Method, device and medium for predicting risk of brain metastasis of limited-stage small cell lung cancer based on histopathology image

Abstract

The application provides a kind of based on histopathology image limited stage small cell lung cancer brain metastasis risk prediction method, device, equipment and medium, it is related to medical image processing technical field.The method is carried out image block division and tissue area screening to pathological whole section image, image block morphological features are extracted using pre-training pathology representation model, feature sequence is constructed according to image block spatial coordinates, after linear projection, spatial position coding, S4D long-range dependence modeling, residual fusion and multi-instance aggregation, brain metastasis risk prediction probability is output;Further, the integral gradient is used to calculate the attribution score of each image block to the target risk output, unsupervised clustering is carried out based on morphological features, and whole section risk contribution heat map and histopathology cluster contribution spectrum are generated by combining image block coordinates, and high-risk driving area, representative tissue morphology and its quantitative contribution are output.The application can realize efficient risk stratification and multi-level interpretation that can be verified by pathologists.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, device, equipment, and medium for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images. Background Technology

[0002] Small cell lung cancer (SCLC) is a highly aggressive neuroendocrine malignancy, accounting for approximately 15% of all lung cancer cases. SCLC is characterized by rapid growth, early metastasis, and poor prognosis. Although initial response rates to platinum-based chemotherapy and radiotherapy are high, disease recurrence is very common, with the brain being one of the most clinically impactful sites of metastasis. Brain metastases (BM) significantly worsen survival and lead to cognitive decline and functional impairment. Brain metastases are often diagnosed only at the time of recurrence, at which point treatment options are limited and the prognosis is extremely poor. Therefore, effective prevention of brain metastases and early risk stratification remain crucial for improving the long-term management of SCLC.

[0003] Historically, prophylactic brain irradiation (PCI) has been used as an intracranial prophylaxis strategy for patients with limited-stage small cell lung cancer (LS-SCLC) who respond to initial therapy. Meta-analytical evidence suggests that PCI can reduce the cumulative incidence of brain metastases from approximately 58.6% to 33.3%, and correspondingly improve the three-year survival rate from 15.3% to 20.7%. Another analysis pooling 28 retrospective studies showed a significant benefit of PCI compared to no PCI, with a median survival of 27.8 months in the PCI group and 18.8 months in the no-PCI group. However, PCI is also associated with clinically relevant toxicities, including cognitive impairment and fatigue, which complicates its benefit-risk profile.

[0004] In the era of modern MRI monitoring, the role of PCI is becoming increasingly individualized, especially considering issues of neurotoxicity, the heterogeneity of baseline brain metastasis risk, and the continuous development of monitoring strategies. Current PCI selection is still largely based on clinicopathological factors and doctor-patient discussions, rather than on individualized predictions of future brain metastasis risk. Current imaging modalities, such as magnetic resonance imaging (MRI) and computed tomography (CT), are primarily used to detect existing intracranial disease, rather than proactively assessing future brain metastasis risk. Therefore, in LS-SCLC, more accurate and clinically practical methods are needed to stratify brain metastasis risk and inform individualized PCI risk discussions.

[0005] In recent years, deep learning-based whole-slice image analysis methods have made significant progress in the fields of tumor diagnosis and prognosis prediction. However, existing deep learning prediction methods based on whole-slice images mostly adopt a multi-instance learning framework, treating the whole-slice image as an unordered collection of independent image patches, and aggregating the features of each image patch through attention mechanisms or pooling strategies to perform slice-level prediction. The following shortcomings exist in this type of method: First, the multi-instance learning framework ignores the spatial relationships between image patches in whole-slice images and the histological continuity between adjacent pathological regions, resulting in the loss of long-distance spatial correlation information between tumor cell-dense areas, stromal activation areas, immune infiltration areas, and necrotic areas, thus limiting the model's ability to model the overall spatial heterogeneity of tumors. Second, although the Transformer architecture used in existing methods can capture global contextual information, the computational complexity of its attention mechanism increases quadratically with the number of image patches, making it difficult to efficiently process whole-slice image sequences containing thousands to tens of thousands of effective image patches. In modeling ultra-long sequences, there are problems of excessive GPU memory consumption and attention dilution. Third, none of the existing methods are designed for the specific clinical scenario of predicting the risk of brain metastasis in limited-stage small cell lung cancer. They lack the technical means to effectively extract the histological morphological features related to the occurrence of brain metastasis from whole-slice images, and cannot provide a reliable risk stratification basis for individualized PCI decisions. Fourth, existing methods typically only provide slice-level risk probabilities or coarse-grained heatmaps based on attention weights, lacking an interpretive link that can simultaneously connect image patch contribution, tissue morphology category, and its spatial location in the whole slice. This makes it difficult to convert the model output into histological evidence that can be verified by pathologists, and also makes it difficult to quantitatively compare the positive or negative contributions of different tissue microenvironment components to the prediction of brain metastasis risk. Summary of the Invention

[0006] This invention provides a method, device, equipment, and medium for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images. The method involves preprocessing whole-slice images and selecting effective image blocks, then extracting high-dimensional morphological features of each image block using a pre-trained pathological characterization model. Based on the two-dimensional spatial coordinates of the image blocks in the whole slice, a spatially continuous image block feature sequence is constructed according to preset spatial scanning rules, effectively preserving the spatial proximity and pathological evolution relationships between tissue regions. Furthermore, a linear projection layer maps each feature vector to a unified hidden feature space, fusing spatial location encoding and tissue cluster encoding to enhance location awareness. Subsequently, a diagonal structured state space model, namely the S4D module, is introduced to model the long-range dependencies of the image block feature sequence through state space equations, efficiently capturing the collaborative relationships between distant pathological regions using its linear computational complexity. Finally, residual fusion, multi-instance aggregation, and a classification predictor output the predicted probability of brain metastasis risk, providing an objective and reliable quantitative basis for assessing the risk of brain metastasis and making individualized PCI decisions for patients with limited-stage small cell lung cancer.

[0007] Furthermore, this invention establishes a multi-level interpretability analysis chain outside the risk prediction model: taking the target output of the brain metastasis risk prediction model as the interpretation object, the signed feature attribution value of each image patch is calculated through the interpolation path from the baseline feature sequence to the input feature sequence using integral gradients; unsupervised clustering is performed based on the high-dimensional morphological features of the image patches to form histopathological clusters with similar tissue morphology; the image patch attribution values, histopathological cluster labels, and two-dimensional spatial coordinates are jointly mapped to generate a whole-slice risk contribution heatmap, a histopathological cluster contribution spectrum, and representative tissue regions. This transforms a single risk probability into an interpretation result combining image patch-level spatial localization, tissue cluster-level morphological semantics, and whole-slice-level quantitative summarization, enabling the model's prediction basis to be reviewed by pathologists.

[0008] In a first aspect, the present invention provides a method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, the method comprising: Acquire stained whole-slice images of pathological sections from patients with limited-stage small cell lung cancer, and perform image block division and tissue region detection on the whole-slice images. Based on the detection results, invalid image blocks are removed to obtain a set of valid image blocks consisting of multiple valid image blocks. Each valid image block in the set of valid image blocks is input into a pre-trained pathological characterization model, and the high-dimensional morphological feature vector of each valid image block is extracted to obtain the feature vector set corresponding to the whole slice image. Based on the two-dimensional spatial coordinates of each valid image block in the full slice image, the high-dimensional morphological feature vectors in the feature vector set are sorted according to a preset spatial scanning rule to construct an image block feature sequence with spatial continuity. Each high-dimensional morphological feature vector in the image block feature sequence is sequentially input into a linear projection layer for linear projection, and the linearly projected features are fused with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features; The enhanced sequence features are input into the S4D module, and long-range dependency modeling is performed through state-space equations to obtain the output representation corresponding to each image block output by the S4D module. The output representation of each image block output by the S4D module is residually fused and normalized with the corresponding linearly projected feature input residual fusion layer to obtain the long-range dependent enhanced image block representation. All long-range dependency-enhanced image patch representations are aggregated into full-slice-level representations, and these full-slice-level representations are input into a classification predictor to obtain the brain metastasis risk prediction probability.

[0009] In one possible design, the preset spatial scanning rules include row-first scanning, Z-order curves, Hilbert curves, or local neighborhood scanning based on the organization's connected regions.

[0010] In one possible design, each high-dimensional morphological feature vector in the image patch feature sequence is sequentially input into a linear projection layer for linear projection. The linearly projected features are then fused with the corresponding spatial location code and tissue cluster code to obtain the location-enhanced sequence features, specifically including: According to the following formula, the first... A high-dimensional morphological feature vector is linearly projected and normalized to map it to a unified hidden feature space, resulting in projected features: in, For the first High-dimensional morphological feature vectors of image patches and These are the learnable projection matrix and the bias term, respectively. For the first Projection features of an image patch; The projection features are fused with the corresponding spatial location code and tissue cluster code according to the following formula to obtain the location-enhanced sequence features: in, For the first Spatial location encoding of image blocks Encoding for organizational clusters, For the first The enhanced sequence features of the location of each image patch.

[0011] In one possible design, the location-enhanced sequence features are input into the S4D module, and long-range dependency modeling is performed using state-space equations to obtain the output representation corresponding to each image patch output by the S4D module, specifically including: Based on the S4D module, the input sequence is modeled using the following state-space equation: in, Indicates the first Enhanced sequence features of the location of each image patch Indicates the current hidden state. Indicates the previous hidden state. Indicates the output of the S4D module. Output representation of each image patch , , , The learnable parameter matrix; The learnable parameter matrix In diagonal form, the recursive state update is transformed into a parallelizable one-dimensional convolution using the Fast Fourier Transform.

[0012] In one possible design, the long-range dependency-enhanced image patch representation is determined by the following formula: in, The first output of the S4D module Output representation of each image patch For the first Enhanced sequence features of the location of each image patch For the first Long-range dependency enhancement of image patch representations for each image patch.

[0013] In one possible design, all long-range dependency-enhanced image patch representations are aggregated into a full-slice-level representation, and this full-slice-level representation is input into a classification predictor to obtain the brain metastasis risk prediction probability, specifically including: Max pooling is performed on all long-range dependency-enhanced image patch representations according to the following formula to obtain full-slice-level representations: in, N The number of valid image patches, , and For the 1st, 2nd and 3rd N Long-range dependency augmented image patch representation of each image patch Full-slice level characterization; The full-slice-level representation is input into the classification predictor according to the following formula, and after processing by the Sigmoid function, the brain metastasis risk prediction probability is obtained: in, and For learnable classification parameters, For the Sigmoid function, To predict the probability of brain metastasis risk.

