and morphological vector extraction on the target patient images as during the training phase, calculating the similarity between the target patient's niche pathological morphological vector and the morphological centroids of each anchor point in the anchor point library to generate patient-level anchor point hit strength. The system then calculates the chemotherapy response probability and drug resistance risk score based on the anchor point hit strength, the anchor point weights corresponding to candidate treatment plans, the treatment scenario encoding, and the baseline clinical information, and outputs potential mechanism hints, similar case evidence, and supplementary review prompts.

[0013] The present invention has the following beneficial effects: 1. This invention limits spatial transcriptomics to training-phase library construction data, rather than the necessary input for the reasoning phase of each target patient, thereby reducing the cost of clinical application and the waiting time for testing.

[0014] 2. This invention establishes niche training samples by softly aligning spatial transcriptome sites with digital pathological image regions, avoiding the forced interpretation of multicellular mixed sites as single-cell morphology, thereby improving the reliability of the morphological-molecular correspondence.

[0015] 3. This invention generates pathological response anchors by using morphological prototypes, molecular prototypes, therapeutic directions, and co-occurrence intensity, so that model interpretation no longer relies solely on black-box attention heatmaps, but can trace back to stable pathological morphology, local molecular mechanisms, and historical case evidence.

[0016] 4. This invention uses anchor point hit strength and a clear response scoring function to output the chemotherapy response probability, which facilitates response reference ranking among different candidate protocols and also facilitates calibration based on newly added real-world outcomes.

[0017] 5. The present invention sets up mechanisms for low-confidence region removal, low-confidence result prompts, similar case retrieval, cross-center correction, and anchor point database updates, which are beneficial to improving the robustness and long-term maintainability of multi-center deployments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 A schematic diagram illustrating the process of mapping spatial transcriptome sites to digital pathology image analysis regions; Figure 3 A schematic diagram illustrating the construction process of niche pathological morphology vectors and niche molecular mechanism vectors; Figure 4 A schematic diagram illustrating the generation process of morphological prototypes, molecular prototypes, and pathological response anchor point libraries; Figure 5 A schematic diagram of the process for anchor point matching, response scoring, and mechanism hint output in the clinical reasoning stage; Figure 6 A schematic diagram illustrating the process of writing back to the real-world ending and updating the anchor point library. Detailed Implementation

[0019] The present invention will be described in detail below with reference to specific embodiments. These embodiments are used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the core idea of ​​the present invention, those skilled in the art can make equivalent substitutions for image segmentation algorithms, clustering algorithms, path scoring methods, similarity functions, parameter thresholds, and storage formats.

[0020] 1. Terminology and Symbol Conventions Digital pathology whole-slide images refer to digital images of pathological slides acquired by a whole-slide scanner, preferably H&E-stained slide images. Spatial transcriptome loci refer to measurement units with 3D spatial coordinates and local gene expression readings. Corresponding analysis regions refer to the regions in the digital pathology whole-slide image corresponding to spatial transcriptome loci after coordinate transformation. Pathological response anchor points refer to intermediate knowledge objects that stably connect a certain type of pathological morphology, a certain type of molecular mechanism state, and a certain therapeutic direction in a specific treatment scenario.

[0021] This article uses The spatial transcriptome locus or its corresponding analysis region number is represented by... The region number in the digital pathology image of the target patient is represented by... Indicates the pathological response anchor point number, using This represents the candidate treatment option number. The niche pathological morphology vector is denoted as... The molecular mechanism vector of the ecological niche is denoted as The anchor point shape and centroid are denoted as The anchor point molecular center of mass is denoted as The patient-level anchor point hit strength is recorded as .

[0022] ; 2. Step S100: Treatment scenario limitation and construction of training case cohort Step S100 is used to define the task boundaries before database construction. The system establishes a treatment scenario dictionary, encoding drug categories, combination regimens, treatment lines, molecular subtypes, sampling time points, and efficacy labeling rules. The treatment scenario encoding can be represented as:

[0023] in, Indicates the drug category code, Indicates the joint scheme coding, Indicates the treatment line code. This indicates the molecular subtype code for breast cancer. Indicates the time point code for material collection. This indicates the coding rules for efficacy labeling. For neoadjuvant anthracycline combined with taxane chemotherapy in triple-negative breast cancer, a separate scenario coding can be set up; for platinum-based regimens, HER2-positive combination regimens, or advanced lineage regimens, scenario coding and corresponding anchor point libraries should be established separately, and they should not be directly mixed into the same training task.

