Tumor microenvironment triage prediction method, device and equipment for small cell lung cancer and storage medium

Abstract

This invention provides a method, apparatus, device, and storage medium for predicting three subtypes of the tumor microenvironment in small cell lung cancer, relating to the field of medical image processing technology. The method includes: acquiring a stained full-view digital pathological image of a small cell lung cancer tumor tissue sample; preprocessing the image to obtain an effective set of patches; extracting patch feature vectors; constructing multi-instance learning features using the full-view image as a package and patch features as instances; constructing a clustering-constrained attention multi-instance learning model; obtaining attention weights in response to the input multi-instance learning features; weighted aggregation to obtain slice-level representations; and outputting the category probabilities and subtyping results of the three subtypes; simultaneously generating attention heatmaps, key patch ranking results, or probability reports. This invention achieves cross-modal label transfer from immunohistochemical typing to H&E images, enabling automatic prediction of the tumor microenvironment state based solely on conventional stained slices, thus assisting clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, apparatus, device, and storage medium for predicting the three types of tumor microenvironment in small cell lung cancer. Background Technology

[0002] Small cell lung cancer (SCLC) is a highly aggressive neuroendocrine malignancy characterized by rapid growth, early metastasis, and sensitivity to initial treatment but a high recurrence rate. In recent years, immune checkpoint inhibitors combined with chemotherapy have become one of the first-line standard treatment regimens for extensive-stage SCLC, but the overall benefit is limited, and there are significant differences in efficacy among different patients. Therefore, developing a subtyping tool that accurately reflects the tumor microenvironment and can be used to predict patient prognosis and the benefit of immunotherapy has significant clinical value.

[0003] Currently, classification methods for small cell lung cancer mainly include transcriptome-based molecular typing, immunohistochemistry (IHC) marker typing, and histological assessment methods based on pathological morphology. Transcriptome typing can comprehensively reflect the state of tumor cells and the characteristics of the tumor microenvironment, but it has high requirements for sample quality, sequencing platforms, detection costs, and bioinformatics analysis procedures, making it difficult to routinely promote in clinical practice. Immunohistochemistry-based typing methods have strong biological interpretability, reflecting the state of the tumor microenvironment through quantitative detection results of T cells, B cells, myeloid cells, macrophages, endothelial cells, and matrix-related markers. However, this method usually requires additional staining experiments, manual interpretation, or semi-quantitative analysis, and suffers from problems such as long detection cycles, increased costs, limited throughput, and insufficient cross-center consistency.

[0004] With the development of digital pathology technology, artificial intelligence models based on whole-slide images (WSI) have provided a new approach for the automated characterization of tumor tissue structure and microenvironment. However, existing deep learning models based on pathology images mostly focus on tumor detection, tissue segmentation, pathological subtype diagnosis, prognostic risk scoring, or single treatment response prediction, and rarely use biologically interpretable immunohistochemical microenvironment typing results as supervisory labels and transfer them to conventional H&E-stained whole-slide images for prediction.

[0005] In summary, the existing technology mainly suffers from the following defects and shortcomings: (1) Transcriptome or multiple immunohistochemistry typing methods are costly and time-consuming, rely on additional experimental platforms or complex detection processes, and are difficult to be directly embedded into routine H&E pathology workflows for rapid deployment.

[0006] (2) Although immunohistochemical microenvironment typing has strong biological interpretability, it requires additional slide staining, marker detection and manual or semi-automatic quantitative analysis, making it difficult to achieve low-cost, rapid and automated typing based on conventional H&E slides.

[0007] (3) Existing artificial intelligence models based on H&E pathological images are mostly used for tumor identification, tissue region segmentation, prognostic risk classification or treatment response prediction, and lack a special modeling scheme for cross-modal prediction of "tumor microenvironment triad results defined by IHC".