[0014] In one possible design, the method further includes constructing a brain metastasis risk prediction model; The brain metastasis risk prediction model includes a linear projection layer, an S4D module, a residual fusion layer, and a classification predictor connected in sequence. Using whole-slice pathological images of patients with historical limited-stage small cell lung cancer, after preprocessing and extraction by a pre-trained pathological characterization model, image patch feature sequences were obtained as training samples. The presence or absence of radiologically confirmed brain metastases was used as the supervision label. A binary classification cross-entropy loss function was constructed according to the following formula. : in, This serves as a true label for whether radiologically confirmed brain metastases have occurred. The probability of brain metastasis risk predicted by the brain metastasis risk prediction model; The learnable parameters in the linear projection layer, the S4D module, the residual fusion layer, and the classification predictor are jointly optimized end-to-end through backpropagation to minimize the loss function until the model converges, thus completing the training of the brain metastasis risk prediction model. The image block feature sequence obtained by preprocessing and extracting from the pathological whole slice image of the patient with limited-stage small cell lung cancer to be predicted is input into the trained brain metastasis risk prediction model, and the corresponding brain metastasis risk prediction probability is output.

[0015] In one possible design, after obtaining the predicted probability of brain metastasis risk, the method further includes performing a histopathological interpretability analysis on the risk prediction results. This interpretability analysis does not change the parameters of the trained brain metastasis risk prediction model, but rather constructs an interpretable processing chain based on the model input features, target risk output, and image patch spatial coordinates.

[0016] Specifically, the input feature sequence consisting of N valid image patch features is denoted as X, the baseline feature sequence with the same dimension as X is denoted as X′, and the output function of the brain metastasis risk prediction model for the target category is denoted as F. Preferably, the baseline feature sequence is a zero vector sequence, or a reference feature sequence determined based on low-risk reference samples.

[0017] in, This represents the integral gradient attribution value corresponding to the j-th dimension feature of the i-th valid image patch. and These represent the input features and baseline features, respectively, and α represents the interpolation coefficients. In actual calculations, multiple discrete interpolation step sizes are set between 0 and 1, and the gradient at each interpolation point is numerically integrated.

[0018] in, Represents the signed attribution score of the i-th valid image patch; A positive value indicates that the image patch contributes to outputs with a high risk of brain metastasis. A negative value indicates that the image patch contributes supportively to the low-risk output. The orientation-indiscriminate contribution strength can also be determined by the L1 or L2 norm of each dimension's attribution value.

[0019] Furthermore, nearest neighbor relationships are constructed based on the high-dimensional morphological feature vectors of each effective image patch, and an unsupervised clustering method based on the K-nearest neighbor graph is used to divide the effective image patches into K histopathological clusters. The value of K can be determined by a combination criterion that increases the Calinski-Harabasz exponent and decreases the Davies-Bouldin exponent; in one specific embodiment, K is 12.

[0020] in, This indicates the histopathological cluster to which the i-th valid image patch belongs. This represents the average attribution contribution of the k-th histopathological cluster. This indicates the percentage of the image patch containing the histopathological cluster in the current whole-slice image. This is an indicator function. Furthermore, high-risk contribution spectra and low-risk support spectra can be constructed based on the sum of positive attribution values, the sum of negative attribution values, and their normalization ratios, respectively.

[0021] Based on the two-dimensional spatial coordinates of each valid image patch in the whole pathological slide image, the attribution score is... and histopathology cluster tags The original whole-slice image is mapped back to generate a risk contribution heatmap, a spatial distribution map of histopathological clusters, and an overlay of the two. The image blocks and histopathological clusters are sorted according to their attribution contribution, and representative image blocks are selected for pathological review.

[0022] The interpretability results include at least the predicted probability of brain metastasis risk, risk level, histopathological clusters with the highest positive contribution, histopathological clusters with the lowest negative contribution, the proportion of image blocks for each cluster, representative tissue regions, and their spatial coordinates. Preferably, the histopathological clusters can also be annotated by pathologists as tissue morphology types such as high tumor cell density, nuclear crowding and pleomorphism, active mitosis, interstitial activation, immune cell infiltration, necrosis, alveolar structures, or bronchial epithelium.

[0023] In implementations with proteomics or other molecular omics data, a histopathological cluster proportion matrix can be further constructed on a patient-by-patient basis, and the correlation between this proportion matrix and functional pathway scores, tumor microenvironment characteristic scores, or molecular expression characteristics can be calculated to verify the biological consistency of risk-related tissue morphology. The molecular omics analysis is used for external biological interpretation of the prediction results and is not a necessary input to the brain metastasis risk prediction model.

[0024] In one possible design, after obtaining the predicted probability of brain metastasis risk, the method further includes: Using the target output corresponding to the predicted probability of brain metastasis risk as the output to be explained, a baseline feature sequence with the same dimension as the image patch feature sequence is set, and multiple interpolation feature sequences are generated between the baseline feature sequence and the input image patch feature sequence. The gradient of the target output relative to the features of each image block in each interpolation feature sequence is calculated, and the gradient is integrated along the interpolation path to obtain the feature attribution value of each valid image block. Based on the feature attribution values ​​of each valid image patch and its two-dimensional spatial coordinates in the whole-slice image, a whole-slice risk contribution heatmap is generated, representing the contribution of different tissue regions to the prediction of brain metastasis risk.

[0025] In one possible design, the method further includes: performing unsupervised clustering based on the high-dimensional morphological feature vectors of each valid image patch to form multiple histopathological clusters; mapping the feature attribution value of each valid image patch to its respective histopathological cluster, and calculating the attribution contribution and image patch proportion of each histopathological cluster; wherein the number of histopathological clusters is determined according to at least one of the Calinski-Harabasz index and the Davies-Bouldin index.

[0026] In one possible design, interpretable results are generated based on the whole-slice risk contribution heatmap, the attribution contribution of each histopathological cluster, and the proportion of image blocks. The interpretable results include the predicted probability of brain metastasis risk, the histopathological cluster with the highest positive contribution and / or the lowest negative contribution, representative image blocks of each histopathological cluster, spatial distribution area, and quantitative contribution information.

[0027] Secondly, the present invention provides a device for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, the device comprising: The region detection module is configured to acquire stained pathological whole-slice images of patients with limited-stage small cell lung cancer, and to perform image block division and tissue region detection on the whole-slice images. Invalid image blocks are removed based on the detection results to obtain a set of valid image blocks consisting of multiple valid image blocks. The feature extraction module is configured to input each valid image block in the set of valid image blocks into a pre-trained pathological characterization model, extract the high-dimensional morphological feature vector of each valid image block, and obtain the feature vector set corresponding to the whole slice image. The feature sequence construction module is configured to sort the high-dimensional morphological feature vectors in the feature vector set according to a preset spatial scanning rule based on the two-dimensional spatial coordinates of each valid image block in the full slice image, thereby constructing an image block feature sequence with spatial continuity. The sequence feature enhancement module is configured to sequentially input each high-dimensional morphological feature vector in the image block feature sequence into a linear projection layer for linear projection, and then fuse the linearly projected features with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features. The long-range dependency modeling module is configured to input the location-enhanced sequence features into the S4D module, perform long-range dependency modeling through the state-space equation, and obtain the output representation corresponding to each image block output by the S4D module. The dependency enhancement module is configured to perform residual fusion and normalization on the output representation corresponding to each image block output by the S4D module and the corresponding linearly projected feature input residual fusion layer to obtain the image block representation after long-range dependency enhancement. The risk prediction module is configured to aggregate all long-range dependency-enhanced image patch representations into a full-slice-level representation, and input the full-slice-level representation into a classification predictor to obtain the brain metastasis risk prediction probability. The risk prediction probability can be further converted into different risk levels, including low-risk, medium-risk, and high-risk levels, based on a preset threshold or model calibration results.

[0028] In one possible design, the device further includes a feature attribution module, a histopathological clustering module, a spatial mapping module, and an interpretation result output module; The feature attribution module is configured to take the target output corresponding to the brain metastasis risk prediction probability as the output to be explained, and calculate the feature attribution value of each effective image block by integral gradient. The histopathology clustering module is configured to perform unsupervised clustering based on the high-dimensional morphological feature vectors of each effective image patch to form multiple histopathology clusters, and calculate the attribution contribution and image patch proportion of each histopathology cluster. The spatial mapping module is configured to map the feature attribution value and histopathological cluster label to the whole pathological slice image based on the two-dimensional spatial coordinates of each valid image block, thereby generating a whole slice risk contribution heatmap and a spatial distribution map of histopathological clusters. The interpretation result output module is configured to output high-risk driving regions, low-risk supporting regions, representative histopathological clusters, representative image blocks, and their quantitative contribution information.

[0029] In one possible design, the device further includes: The feature attribution module is configured to take the target output corresponding to the brain metastasis risk prediction probability as the output to be explained, and calculate the feature attribution value of each valid image block through integral gradient. The histopathology clustering module is configured to perform unsupervised clustering based on the high-dimensional morphological feature vectors of each effective image patch to form multiple histopathology clusters, and calculate the attribution contribution and image patch proportion of each histopathology cluster. The spatial mapping module is configured to map the feature attribution value and histopathological cluster label to the whole pathological slide image based on the two-dimensional spatial coordinates of each valid image block, thereby generating a whole slide risk contribution heatmap and a spatial distribution map of the histopathological cluster. The interpretation output module is configured to output interpretable results including high-risk driving regions, low-risk supporting regions, representative histopathological clusters, and their quantitative contribution information.

[0030] Thirdly, embodiments of the present invention provide an electronic device, comprising: at least one processor and a memory; the memory storing computer-executable instructions; the at least one processor executing the computer-executable instructions stored in the memory, such that the at least one processor performs the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images as described in the first aspect and various possible designs of the first aspect.

[0031] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images as described in the first aspect and various possible designs of the first aspect.

[0032] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images as described in the first aspect and various possible designs of the first aspect.

[0033] In one embodiment, the present invention also provides a system for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images. The system can be deployed on a server, cloud computing platform, or internal computing environment of a medical institution, and includes an image input module, a risk prediction module, an interpretability analysis module, and a result output module. Users can upload digitized whole-slice images of pathology slides via a client or a medical institution. After receiving the images, the system performs image preprocessing, feature extraction, risk prediction, and interpretability analysis, and outputs analysis results including the predicted probability of brain metastasis, risk level, risk contribution region, and histopathological interpretation information.