[0024] The system reads the medication records, molecular subtyping, pathological diagnosis, tissue sampling time points, and follow-up outcomes of training cases. For neoadjuvant chemotherapy scenarios, pathological complete remission is preferred as the primary label, with residual tumor burden or Miller-Payne grade as secondary labels. For advanced chemotherapy scenarios, solid tumor efficacy evaluation classification, progression-free survival, and progression events within a pre-defined time window are preferred as labels. The label time window should be fixed before training and should not be adjusted temporarily for each case during the modeling process.

[0025] Cases with complete digital pathology images, spatial transcriptome locus data, baseline clinical information, and efficacy outcome labels are included as cases for database construction. If the spatial transcriptome data comes from adjacent consecutive slides, it should have tissue outlines or morphological landmarks that can be used for registration. Cases with unconfirmed pairings, severely damaged slides, unclear labels, mixed medication regimens, or inconsistent sampling times are excluded from the supervised database construction queue.

[0026] ; 3. Step S200: WSI and ST pairing, coordinate mapping, and area quality control. Step S200 is used to establish niche training samples. The system reads the center coordinates of the spatial transcriptome sites, the site diameter, the spatial transcriptome reference image, and the digital pathology full-slice image. If the two come from the same slice, rigid transformation combined with local deformation correction can be used; if the two come from adjacent consecutive slices, coarse registration can be performed first based on the tissue outline, blood vessels, glandular structures, necrotic area boundaries, or artificial markers, and then local fine-tuning can be performed.

[0027] Let the first The center coordinates of each spatial transcriptome site are: The coordinate transformation function is Then the center position of this point in the coordinate system of the digital pathology full-slice image is:

[0028] Let the diameter of this point be The soft boundary coefficient is If the value is preferably between 1.0 and 1.4, then the radius of the corresponding analysis region is:

[0029] Corresponding analysis area Defined as Centered on, with radius The circular region can also be set as an elliptical or polygonal region depending on the local deformation of the tissue. A circular region can be represented as:

[0030] The system calculates three quality indicators for each corresponding analysis region. Organizational coverage. The proportion of the effective organized area within the region to the total area of ​​the region; boundary miscellaneousness. Information entropy of the distribution of major organizational categories within the region; registration confidence level This is a weighted result of landmark matching score, local cross-correlation score, and deformation smoothness.

[0031]

[0032]

[0033]

[0034] in, Indicates the first The first region The area percentage of tissue-like components, This indicates the local landmark matching score. This represents the local cross-correlation score of the image. Indicates the deformation smoothness score. , , The weights are non-negative and their sum is 1. If Below the coverage threshold, or Above the confounding threshold, or If the region is below the registration threshold, the system marks it as a low-confidence region. Low-confidence regions can be used for unsupervised preprocessing, but will not proceed to the core anchor binding step.

[0035] ; 4. Step S300: Construction of Niche Pathological Morphology Vectors Step S300 is used to convert the corresponding analysis region into a computable pathological morphological representation. The system first constructs a tissue mask at low magnification and excludes blank background, and then identifies invasive carcinoma areas, in situ lesion areas, stroma areas, necrotic areas, inflammatory infiltration areas, and normal tissue areas at medium and high magnification. For the regions entering feature extraction, it is preferred that the tumor area or the microenvironment area closely related to the tumor response occupies a preset proportion of the region area.

[0036] The system for the region The cell nuclei within the cell are segmented, and cell type is determined where feasible. The basic morphological attributes of each nucleus include nuclear area, nuclear perimeter, long axis, short axis, roundness, chromatin texture, nuclear density, and neighborhood density. Roundness can be calculated using the following formula:

[0037] in, For the first Area of ​​each cell nucleus Let be the perimeter of the nth cell nucleus. The system's region... The mean, standard deviation, skewness, quantiles, and hotspot proportions of all cell nuclei are calculated to form a cell-level statistical vector. .

[0038] The system further extracts organizational structure-level features. The tissue-level features include the degree of gland formation, the size and distribution of tumor nests, the irregularity of tumor boundaries, the proportion of necrotic area, the distribution of vascular structures, and the complexity of the invasion front. Boundary irregularity can be represented by the ratio of the tumor region's contour length to the contour length of a circle of the same area.