[0008] (4) Full-view digital pathology images are extremely large in size, and the morphological clues related to the microenvironment are usually focal, sparse and spatially heterogeneous. If ordinary whole-image classification, sliding window supervision or simple patch averaging aggregation strategy is used, it is easily affected by non-tumor background, necrotic areas, folds, inflammatory reactions, stromal components and large areas of tumor parenchyma.

[0009] (5) Although existing weakly supervised multi-instance learning methods can perform slice-level classification in the absence of pixel-level or region-level annotations, they still have problems such as insufficient key patch recognition ability, large intra-class phenotypic differences, and unstable cross-center generalization performance in the task of transferring IHC microenvironment labels to H&E images.

[0010] (6) Existing technologies usually only output results such as benign or malignant, high or low prognostic risk, or whether or not treatment is responded to. They lack fine-grained, interpretable classification results for tumor microenvironment states such as IE, Desert and Stromal types, and are difficult to directly serve patients’ clinical stratification, prognostic assessment and potential benefits of immunotherapy. Summary of the Invention

[0011] This invention provides a method, device, equipment, and storage medium for predicting the three subtypes of the tumor microenvironment in small cell lung cancer, aiming to solve the following technical problems: how to construct a cross-modal microenvironment subtyping prediction method using immunohistochemical microenvironment subtyping results as labels and conventional H&E full-view digital pathology images as input, so as to obtain three subtypes (IE, Desert, and Stromal) based solely on H&E slices without the need for additional transcriptomic or multiplex immunohistochemical detection; at the same time, under the condition of only slice-level labels, automatically learn discriminative features, identify key regions, and improve the stability of the model under tissue heterogeneity, background noise, and cross-center staining differences.

[0012] In a first aspect, the present invention provides a method for predicting the three subtypes of the tumor microenvironment in small cell lung cancer, the method comprising: Acquire stained full-field digital pathological images of target small cell lung cancer tumor tissue samples; The full-view digital pathological image is subjected to tissue region detection, non-tissue background removal, patch segmentation, patch quality control and / or color standardization to obtain an effective patch set. The effective set of map patches is input into a pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors; Using the full-view digital pathological image as a package, and the multiple image patch feature vectors as multiple instances in the package, a multi-instance learning feature is constructed. A clustering-constrained attention multi-instance learning model is constructed. This model responds to input multi-instance learning features and obtains attention weights for each instance learning feature. Based on these attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on these slice-level representations, the model outputs the category probabilities and classification results for three tumor microenvironment subtypes: IE, Desert, and Stromal. The clustering-constrained attention multi-instance learning model is trained using the three-level classification results of small cell lung cancer tumor microenvironment obtained based on immunohistochemical marker quantification as supervised labels for full-view digital pathology images. Generate an attention heatmap, a key tile ranking result, or a probability report; wherein the attention heatmap is obtained based on the attention weights, the key tile ranking result is obtained by sorting the key tiles in descending order based on the magnitude of the attention weights, and the probability report is generated based on the category probability.

[0013] In one possible design, the full-view digital pathology image undergoes tissue region detection, non-tissue background removal, tile segmentation, tile quality control, and / or color normalization to obtain an effective tile set, including: The low-magnification thumbnail of the full-view digital pathology image is converted from the RGB color space to the HSV color space. Based on the set saturation and brightness thresholds, the initial tissue mask is calculated using the following formula. : in, Indicates the position of image pixels. The saturation threshold, The brightness threshold. Indicates saturation. Indicates brightness. Indicates the initial tissue mask At image pixel location The value at; Based on the initial tissue mask The final tissue mask is calculated using the following formula. : in, For the removal of small connected components. For filling the holes, This is a morphological closing operation. Open operations for morphology.