[0034] The method, apparatus, device, and medium for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images provided by this invention have at least the following beneficial effects: 1. This invention provides a method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images. It can estimate the future risk of brain metastasis in patients with limited-stage small cell lung cancer from conventionally acquired H&E-stained whole-slice images. This method preprocesses the whole-slice images and filters effective image blocks. It extracts high-dimensional morphological features of each image block using a pre-trained pathological characterization model and constructs a spatially continuous image block feature sequence based on the two-dimensional spatial coordinates of the image blocks in the whole slice according to preset spatial scanning rules. This effectively preserves the spatial proximity and pathological evolution relationships between tissue regions, overcoming the shortcomings of traditional multi-instance learning frameworks that treat the whole slice as a disordered set of image blocks, leading to the loss of spatial structural information. This allows the model to learn long-distance spatial relationships between tumor cell-dense areas, stromal activation areas, immune infiltration areas, and necrotic areas.

[0035] 2. This invention introduces a diagonal structured state-space model, namely the S4D module, which performs long-range dependency modeling on image patch feature sequences through state-space equations. The S4D module models long sequences through state-space equations, and its state matrix A is in diagonal form. It also uses Fast Fourier Transform to transform recursive state updates into parallel one-dimensional convolutions, which can efficiently process full-slice image sequences containing thousands to tens of thousands of effective image patches with linear computational complexity. This overcomes the problems of excessive memory consumption and attention dilution caused by the quadratic increase in computational complexity with the number of image patches in the traditional Transformer architecture for modeling ultra-long sequences.

[0036] 3. This invention maps each high-dimensional morphological feature vector to a unified hidden feature space through a linear projection layer, and fuses the projected features with spatial location encoding and tissue cluster encoding to obtain position-enhanced sequence features. This enables the model to not only receive the morphological features of image patches, but also perceive their spatial distribution in the whole slice, enhancing the model's ability to model the relationship between local pathological morphology, global tissue structure and clinical risk. At the same time, by performing residual fusion and normalization between the output representation of the S4D module and the linearly projected features, it is possible to introduce contextual state information within the whole slice while preserving the original local morphological information of the image patch, avoiding the model focusing only on local high-risk areas and ignoring the global tissue structure.

[0037] 4. This invention establishes a multi-level interpretability mechanism that connects image patch level, histopathological cluster level, and whole-slice level. By calculating the image patch attribution value through integral gradient along the continuous path from baseline features to true input features, it can quantitatively distinguish tissue regions that promote high-risk output and tissue regions that support low-risk output. By performing unsupervised clustering on high-dimensional morphological features, image patches with similar histological patterns are grouped into stable histopathological clusters. Then, by combining the two-dimensional spatial coordinates of the image patches, a risk contribution heatmap, a cluster spatial distribution map, and a cluster contribution spectrum are generated, expanding the model output from a single risk probability to histological evidence that can be located, summarized, and reviewed by pathologists.

[0038] 5. The interpretability analysis of this invention can identify mesenchymal activation regions associated with high-risk prediction, such as high tumor cell density, nuclear crowding and pleomorphism, active mitosis, and enrichment of fibroblasts and extracellular matrix, and can identify regions associated with low-risk prediction, such as relatively preserved lung tissue structure, immune cell infiltration, and necrosis. At the same time, through the correlation analysis of histopathological clusters with molecular pathways or tumor microenvironment characteristics, it can provide independent biological consistency evidence for model risk stratification without requiring molecular omics data as a necessary input to the predictive model.

[0039] 6. This invention constructs a binary cross-entropy loss function by using radiologically confirmed brain metastases as a supervision label, and performs end-to-end joint optimization of learnable parameters in the linear projection layer, S4D module, residual fusion layer, and classification predictor. This enables the model to jointly learn whole-slice morphological patterns related to the risk of brain metastases, and demonstrates robust discrimination performance in multiple independent cohorts, exhibiting excellent generalization ability and clinical applicability. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0041] Figure 1The flowchart illustrates a method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, as provided in this embodiment of the invention.

[0042] Figure 2 A schematic diagram of the BraMARS deep learning architecture provided in this embodiment of the invention is shown, wherein (A) a flowchart illustrates patient enrollment, inclusion / exclusion criteria from two independent medical centers (CHCAMS and HMUCH), and cohort allocation for model development, validation, and clinical utility assessment; (B) a schematic diagram of the BraMARS deep learning model architecture, including three main components: WSI preprocessing and plaque-level feature extraction, sequence feature encoding using S4, and MIL aggregation for final prediction; (C) a workflow for model evaluation and interpretability; and (D) the potential clinical impact of BraMARS, demonstrating its integration into the process of personalized PCI recommendation clinical decision-making and the expected improvement in patient outcomes; TP, true positive; TN, true negative; WSI, whole slice image; S4, structured state space of the sequence; MIL, multi-instance learning.

[0043] Figure 3 Performance graphs of BraMARS in predicting brain metastasis risk provided in embodiments of the present invention; wherein, (A) Receiver operating characteristic curves, showing the performance of BraMARS in internal and external validation cohorts. (B) Distribution of BraMARS risk scores between patients with brain metastases (BM+) and patients without metastases (BM-). (C) Confusion matrix, showing the classification performance of BraMARS at the optimal threshold in internal and external validation sets. (D) Patient distribution ranked by BraMARS risk score, demonstrating the model's ability to stratify patients across the entire risk spectrum. (E) Decision curve analysis, comparing the net benefit of the BraMARS-guided strategy with "all treatment" and "no treatment" approaches within clinically relevant threshold probability ranges. (F) Calibration plot, showing the consistency between predicted probabilities and observed results; perfect calibration is represented by a diagonal dashed line.

[0044] Figure 4 The following are prognostic value maps of BraMARS risk stratification provided in this embodiment of the invention: (AB) Kaplan-Meier curves of disease-free survival stratified by BraMARS risk group in the CHCAMS (A) and HMUCH (B) cohorts; (C) an integrated framework combining the BraMARS risk score with the AJCC staging system to create an improved three-category risk classification; and (DE) disease-free survival analysis of the combined three-category risk classification in the CHCAMS (D) and HMUCH (E) cohorts, showing enhanced prognostic stratification.

[0045] Figure 5This is a schematic diagram illustrating the clinical applicability of BraMARS-guided PCI decision-making in an embodiment of the present invention; wherein, (A) a simulation analysis of BraMARS implementation shows that it may reduce unnecessary PCI treatment and improve coverage of high-risk patients. (B) Kaplan-Meier curves of disease-free survival stratified by BraMARS risk group and PCI treatment status. (C) a proposed clinical workflow for BraMARS-assisted PCI decision-making in the management of LS-SCLC. (D) a comparison of disease-free survival between patients managed according to BraMARS recommendations and patients who deviate from BraMARS-guided treatment.

[0046] Figure 6 This is a schematic diagram illustrating the multi-level histopathological interpretability of BraMARS prediction provided in this embodiment of the invention. (A) is a UMAP visualization of 12 histopathological clusters formed based on deep learning morphological features; (B) is a cluster contribution heatmap formed by combining image patch integral gradient attribution values ​​and histopathological cluster labels, used to show the direction and intensity of different histopathological clusters' predictions for high or low risk; (C) is a representative whole-slice image overlaid with histopathological cluster labels and risk contribution information, used to demonstrate the spatial distribution of high-risk driving regions and low-risk supporting regions, and the corresponding clinical outcomes.

[0047] Figure 7 This diagram illustrates the biological basis of BraMARS risk stratification provided in this embodiment of the invention. (A) Volcano plot and unsupervised clustering heatmap show differentially expressed proteins between the high-risk and low-risk BraMARS groups. (B) Functional enrichment analysis of the KEGG pathway. (C) Functional network of gene ontology biological processes. (D) Correlation heatmap between histopathological clusters and functional gene expression tags, revealing associations with tumor microenvironment components. This diagram is used to validate the external consistency of the interpretation results obtained from image attribution and histopathological clustering at the molecular level; molecular omics information is not a necessary input to the risk prediction model.

[0048] Figure 8 This is a structural diagram of a device for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, provided in an embodiment of the present invention.

[0049] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0051] In the technical solution of this invention, the collection, storage, use, processing, transmission, provision and disclosure of relevant data and information comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0052] It should be noted that in the embodiments of the present invention, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of the present invention. However, they do not mean that the inventor has used or necessarily used the solution.

[0053] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0054] Example 1: This invention provides a method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images. The method first uses a pre-trained pathological characterization model to extract high-dimensional morphological features of large-scale image patches from a whole-slice. Then, it introduces an S4D structured state space module to perform long-range dependency modeling on the feature sequences of image patches within the whole-slice, thereby overcoming the limitation of traditional multi-instance learning methods that mainly rely on local discriminative regions. Figure 1 As shown, the specific implementation steps are as follows: S10-S70.

[0055] S10: Obtain stained whole-slice images of patients with limited-stage small cell lung cancer, and perform image block division and tissue region detection on the whole-slice images. Based on the detection results, invalid image blocks are removed to obtain a set of valid image blocks consisting of multiple valid image blocks.

[0056] For example, this embodiment acquires H&E-stained whole-slice images of pathological tissue sections and performs digital scanning at 20× magnification. Each whole-slice image is divided into multiple non-overlapping image blocks according to a preset physical resolution, with each image block preferably measuring 224×224 pixels. Subsequently, a tissue region detection method based on Otsu thresholding is used to automatically remove background, blank areas, folding artifacts, and non-tissue areas, retaining only valid image blocks with a tissue area ratio of not less than 50%. Through this step, a high-resolution whole-slice image can be converted into a set of pathological image blocks consisting of thousands to tens of thousands of valid image blocks.

[0057] S20: Input each valid image block in the set of valid image blocks into the pre-trained pathological characterization model, extract the high-dimensional morphological feature vector of each valid image block, and obtain the feature vector set corresponding to the whole slice image.

[0058] For example, each valid image patch is input into a pre-trained pathological image representation model, preferably using the CTransPath model pre-trained on a large-scale public pathological dataset. This model combines the local texture extraction capability of convolutional neural networks with the contextual modeling capability of Transformers, enabling the extraction of stable features from multi-level pathological morphology such as cell density, nuclear atypia, interstitial reaction, necrosis, immune infiltration, and tissue structure. For the first... For each image patch, the model outputs a high-dimensional morphological feature vector: in, This represents the 768-dimensional pathological morphological characterization of the image patch. Therefore, a whole-slice image can be represented as: in, This represents the number of valid image blocks in the full slice.

[0059] S30: Based on the two-dimensional spatial coordinates of each valid image block in the full slice image, sort the high-dimensional morphological feature vectors in the feature vector set according to the preset spatial scanning rules to construct an image block feature sequence with spatial continuity.