[0039] in, This represents the length of the tumor region's outline. This represents the area of ​​the tumor region. The larger the value, the more irregular the local boundary.

[0040] System extracts microenvironment-level features This includes at least the density of tumor-infiltrating lymphocytes, the proportion of lymphocytes in the stroma, the degree of fibroblast enrichment, the degree of inflammatory cell aggregation in the vicinity of necrotic tissue, the proximity of blood vessels, and the length of the tumor-stromal interface. The density of tumor-infiltrating lymphocytes can be obtained by dividing the number of lymphocytes by the effective tissue area.

[0041]

[0042] The system constructs a cell adjacency graph within the region to extract spatial topological features. In the adjacency graph, nodes represent cells or small regions, and edges represent adjacency relationships where the spatial distance is less than a preset adjacency threshold. For any node... Its local clustering coefficient can be expressed as:

[0043] in, For nodes The number of adjacent nodes, This represents the actual number of connections between these adjacent nodes. When When less than 2, Set to 0. The niche space topology vector can be composed of the mean node degree, the variance of node degree, the nearest neighbor distance between different cell types, the local clustering index, and the mean local clustering coefficient.

[0044] The niche pathological morphology vector is formed by the concatenation of four types of vectors:

[0045] To avoid imbalances caused by different dimensional features in subsequent clustering, the system performs robust standardization on each feature dimension. For the first feature dimension... For each feature dimension, standardization can be calculated using the median and interquartile range:

[0046] in, For the training set The median of the features, This is the interquartile range. If the interquartile range is 0, then this dimension is uniformly set to 0 or removed from the database features.

[0047] ; 5. Step S400: Construction of Niche Molecular Mechanism Vectors Step S400 is used to extract low-redundancy molecular representations related to treatment response and drug resistance mechanisms from spatial transcriptome loci. The system first performs quality control on the locus expression matrix, removing loci with low reads, low gene counts, or abnormal mitochondrial proportions. For the first... Raw readings of gene g at each site The normalized expression value can be calculated using the following formula:

[0048] The system pre-sets a set of gene modules related to the target treatment scenario. This is based on the l-th pathway or gene module. For example, location The pathway activity score can be expressed as:

[0049] In a preferred embodiment, the pathway or gene module includes modules related to DNA damage repair, drug efflux, epithelial-mesenchymal transition, cell cycle, apoptosis escape, immunosuppression, hypoxia response, and fibroblast activation. The system can assign different module weights to different treatment regimens. For example, platinum-based regimens focus more on DNA damage repair status, while taxane-based regimens focus more on cell cycle, microtubule-related processes, and drug efflux status.

[0050] The system performs cellular composition analysis on spatial transcriptome loci. Non-negative least squares or non-negative matrix factorization can be used to represent the locus expression vectors as non-negative combinations of several cell type reference expression profiles. Let... This is a cell type reference expression matrix. For the first Each point represents a vector. Given the cell composition weights to be determined, the following objective function can be solved:

[0051] Seek Normalization was then performed to obtain the composition ratios of tumor cells, T cells, macrophages, fibroblasts, endothelial cells, etc. .

[0052] The system calculates the local cell interaction strength. Taking the interaction between tumor cells and fibroblasts as an example, the local interaction strength can be obtained by multiplying and standardizing the proportion of fibroblasts, the proportion of tumor cells, and the score of fibroblast activation pathways.

[0053] in, This indicates standardization. Similarly, the strength of immunosuppressive interactions, vascular-associated interactions, and macrophage-associated interactions can be calculated.

[0054] The niche molecular mechanism vector is formed by the cascade of pathway activity vector, cellular composition ratio vector, and local interaction strength vector:

[0055] ; 6. Step S500: Weak labeling of regional efficacy direction and setting of sample weights Since efficacy outcomes such as pCR and PFS are typically case-level labels rather than natural location-level labels, directly assigning case-level labels to all regions can easily introduce noise. This invention uses case outcomes and regional molecular resistance indices to jointly determine weak niche labels.