[0014] In the final tissue mask Within the corresponding tissue region, the full-view digital pathology image is segmented according to a preset tile size and step size, and a candidate tile set is obtained using the following formula. ,in, Indicates the first i One candidate patch, Indicates the total number of candidate tiles; For each candidate patch The proportion of its tissue area is calculated using the following formula. : in, Representing candidate plots The total number of pixels; when Remove the candidate tile at that time. ,when The candidate tile is retained at that time. , Preset a threshold for the proportion of organizational areas; For the retained candidate patches, the average brightness value is calculated using the following formula. Luminance variance and Laplace variance : in, Indicates the brightness channel value. Represents the Laplace operator. Used to evaluate tile sharpness This is the variance calculation function; Based on the average brightness of the retained candidate patches Luminance variance and Laplace variance Each tile is compared with its corresponding preset threshold, and tiles that do not meet the threshold requirements are removed to obtain the intermediate set of valid tiles. in, Indicates the first One valid tile, , Indicates the total number of valid tiles; For the valid set of tiles Color standardization is performed, and the resulting set of color-standardized tiles is used as the final valid set of tiles. ,in, Indicates the color-normalized first... If a valid tile is not color-normalized, then .

[0015] In one possible design, the set of effective patches is input into a pre-trained pathological image feature extraction network to obtain multiple patch feature vectors, including: Each valid tile Input the pre-trained pathological image feature extraction network and calculate the patch feature vector using the following formula. : in, For pre-training the pathological image feature extraction network, UNI, CONCH, CTransPath, ResNet, ConvNeXt, or VisionTransformer were selected. For feature dimension, It is the space of real numbers; For the full-view digital pathology image, the feature vectors of all valid patches constitute the feature set of the image.

[0016] In one possible design, the clustering-constrained attention multi-instance learning model responds to the input multi-instance learning features and outputs the category probabilities and classification results of three tumor microenvironment subtypes: IE, Desert, and Stromal, in the following manner: Patch feature vectors from input multi-instance learning features The transformed instance features are calculated using the following formula. : in, and These are the learnable parameters of the feature transformation layer. It is a non-linear activation function; Transformed instance features A gating attention mechanism is used to calculate the attention score using the following formula. : in, , and For the learnable parameters of the attention scoring module, This represents element-wise multiplication. The hyperbolic tangent activation function is used. For activation functions; Based on the attention score Attention weights are calculated using the following formula. : in, , This represents the attention score of the k-th tile. It is the softmax function; Based on the attention weight The slice-level representation is calculated using the following formula. : Slice-level representation The input classification output layer calculates the three-class logical values ​​using the following formula. : in, , , , These are the logits values ​​for IE, Desert, and Stormal types, respectively. and Learnable parameters for the classification output layer; The predicted probabilities of the three microenvironment subtypes are calculated using the following formula: in, This indicates that the full-view digital pathology image belongs to the category. The predicted probability, Set of microenvironment subtypes Elements in; Based on the predicted probability, the final predicted category is obtained using the following formula. : Here, arg max is a function that takes the index corresponding to the maximum value.

[0017] In one possible design, attention heatmaps, key tile ranking results, or probability reports are generated, including: The attention weight of each patch is mapped to its spatial coordinates in the full-view digital pathology image. The attention weights are normalized using the following formula: in, For the first Normalized attention weights for each patch, Attention weights for all tiles The minimum value, The maximum value of the attention weights for all tiles. To prevent constants with a denominator of zero; The normalized attention weights are superimposed on the thumbnail of the full-view digital pathology image according to their corresponding spatial locations to obtain an attention heatmap, which is used to indicate the pathological areas that contribute the most to the triad determination. When the same target has multiple full-view digital pathology images, prediction is performed for each full-view digital pathology image separately, and the prediction is made on the first... The category probability of Zhang Quan's digital pathology image is Image weights are In this case, the target-level category probability is calculated using the following formula: in, The total number of full-view digital pathology images targeting [the target]. According to the Determining the effective tissue area, image quality score, maximum prediction confidence, or attention region intensity of Zhang Quan's digital pathology images; The target-level classification result is obtained using the following formula: in, The final comprehensive classification result of the tumor microenvironment is the target.