[0060] Unlike traditional multi-instance learning methods that treat the entire slice as an unordered set of image patches, this embodiment further utilizes the two-dimensional spatial coordinates of image patches within the original entire slice to organize high-dimensional image patch features into a spatially continuous sequence. Specifically, the horizontal and vertical coordinates of each image patch within the entire slice are recorded. The image blocks are sorted according to a preset spatial scanning rule, which may include row-first scanning, Z-order curves, Hilbert curves, or local neighborhood scanning based on tissue connectivity regions.

[0061] For example, a sequence strategy that maintains spatial proximity is employed, ensuring that adjacent pathological regions remain close together in the feature sequence, thus avoiding the loss of spatial structural information in traditional random bag modeling. For whole slices with discontinuous tissue regions or an excessive number of image patches, the whole slice can be further divided into multiple tissue sub-regions, with each sub-region forming a local sequence. Cross-regional information fusion is then performed through a global state space module. This transforms the whole slice from simply treating image patches as independent instances into a long sequence carrying spatial order and pathological evolution relationships, enabling subsequent models to learn long-distance associations between tumor cell-dense areas, stromal activation areas, immune infiltration areas, and necrotic areas.

[0062] S40: Each high-dimensional morphological feature vector in the image patch feature sequence is sequentially input into the linear projection layer for linear projection, and the linearly projected features are fused with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features.

[0063] For example, before inputting into the S4D module, the feature vector of each image patch is linearly projected and normalized to map it to a unified hidden feature space: in, For the first High-dimensional morphological feature vectors of image patches and These are the learnable projection matrix and the bias term, respectively. For the first Projection features of an image patch.

[0064] Furthermore, the spatial coordinate encoding, tissue region encoding, or pathological clustering encoding of image patches are combined with... The sequences are fused to obtain position-enhanced sequence features: in, Encoding spatial location, Encoding for optional tissue clusters, For the first The enhanced sequence features of each image patch. Through this design, the model not only receives the morphological features of the image patch, but also perceives its spatial distribution in the whole slice, thereby enhancing its ability to model the relationship between local pathological morphology, global tissue structure and clinical risk.

[0065] S50: Input the location-enhanced sequence features into the S4D module, perform long-range dependency modeling through state-space equations, and obtain the output representation corresponding to each image block output by the S4D module.

[0066] For example, the location-enhanced image patch feature sequence is input into the S4D (Diagonal Structured State Space Model) module. S4D models long sequences using state-space equations: in, Indicates the first Input features of an image patch, Indicates the current hidden state. Indicates the previous hidden state. This represents the output representation. This is a learnable parameter matrix. Preferably, the state matrix is... The design is diagonal to reduce the computational complexity of modeling long sequences. Simultaneously, the Fast Fourier Transform (FFT) is used to transform recursive state updates into parallelizable one-dimensional convolutions, making it suitable for large-scale sequence modeling of thousands to tens of thousands of image patches in a full-slice image.

[0067] In this process, the hidden states of S4D are equivalent to "pathological memory units" across image patches, continuously accumulating morphological information from preceding tissue regions and transferring it to subsequent image patches. Thus, the model can capture synergistic relationships between distant pathological regions, such as the association between high-cell-density tumor regions and distal mesenchymal activation regions, the combination patterns between immune-infiltrating regions and necrotic regions, and the combined impact of different tissue microenvironment states on the risk of brain metastasis.

[0068] S60: The output representation corresponding to each image block output by the S4D module is residually fused and normalized with the corresponding linearly projected feature input residual fusion layer to obtain the image block representation after long-range dependency enhancement.

[0069] To improve the model's ability to preserve dependencies on long-distance image patches, the HiPPO high-order polynomial projection initialization method is introduced during the S4D state matrix initialization stage. This initialization method enables the state space module to retain key information from earlier image patches during long sequence propagation, reducing the information decay, excessive memory consumption, or attention dilution problems that may occur in traditional recurrent neural networks or attention models when modeling ultra-long sequences.

[0070] Each image patch output by the S4D module represents Further residual fusion with the original input features: in, This is an image patch representation enhanced for long-range dependencies. The residual structure preserves the original local morphological information of the image patch on the one hand, and introduces contextual state information across the entire slice on the other, preventing the model from focusing only on local high-risk areas while ignoring the global organizational structure.

[0071] S70: Aggregate all long-range dependent enhanced image patch representations into a full-slice level representation, and input the full-slice level representation into the classification predictor to obtain the brain metastasis risk prediction probability.

[0072] For example, all enhanced image patches are represented Aggregation is performed at the full-slice level. In the basic implementation, a max-pooling strategy can be used: This involves extracting the most discriminative risk-related features from all image patches. A further preferred embodiment may employ a combination of max pooling and Top-k weighted aggregation. Specifically, it first selects the image patches with the highest contribution, then performs weighted fusion based on their risk weights to simultaneously retain both locally strong discriminative regions and global pathological background information. This aggregation method can reduce the accidental impact of a single anomalous image patch on the final prediction and improve the model's stability under different slice qualities and different scanning batches.

[0073] Full-slice level characterization Input a classification predictor, output the probability of the patient developing brain metastases in the future: in, For the Sigmoid function, This represents the predicted risk value for brain metastasis. Based on the optimal threshold determined using the Youden index in the training set, patients were divided into high-risk and low-risk groups for brain metastasis. This prediction result can be further used to assist in determining whether patients need enhanced brain MRI follow-up, whether to enter the individualized discussion process for prophylactic cranial irradiation PCI, or whether they need more intensive clinical management.

[0074] In some embodiments, after step S70, the method further includes steps S80-S110, performing multi-level histopathological interpretability analysis on the brain metastasis risk prediction results.

[0075] S80: Using the output of the brain metastasis risk prediction model for the target category as the output to be explained, multiple interpolated feature sequences are constructed between the baseline feature sequence and the input image block feature sequence. The output gradient corresponding to each interpolated feature sequence is calculated and integrated along the interpolation path to obtain the feature attribution value of each effective image block.

[0076] For example, a zero vector sequence is used as the baseline feature sequence, and the integration interval is discretized into multiple equally spaced interpolation steps. For the i-th valid image patch, the integral gradients of its features in each dimension are summed with a signed sum or a weighted sum to obtain the image patch attribution score a. i A positive attribution score indicates that the image patch promotes high-risk output, while a negative attribution score indicates that the image patch supports low-risk output.

[0077] S90: Construct K-nearest neighbor relationships based on the high-dimensional morphological feature vectors of each valid image block and perform unsupervised clustering to form multiple histopathological clusters; determine the number of histopathological clusters according to the Calinski-Harabasz index and the Davies-Bouldin index, and assign a histopathological cluster label to each valid image block.

[0078] In one specific implementation, all image patches are divided into 12 histopathological clusters. These cluster labels are used to categorize consecutive high-dimensional features into morphological classes with similar cellular composition, stroma state, and tissue structure. The histopathological clusters used for interpretive analysis are preferably post-processing labels after prediction, without altering the parameters of the trained risk prediction model.

[0079] S100: Map the attribution score of each valid image patch to its corresponding histopathology cluster, calculate the average attribution contribution, total positive attribution, total negative attribution, and image patch percentage of each histopathology cluster; and based on the two-dimensional spatial coordinates of the image patch, map the attribution score and cluster label back to the original pathological whole slice image to generate a risk contribution heatmap, a spatial distribution map of histopathology clusters, and an overlay map of the two.

[0080] S110: Sort tissue regions based on the attribution contribution of each histopathological cluster, extract representative image patches with the highest positive contribution and the lowest negative contribution, and output an interpretable report including the predicted probability of brain metastasis risk, risk level, main driving histopathological cluster, main supporting histopathological cluster, cluster proportion, representative image patch, and spatial coordinates.

[0081] Furthermore, pathologists can perform morphological annotation on representative histopathological clusters and, when molecular omics data is available, calculate the correlation between cluster proportions and molecular pathways or tumor microenvironment characteristics to evaluate the histological and biological consistency of the interpretation results.

[0082] In some embodiments, after step S110, the method further includes step S120, model training and parameter optimization.

[0083] Specifically, a brain metastasis risk prediction model is constructed. This model comprises a linear projection layer, an S4D module, a residual fusion layer, and a classification predictor, connected sequentially. Image block feature sequences obtained from preprocessed and pre-trained pathological characterization models of whole-slice images from patients with historically localized small cell lung cancer are used as training samples. Radiologically confirmed brain metastases are used as supervisory labels. A binary classification cross-entropy loss function is constructed according to the following formula. : in, This serves as a true label for whether radiologically confirmed brain metastases have occurred. The probability of brain metastasis risk predicted by the brain metastasis risk prediction model; The learnable parameters in the linear projection layer, S4D module, residual fusion layer, and classification predictor are jointly optimized end-to-end through backpropagation to minimize the loss function until the model converges, thus completing the training of the brain metastasis risk prediction model. In practical applications, the model can also be updated and optimized based on new case data and follow-up outcome information to improve its applicability in different medical environments.

[0084] This embodiment preferably uses the AdamW optimizer, with an initial learning rate set to... The weight decay is set to The system is trained using a cosine annealing learning rate scheduling strategy. Through end-to-end training, the S4D module, the MIL aggregation module, and the classifier jointly learn whole-slice morphological patterns associated with brain metastasis risk.

[0085] Finally, the pathological whole-slice images of patients with limited-stage small cell lung cancer to be predicted are preprocessed and the image block feature sequences obtained by the pre-trained pathological characterization model are input into the trained brain metastasis risk prediction model, and the corresponding brain metastasis risk prediction probability is output.

[0086] Example 2: This embodiment will provide a detailed explanation of the feasibility and advancement of the method proposed in Embodiment 1, based on specific experimental data.

[0087] Specifically, this embodiment proposes a BraMARS deep learning architecture for implementing the method described in Embodiment 1 above. Figure 2The diagram illustrates the BraMARS deep learning architecture provided in this embodiment of the invention. (A) A flowchart illustrates patient enrollment, inclusion / exclusion criteria from two independent medical centers (CHCAMS and HMUCH), and cohort allocation for model development, validation, and clinical utility assessment. (B) A schematic diagram of the BraMARS deep learning model architecture, including three main components: WSI preprocessing and plaque-level feature extraction, sequence feature encoding using S4, and MIL aggregation for final prediction. (C) The workflow for model evaluation and interpretability. (D) The potential clinical impact of BraMARS, demonstrating its integration into the process of personalized PCI recommendation clinical decision-making and the expected improvement in patient outcomes. TP, True Positive; TN, True Negative; WSI, Whole Slice Image; S4, Structured State Space of Sequence; MIL, Multi-Instance Learning.