[0056] The system sets weights for drug resistance-related pathways based on the treatment scenario. Calculate the first Molecular resistance index of each region:

[0057] in, For the first Pathway activity, For the first Local interaction strength, and Context-related weights. For cases from response cases and Regions below the lower percentile threshold for the same scenario are marked as supportive response regions by the system; for regions from non-responding cases and... Regions with values ​​above the high percentile threshold in the same scenario are marked as drug resistance support regions by the system; the remaining regions are marked as neutral regions or low-weight regions.

[0058] Regional sample weights It is determined by alignment quality, organization quality, and the reliability of weak annotations:

[0059] in, The normalized boundary disorder. This represents the confidence level for weak labeling. The higher the value, the more suitable the region is for participating in morphological prototypes, molecular prototypes, and anchor point binding.

[0060] 7. Step S600: Construction of morphological prototype, molecular prototype and pathological response anchor library Step S600 is the core step of this invention. Under the same treatment scenario encoding, the system performs clustering on the pathological morphological vectors of high-confidence regions to generate a set of morphological prototypes. For the a-th morphological prototype, its morphological centroid is:

[0061] in, Indicates belonging to the first The system performs clustering on molecular mechanism vectors to generate a set of molecular prototypes. For the ... A prototype molecule, whose molecular center of mass is:

[0062] The co-occurrence intensity of the morphological prototype and the molecular prototype in the system statistics. Let... This represents the number of high-confidence regions that simultaneously belong to both morphological prototype a and molecular prototype b. It belongs to the prototype of form The number of high-confidence regions. It belongs to the molecular prototype The number of high-confidence regions is then used to define the co-occurrence intensity as:

[0063] The higher the value, the more stably a certain morphological pattern co-occurs with a certain molecular mechanism. The system further calculates the consistency of the therapeutic direction between the two. If the high-weighted region within the morphological prototype mainly comes from responding cases and the corresponding molecular prototype has a low drug resistance index, the direction is marked as responsive; if the high-weighted region within the morphological prototype mainly comes from non-responding cases and the corresponding molecular prototype has a high drug resistance index, the direction is marked as resistant; the rest are marked as intermediate.

[0064] when Not lower than the co-occurrence threshold When the morphological prototype and the molecular prototype have the same therapeutic direction and belong to the same treatment scenario coding, the system will bind them as pathological response anchors. The data structure for pathological response anchor points is as follows:

[0065] in, Encoding the treatment scenario, For the center of mass of the form, Center of mass of the molecule, For therapeutic purposes, For the level of evidence, This is a set of coordinates representing the pathological regions. For mechanism pathway label set, This is a collection of historical case indexes.

[0066] Level of evidence Based on internal resampling stability Consistency of external verification direction Score supported by literature or pharmacogenomics evidence In summary:

[0067] in, , , The weights are non-negative and sum to 1. Internal resampling stability can be represented by the proportion of the same anchor point appearing repeatedly after multiple sampling and library construction; external validation direction consistency can be represented by the proportion of anchor points with consistent response directions in the external cohort; the literature or pharmacogenomics evidence support score is used to reflect the degree of consistency between the mechanism pathway label and existing evidence.

[0068] ; 8. Step S700: Anchor point matching and response probability calculation in the clinical reasoning stage. Step S700 is used for clinical inference for the target patient. This stage does not require the target patient to provide spatial transcriptome data, nor does it require immunohistochemistry, liquid biopsy, or panoramic molecular detection as necessary inputs to the model. The system reads the target patient's full-length H&E digital pathology images and baseline clinical information, performing quality control, tissue segmentation, tumor region identification, cell nuclear segmentation, cell classification, and niche pathological morphology vector extraction consistent with the training phase.

[0069] Let the pathological morphology vector of the j-th region of the target patient be... , No. The shape and centroid of each anchor point are The similarity of niche anchor points can be calculated using the following formula:

[0070] in, For the first The anchor points have discrete morphological scales in the training set. To avoid diluting the signal with a large number of irrelevant regions across the entire image, the system selects the anchor points with the highest similarity. The average value across the regions is used to obtain the patient-level anchor point hit strength.

[0071] in, The number of regions in which the target patients participated in the analysis. This indicates taking the first part after sorting. The average value The value is preferably between 1 and 10.

[0072] For candidate treatment options The system reads the anchor point weight table corresponding to the proposed scheme. The response probability is generated using the following scoring function:

[0073] in, For the plan The response probability, For the plan The intercept term, anchor point In the plan The response weights are as follows: This is the baseline clinical information vector. Weights for clinical variables. If anchor point It is drug-resistant and in the program If the manifestation is unfavorable, then Take the negative weight; if the anchor point If it is responsive, then Take positive weights.