[0018] In one possible design, the clustering-constrained attention multi-instance learning model is trained as follows: Acquire the full-view digital pathology image corresponding to each target in the training queue; Acquire immunohistochemical microenvironment triad labels corresponding to the full-view digital pathology images used for training, the labels including IE type, Desert type and Stromal type; Tissue region detection, patch segmentation, patch quality control, and color standardization were performed on the full-view digital pathology images used for training. A pre-trained pathological image feature extraction network was used to encode the image patches to obtain the patch feature matrix; Each training full-view digital pathology image is constructed as a multi-instance learning package and combined with the corresponding immunohistochemical microenvironment label to form training samples. The training samples are input into the clustering-constrained attention multi-instance learning model, and the overall model loss, including slice-level cross-entropy loss and instance-level clustering loss, is calculated. The model parameters are updated using the backpropagation algorithm; Five-fold cross-validation was used for model training, parameter selection, and robustness evaluation in the training set. The model with the best validation performance was selected as the trained clustering-constrained attention multi-instance learning model.

[0019] In one possible design, training samples are input into a clustering-constrained attention multi-instance learning model, and the overall model loss, including slice-level cross-entropy loss and instance-level clustering loss, is calculated, including: For each training session, use a multi-instance learning package. According to attention weight The candidate positive instance sets are obtained using the following formulas. and candidate negative instance set : in, Before selection Operations for finding the maximum value, After selection Operations to find the minimum value The preset number of instances; Input candidate positive instances and candidate negative instances into the instance-level discriminator. The instance-level prediction results are obtained using the following formula; in, For instance-level discriminators, For the learnable parameters of the instance-level discriminator, For the first The predicted probability that an instance is a positive instance; The instance-level clustering loss is calculated using the following formula. : in, It is the natural logarithm function; For the training set j For each sample, the slice-level cross-entropy loss is calculated using the following formula. : in, The one-hot encoding of the true label of the j-th sample. The full-view digital pathology images used for model prediction training belong to the category. c The probability of; The overall loss of the model is calculated using the following formula. : in, For instance-level loss weights, The L2 regularization coefficient is... This represents the set of learnable parameters of the model.

[0020] Secondly, the present invention provides a tumor microenvironment triad prediction device for small cell lung cancer, the device comprising: The data acquisition module is configured to acquire stained full-field digital pathological images of target small cell lung cancer tumor tissue samples; The data preprocessing module is configured to perform tissue region detection, non-tissue background removal, patch segmentation, patch quality control, and / or color standardization on the full-view digital pathology image to obtain an effective patch set. The feature extraction module is configured to input the effective set of map patches into a pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors; The instance feature construction module is configured to construct multi-instance learning features by taking the full-view digital pathology image as a package and the multiple patch feature vectors as multiple instances in the package; The model prediction module is configured to construct a clustering-constrained attention multi-instance learning model. This model responds to input multi-instance learning features by obtaining attention weights for each instance learning feature. Based on these attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on these slice-level representations, the model outputs the category probabilities and classification results for three tumor microenvironment subtypes: IE, Desert, and Stromal. The clustering-constrained attention multi-instance learning model is trained using the three-level classification results of small cell lung cancer tumor microenvironment obtained based on immunohistochemical marker quantification as a full-view digital pathology image-level supervised label. The prediction result generation module is configured to generate an attention heatmap, a key tile ranking result, or a probability report; wherein the attention heatmap is obtained based on the attention weights, the key tile ranking result is obtained by sorting the key tiles in descending order based on the magnitude of the attention weights, and the probability report is generated based on the category probability.

[0021] 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 tumor microenvironment triad prediction method for small cell lung cancer as described in the first aspect and various possible designs of the first aspect.

[0022] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the tumor microenvironment triad prediction method for small cell lung cancer as described in the first aspect and various possible designs of the first aspect.

[0023] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the three-type prediction method for tumor microenvironment in small cell lung cancer as described in the first aspect and various possible designs of the first aspect.