[0088] Specifically, the BraMARS deep learning architecture is divided into four main modules: queue construction, model computation, analysis and validation, and clinical translation, corresponding to the four sub-graphs A to D in the diagram. The composition and operation mechanism of each module are as follows: Figure 2 Part A outlines the research cohort construction and allocation process, providing a patient sample base for model training, validation, and clinical utility analysis. The study included patients with limited-stage small cell lung cancer (LS-SCLC) from two independent medical centers: the Chinese Academy of Medical Sciences Cancer Hospital (CHCAMS) cohort included patients treated between January 2005 and December 2016, and the Harbin Medical University Cancer Hospital (HMUCH) cohort included patients treated between December 2014 and December 2020. Enrollment required meeting three inclusion criteria: histological confirmation of limited-stage LS-SCLC; R0 surgical resection with available H&E-stained tissue sections; and complete clinicopathological data and follow-up data. Three categories of patients were excluded: those with a history of brain metastases at initial diagnosis; those with a history of head irradiation or neurosurgery prior to SCLC diagnosis; and those with substandard whole-section image quality. A total of 354 patients were ultimately included after screening, and divided into two groups according to whether they received prophylactic brain irradiation (PCI): 292 patients who did not receive PCI and 62 patients who did receive PCI. Among them, the patients who did not receive PCI were further divided by center: the CHCAMS cohort contained 188 patients, from which 70 true negative (TN) and true positive (TP) patients were selected for model development and internal testing; the HMUCH cohort contained 104 patients, from which 51 true negative (TN) and true positive (TP) patients were selected for external validation; and the 62 patients who received PCI formed an independent cohort for AI-assisted PCI benefit assessment analysis.

[0089] Figure 2Part B is the core deep learning architecture of the BraMARS model, which is divided into four levels along the data flow: whole-slice image processing, patch-level feature extraction, sequence feature encoding, and risk score output. First, a digitized pathological whole-slice image is input, with a single whole-slice pixel size of approximately 60,000 pixels × 30,000 pixels. Then, patch segmentation and filtering are performed, dividing the whole slice into image blocks of 224 pixels × 224 pixels. Invalid patches with a tissue area ratio of less than 50% are removed, retaining the set of valid image blocks. Valid patches enter the feature extraction module, sequentially passing through a CNN module, a linear embedding and block merging layer, and a four-stage SwinTransformer encoder. Finally, a global pooling operation outputs 768-dimensional global slice features, completing the extraction of patch-level morphological representations, corresponding to the feature extraction process of the pre-trained pathological representation model in the aforementioned method. The extracted slice features are input into the structured sequence state space (S4) module for long-range dependency modeling. The features are first processed by linear projection, ReLU activation, and layer normalization (LayerNorm), and then transformed into sequential state input x(t) using zero-order preservation techniques. The S4 module, centered on a state-space equation, includes a learnable state matrix A, input matrix B, output matrix C, and skip connection parameters D. The state matrix is ​​initialized using the HiPPO high-order polynomial projection method and continuously updated during training. The module performs state recursion through multiplication and summation operations, and combines discretization to convert the recursive computation into a parallelizable form, ultimately outputting the sequential representation y(t) for each plaque. The output representation is fused with the original input features through skip connections, and further processed by max pooling, linear projection, and softmax operations to finally output a brain metastasis (BM) risk score. This architecture, through the S4 module, captures the collaborative associations of distant pathological regions across the entire slice, overcoming the limitations of traditional multi-instance learning that relies solely on local discriminant regions.

[0090] Figure 2Part C of the model evaluation module comprises the system for interpretability analysis and biological mechanism validation, supporting the model's reliability from three dimensions: performance, pathological implications, and molecular basis. The left side is the model evaluation module: it quantifies model performance using receiver operating characteristic (ROC) curves, with the true positive rate (TPR) on the vertical axis and the false positive rate (FPR) on the horizontal axis, calculating the area under the curve (AUC) as the core evaluation indicator; it compares the survival differences between the high-risk and low-risk BraMARS groups using Kaplan-Meier survival curves, with the horizontal axis representing follow-up time; it compares the score distribution differences between patients with negative brain metastases (BM-) and positive brain metastases (BM+) using risk score distribution maps; and it compares the net benefit of BraMARS, AJCC staging, and all-treatment / no-treatment strategies using decision curve analysis, with the horizontal axis representing the threshold probability. The middle section is the interpretability analysis module: based on the original whole-slice images, it calculates the contribution of each plaque to risk prediction using the integral gradient method, generating a node map based on the integral gradient; simultaneously, it performs unsupervised clustering of plaques based on deep learning features, and presents the distribution characteristics of different histopathological clusters through two-dimensional dimensionality reduction visualization, locating pathological morphological patterns related to brain metastasis risk. The right side is the biological mechanism analysis module: it displays differentially expressed proteins between high- and low-risk groups using a volcano plot, with the horizontal axis representing the log2 (fold change) and the vertical axis representing the negative log10 (p-value) of the significance level; further, it generates a heatmap through functional enrichment analysis, revealing the biological pathways and molecular characteristics corresponding to risk stratification, providing biological explanations for model predictions.

[0091] Figure 2 Part D outlines the translational pathway and expected benefits of integrating the BraMARS model into clinical diagnosis and treatment. The clinical process begins with a patient diagnosed with limited-stage small cell lung cancer. After surgery and other supporting treatments, H&E-stained whole-slice images of the surgically removed tumor tissue are prepared. These whole-slice images are input into the BraMARS deep learning model, which outputs brain metastasis risk stratification and corresponding recommendations. Clinical experts, based on the risk references output by the model, formulate individualized PCI treatment decisions for the patient, recommending PCI for patients with high brain metastasis risk and not recommending PCI for low-risk patients, thereby optimizing treatment decisions. The final clinical benefit is validated through survival curves: treatment plans aligned with BraMARS risk recommendations show significantly better patient survival outcomes than those deviating from BraMARS guidance, confirming that this risk stratification framework can reduce overtreatment in low-risk patients while increasing preventative intervention coverage in high-risk patients, ultimately improving overall clinical benefit.

[0092] It should be noted that the experiments in this embodiment were conducted in accordance with the guidelines of the Declaration of Helsinki and obtained ethical approvals from the Cancer Hospital of the Chinese Academy of Medical Sciences (Approval No.: 22 / 250-3452) and the Cancer Hospital Affiliated to Harbin Medical University (Approval No.: KY2022-57). Due to the retrospective nature of the study, both ethics committees waived the requirement for informed consent.

[0093] In this embodiment, a total of 354 patients with limited-stage small cell lung cancer (LS-SCLC) from two independent medical institutions were included, including 245 patients from the Cancer Hospital of the Chinese Academy of Medical Sciences (CHCAMS cohort) between January 2005 and December 2016, and 109 patients from the Affiliated Cancer Hospital of Harbin Medical University (HMUCH cohort) between December 2014 and December 2020. Figure 2 (A)

[0094] The inclusion criteria are as follows: (1) Histologically confirmed LS-SCLC at initial diagnosis; (2) R0 surgical resection was performed, and H&E stained whole section images (WSI) were generated from the formalin-fixed paraffin-embedded (FFPE) tumor specimens resected by the surgery; (3) Clinicopathological and follow-up data were available.

[0095] Exclusion criteria were: (1) obvious brain metastases on imaging at the time of initial diagnosis; (2) history of head irradiation or neurosurgery before the diagnosis of SCLC; and (3) lack of H&E stained pathological sections, incomplete clinical data, or unqualified WSI quality.

[0096] The primary endpoint was radiographically confirmed brain metastases during follow-up. In this embodiment, patients classified as having positive brain metastases were those who developed brain metastases after enrollment in the study, not those with intracranial metastases at baseline. Brain metastasis-free survival (BMFS) was defined as the time from diagnosis to radiographically confirmed brain metastases or the last follow-up visit; death was treated as a competing event where applicable. Secondary endpoints included overall survival (OS), defined as the time from diagnosis to all-cause death or the last follow-up visit, and disease-free survival (DFS), defined as the time from diagnosis to disease relapse or death, or the last follow-up visit.

[0097] Based on the above experimental data, a specific implementation step of the method proposed in Example 1 is as follows: Step 1: Digital WSI Acquisition and Preprocessing.

[0098] FFPE tumor sections from surgically resected specimens of all study participants were stained with H&E and digitized into WSIs at 20× resolution using a digital slide scanner. The acquired WSIs were segmented into 224×224 pixel non-overlapping patches at a ratio of 2 μm / pixel. Otsu thresholding was applied to filter artifacts and non-tissue areas, retaining only patches containing ≥50% tissue for subsequent analysis.

[0099] Step 2: Development of deep learning models based on WSI.

[0100] This embodiment develops a deep learning model called BraMARS, based on a structured state-space model combined with multi-instance learning (S4MIL), for assessing the risk of future brain metastases in the WSI of LS-SCLC patients who have never undergone PCI. Figure 2 The deep learning framework shown in B consists of three consecutive components: WSI preprocessing and patch-level feature extraction, S4-based sequential feature encoding, and multi-instance learning aggregation for slice-level classification.

[0101] Step 3: Data preprocessing and patch-level feature extraction.

[0102] To generate robust and information-rich patch-level features, the CTransPath model was used to extract features from each 224×224 pixel patch. This model was pre-trained on a large-scale public pathology dataset, combining the local feature extraction capabilities of a convolutional neural network (CNN) with the global context modeling advantages of a Transformer. Ultimately, this pre-trained model generated a 768-dimensional morphological feature representation for each patch.

[0103] Step 4: Feature Encoding using S4D The extracted patch-level features are then processed by a diagonal structured state-space model (S4D). The neural state-space model (SSM) uses continuous-time state equations to describe the dynamics of the input data and model time dependencies.

[0104] For discrete-time implementation, this embodiment adjusts the state-space equation to a recursive form similar to that of a recurrent neural network (RNN).

[0105] The S4D variant simplifies computation by approximating the state matrix A as a diagonal and utilizing the Fast Fourier Transform (FFT) to transform recursive state updates into parallelizable convolutions. HiPPO initialization is applied to optimize the model's ability to capture long-range dependencies within the WSI, which is crucial for modeling the spatial propagation of pathological features.

[0106] Step 5: Feature aggregation and classification using MIL The final output signal y(t) for each patch is generated by a skip connection combining the original input u and the transformed hidden state x(t). The resulting patch-level representation sequence y(t) is then aggregated into a single slice-level representation through a max-pooling operation, a standard MIL method that identifies the most informative patches for the final prediction. The aggregated feature vector is then processed by a classifier to generate a brain metastasis risk prediction.

[0107] Step 6: Organize proteomics analysis.