[0074] The system can also calculate drug resistance risk scores. The drug resistance risk score is determined by the strength of the drug resistance anchor, the level of evidence for the drug resistance anchor, and clinically high-risk variables.

[0075] The system according to , and confidence level A reference ranking of candidate treatment responses is established. The overall benefit index can be expressed as:

[0076] in, This is the penalty coefficient for drug resistance risk. For confidence level weights. The system follows... or The candidate treatments were ranked, and the report should clearly state that the ranking is an auxiliary evaluation result and does not replace the clinical physician's judgment.

[0077] 9. Step S800: Mechanism hints, visualization, and similar case retrieval The system generates potential drug resistance mechanism suggestions based on the hit anchor points. For any mechanism pathway... The system aggregates all hit anchors carrying the corresponding pathway label and calculates the score accordingly.

[0078] in, For the carry mechanism pathway The set of anchor points for tags Used to indicate the direction of the anchor point. For resistant anchor points, Take positive resistance indications; for responsive anchors, Retrieve response support prompts. The system based on... Potential mechanisms suggested by the study include enhanced drug efflux, enhanced epithelial-mesenchymal transition, enhanced DNA damage repair, immunosuppressive microenvironment, fibroblast-related drug resistance, and vascular-related remodeling.

[0079] The visualization results are generated based on the similarity of niche anchor points. The system maps highly similar regions back to the coordinates of the digital pathology full-slice image, forming a heatmap or bounding box map of key regions. Unlike simple attention heatmaps, the visualization regions of this invention are simultaneously connected to specific anchor points, representative pathological regions of the anchor points, molecular centroids, and mechanism pathway labels, facilitating physician review.

[0080] The system performs a similar case search. Let the target patient be... Historical cases are The overall similarity between the two can be expressed as:

[0081] in, Similarity is represented by patient-level morphology. For anchor point hit spectrum similarity, To score the consistency of treatment scenarios, For baseline clinical variable similarity, to The weights are non-negative and the sum is The system first recalls candidate cases in the patient-level representation space, then reorders them according to comprehensive similarity, and outputs the top 3 to 10 similar cases and their efficacy outcome summaries.

[0082] ; 10. Step S900: Low-confidence control, real-world update, and cross-center verification The system performs confidence control on the inference results. If the target patient's slide quality is low, the overall anchor point hit strength is low, the anchor point hit direction is scattered, the number of similar cases is insufficient, or the distribution of the target patients deviates significantly from the database, the system will reduce the automatic output level and provide supplementary review suggestions. Supplementary review suggestions may include immunohistochemistry, fluorescence in situ hybridization, circulating tumor DNA detection, short-cycle image review, pathological slide rescanning, or manual pathological review.

[0083] The system continuously receives real-world treatment outcomes, including post-treatment pathological review, imaging responses, progression-free survival events, recurrence and metastasis events, and available molecular verification results. New data is not directly added to the formal anchor database; instead, it first enters a candidate update pool. The system performs distribution drift monitoring, intercenter correction, and one-center-only validation on the candidate update pool. Anchor database updates are only performed when the differences in staining style, scanner origin, tissue composition, anchor hit distribution, and label incidence between new cases and original cases are within acceptable limits, and external validation results meet preset criteria.

[0084] For high-value anchors, the system can establish orthogonal validation pathways. For example, immunohistochemistry, fluorescence in situ hybridization, in situ hybridization, or spatial proteomics can be performed on adjacent consecutive slices to confirm whether high-contribution pathological regions spatially correspond to candidate molecular events. This orthogonal validation is not a necessary input for routine reasoning about the target patient, but rather an enhancement step used to improve the level of anchor evidence and the model's credibility.

[0085] 11. Optional Implementation: Lightweight Morphology-to-Molecular Pathway Estimator In implementations with sufficient data, the system can train a lightweight morphology-to-molecular pathway estimator outside the anchor point library. This estimator takes niche pathological morphology vectors as input and outputs several molecular pathway activity estimates. This estimator serves only as an auxiliary verification module for anchor point matching and is not a necessary component of the main solution of this invention. This setup avoids using a complex cross-modal distillation network as a core prerequisite, reducing the difficulty of practical deployment.