[0024] The method, apparatus, device, and storage medium for predicting the three types of tumor microenvironment in small cell lung cancer provided by this invention have at least the following beneficial effects: (1) This invention can predict the three types of tumor microenvironment based solely on conventional H&E full-view digital pathology images, without the need for additional transcriptome sequencing or multiplex immunohistochemistry detection, which significantly reduces costs and application barriers, and facilitates rapid deployment in conventional pathology workflows. By using the immunohistochemistry three-type results as slice-level supervisory labels, cross-modal label migration from the IHC classification system to H&E digital pathology images is achieved.

[0025] (2) The present invention adopts a clustering-constrained attention multi-instance learning model. The attention weight of each tile instance is calculated through the attention scoring module. Under the condition that there are only slice-level labels and lack pixel-level or region-level annotations, the key regions most related to the microenvironment phenotype can be automatically learned from a large number of tiles, which effectively improves the model's ability to adapt to organizational heterogeneity.

[0026] (3) This invention calculates instance-level attention weights through a gated attention mechanism and introduces an instance-level clustering constraint module. By selecting high-attention instances and low-attention instances and calculating instance-level clustering loss, it can more accurately capture the local pathological structures that determine the classification results compared to traditional multi-instance learning methods, thereby improving prediction accuracy and model interpretability. At the same time, a pre-trained pathological image feature extraction network is used to encode the patches, enhancing the image representation ability and cross-center generalization ability.

[0027] (4) This invention uses the three-level subtyping results of immunohistochemistry microenvironment as slice-level supervised labels to train a clustering-constrained attention multi-instance learning model, realizing the technology transfer from the existing interpretable IHC subtyping system to the conventional H&E digital pathology automatic prediction, and has a clear technology inheritance relationship and clinical translation value.

[0028] (5) The three types of results output by this invention include IE type (immune enrichment type), Desert type (immune desert type) and Stromal type (matrix enrichment type), which correspond to different pathological microenvironment states and have clear biological and clinical interpretive significance. They can be directly used for patient stratification, prognostic assessment and prediction of potential benefits of immunotherapy.

[0029] (6) In the training process, the present invention uses a total loss function that includes slice-level classification loss, instance-level clustering loss and L2 regularization term to optimize the model, and uses 5-fold cross-validation in the training set. Furthermore, the model is validated in multiple independent outer center queues, which can prove that the model has good stability, repeatability and cross-center generalization ability.

[0030] (7) The present invention can generate an attention heatmap based on the attention weight of the image patch, and superimpose the normalized attention weight onto the full-view digital pathology image thumbnail according to the spatial position. It can also output the key image patch sorting results and probability reports, which facilitates the review by pathologists and significantly enhances clinical credibility and interpretability of results.

[0031] (8) This invention can be deployed as an independent software algorithm on a digital pathology platform, or as a hardware module to interface with existing hospital scanners, image management systems and clinical information systems. In addition, when the same patient has multiple H&E slides, this invention provides a patient-level fusion step, which obtains the final patient-level classification result by weighted fusion of the probabilities of each slide category, and has good prospects for engineering implementation and flexibility in clinical application. Attached Figure Description

[0032] 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.

[0033] Figure 1 A flowchart illustrating a method for predicting the three types of tumor microenvironment in small cell lung cancer, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of full-view pathological image preprocessing, tissue detection, and image segmentation provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the clustering-constrained attention multi-instance learning model provided in an embodiment of the present invention; Figure 4A schematic diagram of representative pathological images of three pathological microenvironment subtypes provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of a ternary 5-fold cross-validation confusion matrix for small cell lung cancer tumor microenvironment provided in an embodiment of the present invention. Figure 6 ROC curves for pathological microenvironment subtypes in the test set provided in this embodiment of the invention; Figure 7 Kaplan-Meier survival curves for patients with different tumor microenvironment subtypes provided in this embodiment of the invention; Figure 8 This is a structural diagram of a tumor microenvironment triad prediction device for small cell lung cancer provided in an embodiment of the present invention.

[0034] 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

[0035] 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.

[0036] 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.

[0037] 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.