[0108] Proteomics analysis was performed on tumor specimens from 56 SCLC patients. Data were obtained from the National Center for Biotechnology Information, Beijing Institute of Genomics, Chinese Academy of Sciences (https: / / ngdc.cncb.ac.cn / omix:accession number OMIX007230)

[24] . Protein abundance values ​​were normalized by centering on the median and then transformed by log2. Proteins with missing values ​​in more than 25% of the samples were excluded from subsequent analysis. The remaining missing values ​​were imputed using the k-nearest neighbor (k-NN) algorithm in the DreamAIR package

[25] . Differentially expressed proteins (DEPs) between the low-risk and high-risk groups of BraMARS were determined using the Wilcoxon test. A significant DEP was defined as a p-value < 0.05 and log2 (fold change) > 0.3 or < (-0.3).

[0109] Step 7: Functional characteristics of differentially expressed proteins.

[0110] Functional enrichment analysis of differentially expressed proteins was performed using the clusterProfilerR package across Kyoto Gene and Genome Encyclopedia (KEGG) pathways and Gene Ontology Bioprocesses (GOBP)

[26] . The enriched functional network of GOBP terms was constructed using Metascape

[27] . Twenty-nine knowledge-based functional gene expression tags (Fges) representing major functional components of tumors, as well as immune, stromal, and other cell populations, were obtained from Bagaev’s study

[28] . Single-sample gene set enrichment analysis (ssGSEA) of the 29 Fges was performed using GSVA (version 2.0.1).

[0111] Step 8: Multi-level pathological visualization and interpretation analysis.

[0112] To elucidate the histopathological basis of BraMARS predictions and enhance the interpretability of the model, this embodiment employs an analytical approach combining computational pathology, feature attribution mapping, and tissue morphology clustering. First, the predicted value or classification logit output by BraMARS for the high-risk category is used as the objective function F(X). A zero vector sequence with the same dimension as the input image patch feature sequence is used as the baseline X′. Multiple interpolation points are set between X′ and the true input X, and the gradient of the objective function with respect to the features of each image patch is calculated.

[0113] For the i-th image patch, the integral gradients obtained along the interpolation path for each feature dimension are summarized to obtain the signed attribution score for that image patch. Image patches with higher positive attribution scores are identified as the main driving regions for high-risk prediction, while image patches with lower or negative attribution scores are identified as supporting regions for low-risk classification. At the same time, the absolute attribution value is retained as the contribution strength to reduce the impact of positive and negative cancellation of different feature dimensions on the spatial localization results.

[0114] Subsequently, K-nearest neighbor relationships between image patches were constructed based on the 768-dimensional morphological features generated by CTransPath, and unsupervised clustering was performed. The number of candidate clusters was jointly evaluated using the Calinski-Harabasz index and the Davies-Bouldin index, selecting the number of clusters with high inter-cluster separation and good intra-cluster compactness; in this embodiment, 12 histopathological clusters were ultimately formed. UMAP was only used for two-dimensional visualization of high-dimensional features and clustering results and was not used for brain metastasis risk prediction.

[0115] The integral gradient attribution score of each image patch is associated with its histopathological cluster label and two-dimensional spatial coordinates. The signed attribution scores of image patches within the same cluster are averaged and normalized to obtain the cluster-level risk contribution. Simultaneously, the proportion of image patches from different clusters in each patient's whole slice is calculated to form a patient-histopathological cluster proportion matrix. Based on the image patch coordinates, the attribution scores and cluster labels are mapped back to the whole slice image to generate attribution heatmaps, cluster distribution maps, and overlay maps.

[0116] Pathologists performed blinded review and morphological annotation of high-contribution histopathological clusters and their representative image blocks. For high-risk clusters, the focus was on evaluating tumor cell density, nuclear crowding, pleomorphism, mitotic activity, fibroblast proliferation, extracellular matrix, and chondrogenic structures. For low-risk clusters, the focus was on evaluating morphological patterns such as alveolar septum and bronchial epithelial preservation, lymphocyte infiltration, and necrosis.

[0117] In patients with proteomics data, Pearson correlation coefficients were calculated based on single-sample gene set enrichment analysis scores using histopathological cluster proportion matrices and expression tags of 29 functional genes. These coefficients, combined with differentially expressed protein and pathway enrichment results, were used to evaluate the consistency between risk-related histopathological clusters and biological processes such as cancer-related fibroblasts, matrix remodeling, immune cell status, tumor proliferation, epithelial-mesenchymal transition, mitochondrial energy metabolism, reactive oxygen species detoxification, and DNA repair. This analysis served as a biological validation step for the prediction results; proteomics data were not input into the BraMARS risk prediction model.

[0118] Implementation Details: The BraMARS deep learning model was implemented and trained using data from the CHCAMS queue within the following computational framework. Model optimization was performed using the AdamW optimizer with a batch size of 1. The initial learning rate used in the training protocol was 1×10⁻⁶. -5 The weight decays to 1×10 -4 The training process was extended for over 200 epochs to ensure convergence. A cosine annealing scheme with momentum was used to dynamically adjust the learning rate throughout the training process. A binary cross-entropy loss with logarithmic parameters was used as the objective function. All computational experiments were performed on a Windows workstation equipped with an NVIDIA® GeForce RTX 4060 GPU, using PyTorch (version 2.1.0) as the deep learning framework.

[0119] Statistical Analysis Notes: All statistical analyses were performed using R software (version 4.1.3) and related R packages. Continuous variables were summarized using the mean and standard deviation or median and interquartile range (IQR), and categorical variables were expressed as frequencies and percentages. Survival curves were generated using the Kaplan–Meier method, and differences between survival distributions were assessed using the 'survminer' R package (log-rank test). The prognostic value of BraMARS was assessed using univariate and multivariate Cox proportional hazards regression models. The predictive performance of BraMARS was quantified by the area under the receiver operating characteristic (AUC) curve. A two-sided p-value <0.05 was considered statistically significant.

[0120] Based on the above steps, this embodiment conducted a retrospective, multicenter study, which included 354 patients with resected LS-SCLC, of ​​whom 245 patients were from the CHCAMS cohort and 109 patients were from the HMUCH cohort. Figure 2 ).

[0121] The mean age in the CHCAMS cohort was 56.57 ± 10.06 years, and in the HMUCH cohort it was 59.26 ± 7.81 years. Both cohorts were predominantly male. Smoking history was common in both cohorts, accounting for 64.49% and 72.47% of cases, respectively. Disease stage distribution differed between the two cohorts, with a higher proportion of stage III disease (40.4%) in the CHCAMS cohort than in the HMUCH cohort (22.9%). Since all enrolled patients had surgically resected tumor specimens available for WSI analysis, this cohort primarily enriched patients receiving surgery-based multimodal treatment strategies. The main treatment strategies are summarized in Table 1. In the CHCAMS cohort, 187 patients received postoperative chemoradiotherapy, 24 received preoperative chemoradiotherapy, 5 underwent surgery only, and treatment information was missing for 29 patients. In the HMUCH cohort, 59 patients received postoperative chemoradiotherapy, 17 received preoperative chemoradiotherapy, and 33 underwent surgery only.

[0122] Of all participants, 292 patients (82.5%) did not undergo PCI, while 62 patients (17.5%) did. During follow-up, 42 patients (11.9%) developed radiographically confirmed brain metastases, while 312 patients (88.1%) did not. Among the patients who later developed brain metastases, three underwent PCI before radiographic detection of intracranial disease. The relatively low PCI rate and the coarse incidence of brain metastases should be interpreted within the context of this retrospective, surgically resected, and pathologically documented cohort, rather than as a representative estimate of all LS-SCLC patients receiving definitive chemoradiotherapy.

[0123] To develop a binary model and perform direct discriminative assessment, this embodiment constructs a high-confidence subset of endpoints in patients who have not undergone PCI. Patients who develop radiographically confirmed brain metastases during follow-up, whether the metastases occur as an isolated intracranial recurrence or concurrent with extracranial progression, are designated as true positive cases, while patients with DFS ≥ 5 years and no evidence of metastasis during follow-up are designated as true negative cases. This definition aims to reduce endpoint misclassification in a retrospective setting.

[0124] Using these criteria, 70 eligible patients were identified from the CHCAMS cohort and 51 from the HMUCH cohort.

[0125] The CHCAMS cohort was randomly assigned in a 2:1 ratio to a training set (n=46) for model development and an internal validation set (n=24). The HMUCH cohort (n=51) served as a separate external validation cohort to assess the model's generalization ability. Additionally, 62 patients from two institutions who had undergone PCI were pooled into a single PCI cohort for exploratory clinical utility analyses related to PCI.

[0126] BraMARS is designed to assess the risk of future brain metastases from histopathological images. The model is trained on a specified training set and then evaluated through independent validation in internal and external validation cohorts.

[0127] like Figure 3 As shown, BraMARS demonstrated good discriminative ability, distinguishing between patients who subsequently developed brain metastases and those who did not develop metastases during follow-up. The model achieved an AUC of 0.907 (95% CI: 0.787–1.000) on the training set, 0.944 (0.856–1.000) on internal validation, and 0.738 (0.593–0.883) on external validation. Figure 3 (A)

[0128] Patients who subsequently developed brain metastases showed significantly higher BraMARS risk scores than those who did not develop brain metastases during follow-up. Figure 3 (B)

[0129] At the optimal risk threshold determined by maximizing the Youden exponent in the training set, BraMARS achieved sensitivity of 100% and 81.0% in internal and external validation cohorts, and specificity of 66.7% and 70.0%, respectively. Figure 3 C and Figure 3 (D).

[0130] These results suggest that BraMARS can stratify patients based on their varying risk of subsequent brain metastases after resection of LS-SCLC.

[0131] Decision curve analysis demonstrates that, within the threshold probability range of 5% to 70% in the CHCAMS cohort and 30% to 70% in the HMUCH cohort, BraMARS provides a higher net benefit than the "all treatment" and "no treatment" strategies. Figure 3 (E).

[0132] Calibration analysis further revealed a good agreement between the predicted probabilities and the observed brain transfer outcomes. For example... Figure 3 As shown in Figure F, the calibration curves for the two cohorts typically follow the diagonal reference line, demonstrating the reliability of BraMARS for personalized brain metastasis risk assessment.

[0133] The prognostic value of BraMARS was evaluated in patients who did not undergo PCI in the CHCAMS and HMUCH cohorts. Patients were divided into high-risk and low-risk groups based on BraMARS predictions.