[0086] The lightweight estimator can be implemented using linear regression, ridge regression, gradient boosting trees, or shallow multilayer perceptrons. The training objective is to minimize the error between the estimated pathway activity and the actual pathway activity in the spatial transcriptome. If the output of the lightweight estimator is consistent with the mechanism cues from the anchor point, the confidence level of the mechanism cues is increased; if the two are significantly inconsistent, the system lowers the level of the mechanism cues and suggests further verification.

[0087] Example 1: Anchor point library construction in the context of neoadjuvant anthracycline combined with taxanes in triple-negative breast cancer In one specific embodiment, the system uses neoadjuvant anthracycline combined with taxane chemotherapy in triple-negative breast cancer as the target scenario, and the treatment scenario code is set as SC-NACT-TNBC-AT. Training cases are all derived from pre-treatment biopsy samples and include H&E digital pathology full-section images, spatial transcriptomic data of adjacent continuous sections, baseline clinical information, and a postoperative pathological complete remission label.

[0088] The system maps spatial transcriptome loci to the H&E image coordinate system, extracting nuclear atypia, tumor nest boundaries, TIL density, stroma ratio, CAF enrichment, and cell adjacency map features within the corresponding region of each locus to form a niche pathological morphology vector. Simultaneously, the system calculates scores for DNA damage repair, EMT, drug efflux, cell cycle, immunosuppression, and CAF activation pathways from the spatial transcriptome loci, and analyzes the composition ratios of tumor cells, T cells, macrophages, CAF, and endothelial cells to form a niche molecular mechanism vector.

[0089] Within the same scenario, the system clusters morphological vectors and molecular vectors separately. If a morphological prototype exhibits high nuclear atypia, active border invasion, scarce TILs, and significant interstitial activation, and stably co-occurs with molecular prototypes of enhanced EMT, enhanced CAF interaction, and enhanced drug efflux, and this co-occurrence mainly originates from non-pCR cases, then the system generates a drug-resistant pathological response anchor point for this binding result. If a morphological prototype exhibits high tumor regional immune infiltration, low interstitial pressure, and stably co-occurs with immune-activating molecular prototypes, and mainly originates from pCR cases, then the system generates a responsive pathological response anchor point for it.

[0090] Example 2: Single-modal pathological reasoning of target patients During the clinical inference phase, the target patient only provides pre-treatment H&E digital pathology full-slice images and baseline clinical information. The system performs tissue segmentation, tumor region identification, cell nucleus segmentation, and niche morphological vector extraction, and calculates the morphological centroid similarity between each region and anchor points in the same scene library. If multiple high-contribution regions of the target patient hit drug resistance anchor points, and the baseline clinical information shows a high tumor burden or adverse risk factors, the system assigns a high drug resistance risk score.

[0091] The report displays the coordinates of high-contribution regions, corresponding pathological concepts, anchor point numbers, similar historical cases, and potential mechanism hints. If the main anchor point corresponds to EMT enhancement and CAF interaction enhancement, the system describes it in the report as indicating the presence of EMT-related or matrix-related drug resistance clues, rather than stating the clues as a confirmed causal mechanism.

[0092] Example 3: Multi-center validation and enhanced anchor evidence level In multi-center applications, the system first establishes a pathological response anchor database at center A, and then uses untested cases from centers B and C for external validation. If an anchor is resistant in center A but still primarily hits non-responsive cases in centers B and C, the consistency of the external validation direction for that anchor is improved. If the mechanism pathway corresponding to the anchor is reasonably supported in publicly available literature or pharmacogenomics data, the level of evidence is further enhanced.

[0093] For anchor points with high levels of evidence and strong clinical significance, the system can select adjacent consecutive sections from representative cases for immunohistochemical or in situ hybridization validation. For example, for CAF-related drug resistance anchor points, fibroblast activation markers can be validated on adjacent sections; for drug efflux-related anchor points, the expression of relevant transport proteins can be validated. Validation results are used for updates. However, this does not change the input requirement that the inference stage relies solely on H&E images.