[0038] 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.

[0039] This invention provides a method for predicting the three subtypes of tumor microenvironment in small cell lung cancer. The method uses the three subtypes of the immunohistochemical microenvironment as the label source and conventional H&E stained full-view digital pathological images as the model input. It learns the histological characteristics in the H&E images corresponding to the IHC microenvironment classification results through a weakly supervised multi-instance learning model, thereby realizing cross-modal label transfer from the IHC classification system to the H&E digital pathological image prediction model, and outputs three tumor microenvironment subtypes: IE, Desert, and Stromal.

[0040] The core of this invention is not simply applying existing multi-instance learning models to ordinary three-class classification tasks, but rather addressing the problem that IHC microenvironment typing results are difficult to obtain quickly and directly from conventional H&E slices. It establishes a cross-modal prediction framework that integrates immunohistochemical quantitative typing labels with morphological representations of H&E full-view images. Since IHC labels reflect complex microenvironment states such as immune cell infiltration, myeloid cell distribution, matrix activation, vascularity, and immunosuppression-related components, and this information appears as scattered, focal, and spatially heterogeneous morphological cues in H&E images, this invention employs a weakly supervised multi-instance learning approach. The entire WSI is constructed as a package, and patch features are constructed as multiple instances. Attention mechanisms and instance-level clustering constraints are used to filter key regions related to the microenvironment state from a massive number of patches.

[0041] Specifically, such as Figure 1 As shown, the tumor microenvironment triad prediction method for small cell lung cancer includes the following steps S10~S60.

[0042] S10: Acquire stained full-field digital pathological images of the target small cell lung cancer tumor tissue sample.

[0043] It should be noted that the target in this embodiment can be a patient with small cell lung cancer, and the tumor tissue sample is preferably surgically removed tissue, biopsy tissue, or other paraffin-embedded tissue (FFPE) samples from the patient. After routine H&E staining, the sample is digitally scanned using a full-field pathology scanner to generate a WSI image file. The scanning magnification is preferably 20× or 40×. The stained full-field digital pathology image, i.e., the digital slide, can be in SVS, NDPI, TIFF, MRXS, or other resolvable full-field image formats. For each patient, a single slide or multiple slides can be input; in the case of multiple slides, they can be predicted separately and then fused, or a comprehensive judgment can be made at the patient level.

[0044] In some embodiments, a full-field digital pathological image (WSI) stained with H&E is acquired, denoted as... The WSI can be in SVS, NDPI, TIFF, MRXS, or other digital pathology image formats. Based on the scan magnification and pixel physical resolution, the WSI is standardized to a preset analysis magnification, such as 20× magnification or an equivalent per-pixel physical scale.

[0045] If the original scanning magnification is 40×, a 20× analysis image can be obtained through downsampling. If images produced by different scanners have different mpp values, scale normalization is performed based on the mpp information so that different slices can be segmented and feature extracted at the same physical scale.

[0046] S20: Perform tissue region detection, non-tissue background removal, patch segmentation, patch quality control, and / or color standardization on the full-view digital pathology image to obtain an effective patch set.

[0047] In some embodiments, such as Figure 2 The diagram illustrates the full-view pathological image preprocessing, tissue detection, and image patch segmentation provided in this embodiment of the invention. First, the original whole-slice image is acquired, and a tissue region detection operation is performed to accurately identify the effective tissue regions in the image and remove non-tissue background. Next, image patch segmentation is performed within the detected tissue regions, uniformly dividing the large-size full-view digital pathological image into multiple 256×256 pixel image patches. Subsequently, quality control is performed on all segmented image patches, selecting retained patches while removing those that do not meet the analysis requirements. Specifically, removed image patches include blank image patches, blurred image patches, tissue folding image patches, contamination artifact image patches, and large-area necrosis image patches. Finally, the qualified image patches obtained after quality control are used as input to a feature encoder for feature encoding. The feature encoder outputs a high-dimensional feature vector for the corresponding image patch, with the high-dimensional feature vector having a dimension of, for example, 1024.