[0134] In the CHCAMS cohort, BraMARS classified 88 patients (46.8%) as high-risk and 100 patients (53.2%) as low-risk; in the HMUCH cohort, 58 patients (55.8%) were classified as high-risk and 46 patients (44.2%) were classified as low-risk. In the CHCAMS cohort, patients classified as high-risk by BraMARS showed significantly worse DFS compared to low-risk patients (HR, 2.36; 95% CI, 1.55–3.58; log-rank test P < 0.0001). Figure 4 (A)

[0135] This prognostic pattern was also observed in the external HMUCH cohort, where high-risk patients also had shorter DFS (HR=2.52, 95%CI: 1.32-4.81; log-rank test P=0.0051). Figure 4 (B)

[0136] Multivariate Cox regression analysis, adjusted for established clinical factors, confirmed that the BraMARS high-risk category remained independently associated with poor DFS in both the CHCAMS cohort (HR, 2.38; 95% CI, 1.56–3.63; P < 0.001) and the HMUCH cohort (HR, 2.47; 95% CI, 1.27–4.81; P = 0.0078; Table S3).

[0137] In addition to DFS, BraMARS also demonstrated robust prognostic value in OS and BMFS.

[0138] In the CHCAMS cohort, high-risk patients showed significantly shorter OS (HR, 2.19, 95% CI, 1.39–3.46, P < 0.001) and BMFS (HR, 9.82, 95% CI, 2.23–43.29, P = 0.0026) compared to low-risk patients.

[0139] Consistent findings were observed in the OS (HR, 2.77, 95% CI, 1.43–5.36, P=0.0025) and BMFS (HR, 4.31, 95% CI, 1.45–12.87, P=0.0088) of the HMUCH cohort.

[0140] Multivariate Cox regression analysis further showed that, after adjusting for clinical factors in both cohorts, BraMARS risk classification remained independently associated with OS and BMFS.

[0141] This embodiment then evaluates whether BraMARS can supplement traditional AJCC staging. By combining the risks predicted by BraMARS with the traditional AJCC staging system, an improved three-category risk classification system is established. Figure 4 (C). In both the CHCAMS and HMUCH cohorts, this comprehensive classification improved the separation of DFS curves (log-rank tests showed P < 0.0001 and P = 0.0026, respectively). Figure 4 Patients with both BraMARS high-risk classification and stage III disease had the worst outcomes, while those with BraMARS low-risk classification and stage I-II disease had the best outcomes.

[0142] Consistently, the integrated tri-class classification also provided enhanced prognostic stratification for OS and BMFS in both cohorts. Importantly, BraMARS preserved prognostic value within AJCC staging subgroups, identifying patients with poor outcomes, even those at an earlier stage who might have been considered lower risk based solely on traditional staging. These findings support the complementary value of histopathology-driven deep learning assessments beyond traditional clinical staging.

[0143] To explore the potential clinical applicability of BraMARS in personalized PCI risk discussions, this example presents an exploratory retrospective simulation in a pooled cohort comprising 62 patients who received PCI and 292 who did not.

[0144] BraMARS risk scores were generated from pre-treatment pathological images, and patients were stratified based on the brain metastasis risk predicted by BraMARS and the actual PCI treatment status.

[0145] Combining patients who did not receive PCI treatment (untreated group), all cases were categorized into four distinct groups based on BraMARS prediction and actual treatment received. Figure 5 (A) Patients identified as high-risk by BraMARS and subsequently underwent PCI (predicted high-risk treatment group); 2) Patients predicted as low-risk by BraMARS but still underwent PCI (predicted low-risk treatment group); 3) Patients predicted as high-risk by BraMARS but did not undergo PCI (predicted high-risk untreated group); 4) Patients predicted as low-risk by BraMARS but did not undergo PCI (predicted low-risk untreated group).

[0146] like Figure 5As shown in Figure A, the simulation results indicate that BraMARS-based stratification can potentially avoid unnecessary PCI in 19.3% of patients (n=35, BraMARS low-risk treatment group) without affecting clinical outcomes, while also identifying 84.4% of high-risk patients (n=146, BraMARS high-risk untreated group) who are suitable candidates for PCI but would likely be overlooked under conventional protocols.

[0147] These findings suggest that BraMARS can provide a risk stratification framework to identify patients for whom there is a discrepancy between the predicted risk of brain metastases and the actual PCI usage.

[0148] Kaplan–Meier analysis showed that high-risk patients who did not undergo PCI had worse DFS than those who underwent PCI (log-rank test, P<0.0001). Figure 5 (B)

[0149] Conversely, in patients predicted to be at low risk, the disease-free survival (DFS) was relatively comparable between the PCI-treated and untreated groups. Figure 5 (B)

[0150] Parallel analyses with BMFS as the endpoint yielded consistent results: the BMFS of predicted high-risk patients who did not undergo PCI was shorter than that of predicted high-risk patients who underwent PCI, while predicted low-risk patients showed comparable BMFS regardless of whether they underwent PCI.

[0151] This embodiment further compares actual management with patients whose recommendations are consistent with BraMARS and patients whose management is inconsistent. Figure 5 (C)

[0152] The adherence group (including patients with predicted low risk who did not receive PCI and patients with predicted high risk who received PCI) showed significantly better DFS than the non-adherence group (including patients with predicted low risk who received PCI and patients with predicted high risk who did not receive PCI) (log-rank test P < 0.0001). Figure 5 (D). Similar benefits were observed for BMFS, with the group conforming to BraMARS recommendations showing better outcomes without brain metastases compared to the inconsistent group.

[0153] These exploratory findings suggest that BraMARS may help refine discussions of PCI-related risks by identifying high-risk patients with brain metastases who may require consideration of intracranial prophylaxis strategies, while potentially protecting predicted low-risk patients from unnecessary treatment-related toxicities.

[0154] Example 3: This embodiment further verifies the interpretability of the risk results output by the prediction method described in Embodiment 1 through multi-level interpretation. This interpretation processing chain is executed after the brain metastasis risk probability is generated, and transforms the model prediction into localizable and quantifiable histological evidence through image patch attribution, histopathological clustering, and spatial mapping; unless otherwise stated, the interpretive histopathological clusters are not required inputs to the main prediction model.

[0155] To quantitatively analyze the relationship between histopathological features and BraMARS risk, this embodiment first clusters WSI plaques into 12 distinct histopathological clusters (HCs) based on the learned morphological feature representations. Figure 6 (A)

[0156] Then, in this embodiment, the integral gradient (IG) value of each patch is calculated and these values ​​are mapped to the corresponding HC to evaluate its contribution to the model's risk prediction. Figure 6 (B)

[0157] Pathological evaluation showed that the characteristics of HCs (HC8, HC6 and HC2) associated with high risk prediction of BraMARS are common features in invasive tumors.

[0158] Specifically, the HC2 region showed high cell density, accompanied by images of nuclear crowding, pleomorphism, and frequent mitosis.

[0159] HC6 and HC8 showed significant mesenchymal activation, characterized by proliferating fibroblasts and lung cartilage fragments within a dense extracellular matrix.

[0160] These histopathological patterns are recognized markers of invasive potential and metastatic capacity, and support the classification of these areas as high-risk areas for BraMARS, which is associated with an increased risk of subsequent brain metastases.

[0161] Conversely, the HCs (HC5, HC9, HC1, and HC11) associated with low risk prediction by BraMARS showed characteristics associated with good prognosis.

[0162] HC5 and HC9 maintained relatively intact lung structure, with identifiable alveolar septa and bronchial epithelium. HC1 showed extensive infiltration of immune cells, primarily lymphocytes, suggesting active antitumor immunity. HC11 contained extensive areas of necrosis, which in this context may indicate a response to treatment or impaired tumor vitality.

[0163] Furthermore, the spatial distribution and incidence of these HCs differed significantly among patients with different BraMARS risk categories.

[0164] High-risk patients exhibited clear aggregation of invasive histology (HC2, HC6, HC8) with correspondingly high IG values, indicating that these areas are a major driver of poor prognosis. Figure 6 (C)

[0165] In contrast, low-risk patients exhibited a more heterogeneous tissue landscape dominated by favorable HC, with dispersed spatial patterns, which was associated with better clinical outcomes, as evidenced by significantly longer disease-free survival in representative cases. Figure 6 (C). This comprehensive pathological analysis confirms that BraMARS risk stratification captures fundamental differences in the tissue, cellular composition, and structural patterns of the tumor microenvironment.

[0166] To explore the potential biological basis of BraMARS prediction, this study performed proteomic analysis on tumor samples from 56 patients with LS-SCLC. Based on the BraMARS prediction risk score, these patients were divided into high-risk and low-risk groups.

[0167] Differential protein expression analysis identified 79 proteins that were significantly upregulated in the high-risk BraMARS group and 76 proteins that were upregulated in the low-risk group. Figure 7 (A)

[0168] Unsupervised clustering analysis showed that the expression patterns of these differentially expressed proteins (DEPs) effectively separated the two risk components. Figure 7 (A) This indicates that the risk stratification defined by BraMARS is associated with different molecular maps.

[0169] Functional enrichment analysis of differentially expressed proteins further revealed the biological pathways associated with BraMARS risk classification. Figure 7 (B)

[0170] These DEPs are significantly enriched in mitochondrial energy production pathways, transcriptional regulation in cancer, reactive oxygen species (ROS) detoxification, and DNA repair. Figure 7 (C)

[0171] These findings suggest that high-risk BraMARS tumors may have metabolically active and stress-adaptive phenotypic features.

[0172] To link molecular characteristics with histological patterns, this embodiment used ssGSEA with 29 functionally defined gene expression tags and examined their association with histopathological clusters.

[0173] This analysis revealed a coherent relationship between BraMARS-related histological patterns and tumor microenvironment status. Figure 7 (D).

[0174] High-risk histological clusters (HC6 and HC8) were positively correlated with cancer-associated fibroblasts, matrix remodeling, and pro-tumor immune tags, while low-risk clusters (HC1 and HC5) were associated with immune-related traits, including T-cell and B-cell tags.

[0175] In addition, the high-risk group showed higher scores in tumor proliferation and epithelial-mesenchymal transition (EMT) characteristics, which is consistent with a more aggressive tumor state and an active microenvironment.