Claims

1. A breast cancer chemotherapy prediction method based on pathology and spatial transcriptome anchor library, characterized in that, Includes the following steps: Acquire digital pathology full-slice images, spatial transcriptome locus data, treatment scenario coding, baseline clinical information, and efficacy outcome labels for training cases; Based on the coordinate transformation relationship between the coordinates of the spatial transcriptome loci and the coordinates of the digital pathology full-slice image, each spatial transcriptome locus is mapped to a corresponding analysis region in the digital pathology full-slice image. Cell-level morphological features, tissue-level features, microenvironment-level features, and spatial topology-level features are extracted from the corresponding analysis region to generate a niche pathological morphology vector. From the spatial transcriptome locus data, drug resistance-related pathway activity, cell composition ratio, and local cell interaction intensity are extracted to generate a niche molecular mechanism vector. Prototype extraction is performed on the niche pathological morphology vector and the niche molecular mechanism vector to generate morphological prototypes and molecular prototypes, respectively. Based on the co-occurrence intensity, efficacy direction consistency, and treatment scenario consistency of morphological prototypes and molecular prototypes, morphological prototypes and molecular prototypes that meet preset conditions are bound together to generate a pathological response anchor point library. During the inference phase, the system receives H&E-stained digital pathology full-slice images and baseline clinical information of the target patient and extracts the niche pathological morphology vector of the target patient. Calculate the similarity between the target patient's niche pathological morphology vector and the centroid of each anchor point in the pathological response anchor point library, and generate the patient-level anchor point hit strength. Based on the patient-level anchor hit strength, treatment scenario coding, and baseline clinical information, the system outputs chemotherapy response probability, drug resistance risk stratification results, and potential drug resistance mechanism hints corresponding to the hit anchors.

2. The method of claim 1, wherein, The treatment scenario coding consists of drug category coding, combination regimen coding, treatment line coding, molecular subtyping coding, sampling time point coding, and efficacy label rule coding, which are used to limit the training samples, anchor library, and inference task to be in the same treatment context. Among them, the efficacy outcome labels for the neoadjuvant chemotherapy scenario include pathological complete remission, residual tumor burden, or Miller-Payne grade, while the efficacy outcome labels for the advanced chemotherapy scenario include solid tumor efficacy evaluation classification, progression-free survival, or progression events within a preset time window.

3. The method according to claim 1, characterized in that, The steps of mapping each spatial transcriptome locus to a corresponding analysis region in a digital pathology full-slice image include: reading the center coordinates and diameter of the spatial transcriptome locus; establishing a coordinate transformation function from the spatial transcriptome reference image to the digital pathology full-slice image based on tissue contours, vascular structures, glandular structures, necrotic zone boundaries, or artificial markers; inputting the center coordinates of the locus into the coordinate transformation function to obtain the coordinates of the digital pathology full-slice image; and constructing a circular, elliptical, or polygonal analysis region with the transformed locus center as the center and the radius obtained by multiplying the locus diameter by a soft boundary coefficient. A quality screening is performed on the corresponding analysis region. The quality screening includes at least tissue coverage calculation, boundary confounding calculation, and registration confidence calculation. When the tissue coverage is lower than a preset coverage threshold, the boundary confounding is higher than a preset confounding threshold, or the registration confidence is lower than a preset registration threshold, the corresponding analysis region is downweighted or removed.

4. The method according to claim 1, characterized in that, The niche pathological morphology vector is generated by a cascade of cellular-level statistical vectors, tissue structure vectors, microenvironment composition vectors, and spatial topological vectors. The cellular-level statistical vectors include nuclear area, nuclear perimeter, nuclear roundness, nuclear aspect ratio, chromatin texture, and the mean, standard deviation, quantiles, or hotspot proportions of mitosis-related counts. The tissue structure vectors include gland formation degree, tumor nest size distribution, necrosis area proportion, and infiltration boundary complexity. The microenvironment composition vectors include tumor-infiltrating lymphocyte density, stroma proportion, fibroblast enrichment degree, and vascular proximity. The spatial topological vectors include nearest neighbor distances between different cell types, local aggregation indices, cell adjacency map degree centrality, and local clustering coefficients. The niche molecular mechanism vector is generated by a cascade of drug resistance-related pathway activity, cell composition ratio, and local cell interaction intensity; wherein, the drug resistance-related pathway includes at least one of the following: DNA damage repair pathway, drug efflux pathway, epithelial-mesenchymal transition pathway, cell cycle pathway, apoptosis escape pathway, and immunosuppression pathway; the cell composition ratio includes at least one of tumor cells, T cells, macrophages, fibroblasts, and endothelial cells; and the local cell interaction intensity includes the interaction intensity between tumor cells and fibroblasts, the interaction intensity between tumor cells and immune cells, or the interaction intensity of blood vessels.