[0048] Specifically, in combination Figure 2 As shown, based on the full-view digital pathology image obtained in step S10, an effective set of patches can be obtained through the preprocessing steps S201-S204.

[0049] S201: Tissue region detection and non-tissue background removal.

[0050] Tissue region detection was performed on a low-magnification thumbnail. The thumbnail was converted from the RGB color space to the HSV color space to obtain the hue channel (H), saturation channel (S), and brightness channel (V). An initial tissue mask was constructed based on saturation and brightness thresholds. : in, Indicates the position of image pixels. The saturation threshold, The brightness threshold. Indicates saturation. Indicates brightness. Indicates the initial tissue mask At image pixel location The value at that point. Saturation threshold. and brightness threshold It can be determined based on the statistical results of the training set, the Otsu threshold method, or an empirical threshold.

[0051] Subsequently, the initial tissue mask was applied. Morphological opening and closing operations, hole filling, and small connected component removal are performed to obtain the final tissue mask. : in, For the removal of small connected components. For filling the holes, This is a morphological closing operation. Open operations for morphology.

[0052] The final tissue mask Used to define the area to be divided into subsequent tiles, thereby removing blank backgrounds.

[0053] In other implementations, tissue region detection can also be achieved using Otsu thresholding, color deconvolution, deep learning tissue segmentation models, or other tissue mask extraction methods.

[0054] S202: Plot segmentation.

[0055] In the final tissue mask Within the corresponding organizational area, according to the preset block size and step length The WSI is segmented into tiles to obtain a candidate tile set: in, Indicates the first One candidate patch, This indicates the number of candidate tiles. The tile size... Preferably 256 or 512 pixels, the step size Can be equal to To achieve non-overlapping segmentation, it can also be smaller than... To achieve overlapping segmentation.

[0056] For each candidate patch Calculate its organizational area proportion : in, Representing candidate plots The total number of pixels. When Remove the candidate tile at that time. ;when The candidate tile is retained at that time. . Set a threshold for the proportion of organizational regions. It can be 0.3, 0.5, 0.6 or other preset thresholds, preferably 0.5.

[0057] S203: Block quality control.

[0058] Further quality control was performed on the retained candidate tiles. For the retained candidate tiles... Calculate the average brightness Luminance variance Mean saturation and Laplace variance : in, Indicates the brightness channel value. Represents the Laplace operator. Used to evaluate the sharpness of patches. If a patch meets the rejection criteria such as excessive brightness, insufficient tissue proportion, excessive blur, abnormal staining, fold contamination, or excessive proportion of necrotic areas, the patch will be removed from the candidate set.

[0059] After quality control, the intermediate valid tile set is obtained. : in, Indicates the first One valid tile, .

[0060] S204: Color standardization.

[0061] Optionally, for the intermediate valid tile set Color standardization is performed to reduce color differences between different staining batches, different scanning devices, and samples from different centers. Color standardization can employ the Macenko method, Reinhard method, Vahadane method, or deep learning-based color normalization methods.

[0062] The set of tiles after color standardization is used as the final valid set of tiles. , represented as: in, Indicates the color-normalized first... One valid tile. If color normalization is not performed, then... .

[0063] It should be noted that step S204 is not a mandatory step. In some embodiments, color standardization may not be performed, and the intermediate set of valid tiles is used as the final set of valid tiles.

[0064] S30: Input the effective set of map patches into the pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors.

[0065] In some embodiments, patch feature extraction can be performed in the following manner.

[0066] Each valid tile Input pre-trained pathological image feature extraction network Obtain the feature vector of the map patch. : in, It can be UNI, CONCH, CTransPath, ResNet, ConvNeXt, VisionTransformer or other deep neural networks suitable for encoding pathological images; The feature dimension can be 512, 768, 1024, or other preset dimensions. In a preferred embodiment, For the UNI model, d =1024.