[0176] This invention provides a device for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, used to implement the methods described in any of the above embodiments, such as... Figure 8 As shown, the device includes: The region detection module 801 is configured to acquire stained pathological whole slide images of patients with limited-stage small cell lung cancer, and to perform image block division and tissue region detection on the whole slide images, and to remove invalid image blocks according to the detection results to obtain a set of valid image blocks consisting of multiple valid image blocks. The feature extraction module 802 is configured to input each valid image block in the set of valid image blocks into a pre-trained pathological characterization model, extract the high-dimensional morphological feature vector of each valid image block, and obtain the feature vector set corresponding to the whole slice image. The feature sequence construction module 803 is configured to sort each high-dimensional morphological feature vector in the feature vector set according to a preset spatial scanning rule based on the two-dimensional spatial coordinates of each valid image block in the full slice image, thereby constructing an image block feature sequence with spatial continuity. The sequence feature enhancement module 804 is configured to sequentially input each high-dimensional morphological feature vector in the image block feature sequence into a linear projection layer for linear projection, and then fuse the linearly projected features with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features. The long-range dependency modeling module 805 is configured to input the location-enhanced sequence features into the S4D module, perform long-range dependency modeling through the state-space equation, and obtain the output representation corresponding to each image block output by the S4D module. The dependency enhancement module 806 is configured to perform residual fusion and normalization on the output representation corresponding to each image block output by the S4D module and the corresponding linearly projected feature input residual fusion layer to obtain the image block representation after long-range dependency enhancement. The risk prediction module 807 is configured to aggregate all long-range dependency-enhanced image patch representations into a full-slice-level representation, and input the full-slice-level representation into a classification predictor to obtain the brain metastasis risk prediction probability.

[0177] In some embodiments, the apparatus further includes an interpretability processing unit, which may serve as a post-processing unit of the risk prediction module 807, and includes: The feature attribution module 808 is configured to use the target risk value output by the risk prediction module 807 as the output to be explained, and calculate the feature attribution value of each valid image block by integral gradient. The histopathology clustering module 809 is configured to construct nearest neighbor relationships based on the high-dimensional morphological feature vectors of each effective image patch and perform unsupervised clustering to generate histopathology cluster labels, as well as the attribution contribution and image patch proportion of each histopathology cluster. The spatial mapping and interpretation output module 810 is configured to map feature attribution values ​​and histopathological cluster labels back to the whole pathological slice image based on the two-dimensional spatial coordinates of the image patch, generating a risk contribution heatmap, a spatial distribution map of histopathological clusters, and an interpretability report.

[0178] This invention provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0179] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0180] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0181] The electronic device provided in this embodiment of the invention can be the terminal device described in the above embodiments.

[0182] This invention also provides a computer-readable storage medium storing computer instructions that, when executed on a computer, cause the computer to perform the technical solution of the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images described in the above embodiments.

[0183] This invention also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images described in the above embodiments.

[0184] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0185] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0186] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0187] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute certain steps of the methods of the various embodiments of the present invention.

[0188] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0189] The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0190] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0191] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0192] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0193] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, characterized in that, The method includes: Acquire stained whole-slice images of pathological sections from patients with limited-stage small cell lung cancer, and perform image block division and tissue region detection on the whole-slice images. Based on the detection results, invalid image blocks are removed to obtain a set of valid image blocks consisting of multiple valid image blocks. Each valid image block in the set of valid image blocks is input into a pre-trained pathological characterization model, and the high-dimensional morphological feature vector of each valid image block is extracted to obtain the feature vector set corresponding to the whole slice image. Based on the two-dimensional spatial coordinates of each valid image block in the full slice image, the high-dimensional morphological feature vectors in the feature vector set are sorted according to a preset spatial scanning rule to construct an image block feature sequence with spatial continuity. Each high-dimensional morphological feature vector in the image block feature sequence is sequentially input into a linear projection layer for linear projection, and the linearly projected features are fused with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features; The enhanced sequence features are input into the S4D module, and long-range dependency modeling is performed through state-space equations to obtain the output representation corresponding to each image block output by the S4D module. The output representation of each image block output by the S4D module is residually fused and normalized with the corresponding linearly projected feature input residual fusion layer to obtain the long-range dependent enhanced image block representation. All long-range dependency-enhanced image patch representations are aggregated into full-slice-level representations, and these full-slice-level representations are input into a classification predictor to obtain the brain metastasis risk prediction probability.

2. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, The preset spatial scanning rules include row-first scanning, Z-order curves, Hilbert curves, or local neighborhood scanning methods based on tissue connectivity regions.

3. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, Each high-dimensional morphological feature vector in the image patch feature sequence is sequentially input into a linear projection layer for linear projection. The linearly projected features are then fused with the corresponding spatial location code and tissue cluster code to obtain the location-enhanced sequence features, specifically including: According to the following formula, the first... A high-dimensional morphological feature vector is linearly projected and normalized to map it to a unified hidden feature space, resulting in projected features: in, For the first High-dimensional morphological feature vectors of image patches and These are the learnable projection matrix and the bias term, respectively. For the first Projection features of an image patch; The projection features are fused with the corresponding spatial location code and tissue cluster code according to the following formula to obtain the location-enhanced sequence features: in, For the first Spatial location encoding of image blocks Encoding for organizational clusters, For the first The enhanced sequence features of the location of each image patch.

4. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, The enhanced sequence features are input into the S4D module, and long-range dependency modeling is performed using state-space equations to obtain the output representation corresponding to each image patch output by the S4D module, specifically including: Based on the S4D module, the input sequence is modeled using the following state-space equation: in, Indicates the first Enhanced sequence features of the location of each image patch Indicates the current hidden state. Indicates the previous hidden state. Indicates the output of the S4D module. The output representation of each image patch , , , The learnable parameter matrix; The learnable parameter matrix In diagonal form, the recursive state update is transformed into a parallelizable one-dimensional convolution using the Fast Fourier Transform.

5. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, The long-range dependency-enhanced image patch representation is determined by the following formula: in, The first output of the S4D module The output representation of each image patch For the first Enhanced sequence features of the location of each image patch For the first Long-range dependency enhancement of image patch representations for each image patch.

6. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, All long-range dependency-enhanced image patch representations are aggregated into full-slice-level representations, and these full-slice-level representations are input into a classification predictor to obtain the brain metastasis risk prediction probability, specifically including: Max pooling is performed on all long-range dependency-enhanced image patch representations according to the following formula to obtain full-slice-level representations: in, N The number of valid image patches, , and For the 1st, 2nd and 3rd N Long-range dependency augmented image patch representation of each image patch Full-slice level characterization; The full-slice-level representation is input into the classification predictor according to the following formula, and after processing by the Sigmoid function, the brain metastasis risk prediction probability is obtained: in, and For learnable classification parameters, For the Sigmoid function, To predict the probability of brain metastasis risk.

7. The method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images according to claim 1, characterized in that, The method also includes constructing a brain metastasis risk prediction model; The brain metastasis risk prediction model includes a linear projection layer, an S4D module, a residual fusion layer, and a classification predictor connected in sequence. The image block feature sequences obtained by preprocessing and pre-training the pathological characterization model from the whole pathological slide images of patients with historical limited-stage small cell lung cancer were used as training samples, and whether or not radiologically confirmed brain metastasis events occurred during the patient's follow-up period were used as supervision labels. The training samples may come from a single medical institution or multiple medical institutions. The methods for confirming brain metastasis events include, but are not limited to, MRI, enhanced CT, or other imaging assessments that conform to clinical diagnostic standards. A binary classification cross-entropy loss function is constructed according to the following formula. : in, This serves as a true label for whether radiologically confirmed brain metastases have occurred. The probability of brain metastasis risk predicted by the brain metastasis risk prediction model; The learnable parameters in the linear projection layer, the S4D module, the residual fusion layer, and the classification predictor are jointly optimized end-to-end through backpropagation to minimize the loss function until the model converges, thus completing the training of the brain metastasis risk prediction model. The image block feature sequence obtained by preprocessing and extracting from the pathological whole slice image of the patient with limited-stage small cell lung cancer to be predicted is input into the trained brain metastasis risk prediction model, and the corresponding brain metastasis risk prediction probability is output. After obtaining the predicted probability of brain metastasis risk, the method further includes: Using the target output corresponding to the predicted probability of brain metastasis risk as the output to be explained, a baseline feature sequence with the same dimension as the image patch feature sequence is set, and multiple interpolation feature sequences are generated between the baseline feature sequence and the input image patch feature sequence. The gradient of the target output relative to the features of each image block in each interpolation feature sequence is calculated, and the gradient is integrated along the interpolation path to obtain the feature attribution value of each valid image block. Based on the feature attribution values ​​of each effective image patch and its two-dimensional spatial coordinates in the full-slice image, a full-slice risk contribution heatmap representing the contribution of different tissue regions to the prediction of brain metastasis risk is generated. The method further includes: performing unsupervised clustering based on the high-dimensional morphological feature vectors of each effective image block to form multiple histopathological clusters; mapping the feature attribution value of each effective image block to its respective histopathological cluster, and calculating the attribution contribution and image block proportion of each histopathological cluster; wherein the number of histopathological clusters is determined according to at least one of the Calinski-Harabasz index and the Davies-Bouldin index. Based on the whole-slice risk contribution heatmap, the attribution contribution of each histopathological cluster and the proportion of image blocks, interpretable results are generated. The interpretable results include the predicted probability of brain metastasis risk, the histopathological cluster with the highest positive contribution and / or the lowest negative contribution, the representative image blocks of each histopathological cluster, the spatial distribution area and quantitative contribution information.

8. A device for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images, characterized in that, The device includes: The region detection module is configured to acquire stained pathological whole-slice images of patients with limited-stage small cell lung cancer, and to perform image block division and tissue region detection on the whole-slice images. Invalid image blocks are removed based on the detection results to obtain a set of valid image blocks consisting of multiple valid image blocks. The feature extraction module is configured to input each valid image block in the set of valid image blocks into a pre-trained pathological characterization model, extract the high-dimensional morphological feature vector of each valid image block, and obtain the feature vector set corresponding to the whole slice image. The feature sequence construction module is configured to sort the high-dimensional morphological feature vectors in the feature vector set according to a preset spatial scanning rule based on the two-dimensional spatial coordinates of each valid image block in the full slice image, thereby constructing an image block feature sequence with spatial continuity. The sequence feature enhancement module is configured to sequentially input each high-dimensional morphological feature vector in the image block feature sequence into a linear projection layer for linear projection, and then fuse the linearly projected features with the corresponding spatial location code and tissue cluster code to obtain the position-enhanced sequence features. The long-range dependency modeling module is configured to input the location-enhanced sequence features into the S4D module, perform long-range dependency modeling through the state-space equation, and obtain the output representation corresponding to each image block output by the S4D module. The dependency enhancement module is configured to perform residual fusion and normalization on the output representation corresponding to each image block output by the S4D module and the corresponding linearly projected feature input residual fusion layer to obtain the image block representation after long-range dependency enhancement. The risk prediction module is configured to aggregate all long-range dependency-enhanced image patch representations into a full-slice-level representation, and input the full-slice-level representation into a classification predictor to obtain the brain metastasis risk prediction probability.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method for predicting the risk of brain metastasis in limited-stage small cell lung cancer based on histopathological images as described in any one of claims 1-7.

Images