5. The method according to claim 1, characterized in that, The steps for generating morphological and molecular prototypes include: performing clustering on the niche pathological morphology vectors under the same treatment scenario encoding to obtain the morphological prototypes and their morphological centroids; performing clustering on the niche molecular mechanism vectors to obtain the molecular prototypes and their molecular centroids; and labeling the morphological and molecular prototypes with response directions, resistance directions, or intermediate directions according to the efficacy outcome labels of the cases, the regional molecular drug resistance index, and the sample distribution within the prototypes. The steps of binding morphological prototypes and molecular prototypes to generate a pathological response anchor library include: counting the number of regions that simultaneously fall into a certain morphological prototype and a certain molecular prototype, calculating the prototype co-occurrence intensity between the morphological prototype and the molecular prototype; when the prototype co-occurrence intensity is not lower than a preset co-occurrence threshold, the therapeutic directions of the two are consistent, and the treatment scenario codes of the two are consistent, binding the morphological prototype and the molecular prototype as pathological response anchors.

6. The method according to claim 1, characterized in that, Each pathological response anchor in the pathological response anchor library includes at least a treatment scenario code, morphological centroid, molecular centroid, therapeutic direction, evidence level, set of representative pathological region coordinates, set of mechanism pathway labels, and set of historical case indexes; the evidence level is calculated based on the internal resampling stability, consistency of external validation direction, and support scores from literature or pharmacogenomics evidence.

7. The method according to claim 1, characterized in that, The steps for generating patient-level anchor point hit strength include: performing tissue region segmentation, tumor region identification, cell nucleus segmentation, cell classification, and niche pathological morphology vector extraction on the target patient's digital pathology full-screen image, consistent with the training phase; calculating the distance between the pathological morphology vector of each niche of the target patient and the centroid of each pathological response anchor point morphology; converting the distance into niche anchor point similarity; and taking the average value of the first preset proportion of regions after sorting the similarity of each anchor point as the patient-level anchor point hit strength.

8. The method according to claim 1, characterized in that, The steps for outputting chemotherapy response probabilities include: reading the anchor weight table corresponding to the candidate treatment plan, inputting the patient-level anchor hit strength, baseline clinical information and treatment scenario encoding into the response scoring function to generate the response probability of the candidate treatment plan; and outputting the response reference ranking of the candidate treatment plans based on the response probability, drug resistance risk score and model confidence. The potential drug resistance mechanism indication is generated by the set of mechanism pathway tags corresponding to the hit anchor point; for any mechanism pathway, the hit strength, evidence level and efficacy direction of the hit anchor points carrying the mechanism pathway tag are summarized to obtain the indication score of the mechanism pathway, and based on this, at least one mechanism indication is output, including enhanced drug efflux, enhanced epithelial-mesenchymal transition, enhanced DNA damage repair, immunosuppressive microenvironment, fibroblast-related drug resistance or vascular-related remodeling.

9. The method according to claim 1, characterized in that, It also includes a similar case retrieval step, which includes: recalling candidate cases in the patient-level morphological representation space; re-ranking the candidate cases according to anchor hit spectrum similarity, treatment scenario consistency, pathological concept consistency and baseline clinical variable similarity; outputting the re-ranked set of similar cases and its efficacy outcome summary, and using the set of similar cases to calibrate the response probability of the target patient; When the model confidence level is lower than the preset confidence threshold, the target patient anchor point hit distribution is too scattered, or the target patient sample quality is lower than the preset quality threshold, supplementary review prompt information is output; the supplementary review prompt information includes at least one of immunohistochemistry, fluorescence in situ hybridization, circulating tumor DNA detection, short-cycle image review, or pathological slide rescan.

10. The method according to claim 1, characterized in that, It also includes a pathological response anchor database update step: continuously receiving real-world efficacy outcomes, pathological re-evaluation results, imaging responses, and available molecular verification results; performing distribution drift monitoring, inter-center correction, one-center validation, and anchor evidence level recalculation for new cases; and adding new cases to the anchor database update queue when the distribution difference between new cases and the original anchor database is within the allowable range and the external validation results meet the preset conditions.

Images