[0067] For a WSI, the feature vectors of all valid patches constitute the feature set of that slice: in, For feature set, This represents the number of feature vectors of the patch.

[0068] S40: Using a full-view digital pathology image as a package, and multiple patch feature vectors as multiple instances in the package, construct multi-instance learning features.

[0069] In some embodiments, given the extremely large size of full-view pathological images and the difficulty in pixel-by-pixel annotation, a Multiple Instance Learning (MIL) strategy is employed to construct training samples. This strategy enables the model to automatically focus on the local tissue regions most relevant to the three-part pathological microenvironment under weak supervision, thereby reducing the reliance on detailed manual annotation.

[0070] Specifically, in this embodiment, a single H&E full-view digital pathology image is considered as a package, and the multiple patch feature vectors extracted from that slice are considered as multiple instances within the package. For the first... Zhang WSI, its multi-instance learning input representation is: in, Indicates the first The packet corresponding to Zhang WSI Indicates the first Zhang WSI Middle School Feature vectors of each patch Indicates the first Number of valid tiles in a WSI sheet.

[0071] With the first The supervision label corresponding to Zhang WSI is denoted as: Where 0 represents IE type, 1 represents Desert type, and 2 represents Stormal type. (Supervisory tag) The results are derived from the immunohistochemical microenvironment trigenization results corresponding to the WSI, rather than from manual visual interpretation of H&E images.

[0072] Therefore, the training dataset can be represented as: in, This indicates the number of WSIs in the training set. Using this data organization method, this embodiment achieves cross-modal weakly supervised learning from H&E images to IHC microenvironment typing results, even with only slice-level IHC typing labels and lacking tile-level labels.

[0073] S50: Construct a clustering-constrained attention multi-instance learning model. The clustering-constrained attention multi-instance learning model responds to the input multi-instance learning features and obtains the attention weights of each instance learning feature. Based on the attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on the slice-level representations, the class probabilities and classification results of three tumor microenvironment subtypes, namely IE type, Desert type and Stromal type, are output. Among them, the clustering-constrained attention multi-instance learning model is trained with the three-level classification results of small cell lung cancer tumor microenvironment obtained based on the quantitative results of immunohistochemical markers as the full-view digital pathology image-level supervised label.

[0074] This embodiment employs a clustering-constrained attention multi-instance learning model to establish a predictive mapping relationship between H&E image features and IHC microenvironment triad labels. The preferred model is the CLAM-SB model, but CLAM-SB is only a preferred implementation. The core of this invention lies in a cross-modal weakly supervised prediction framework that uses IHC microenvironment typing results as supervisory labels and H&EWSI images as input.

[0075] In some embodiments, such as Figure 3 The diagram shown illustrates the structure of the clustering-constrained attention multi-instance learning model provided in this embodiment of the invention. This model employs a weakly supervised learning framework, using WSI as a bag and image patches as instances. The input consists of WSI slices and image patch features, where the WSI slices serve as the entire bag input. Each image patch corresponds to one instance, and the WSI is the entire bag, trained using a weakly supervised approach.

[0076] The input first enters the attention scoring module, also known as the attention module, which calculates attention weights for each image patch through an attention network. α i The model distinguishes between high-attention and low-attention image patches based on their attention weights, and high-attention image patches are given priority by the model.

[0077] The output of the attention scoring module is fed into the instance-level clustering constraint module, also known as the instance-level constraint module. This module uses high-attention and low-attention image patches to obtain instance-level constraints, resulting in more discriminative image patch representations. High-attention image patches are discriminative instances, while low-attention image patches are non-discriminative instances. Clustering constraints and contrast constraints are applied to both types of image patches respectively. Clustering constraints are used to bring image patches of the same class closer together, while contrast constraints are used to widen the distance between image patches of different classes. This instance-level clustering constraint module employs instance-level constraint supervision, improving the discriminativeness of image patch representations without additional annotations.

[0078] The features processed by instance-layer clustering constraints are fed into the bag weighted aggregation module, also known as the attention weighting or bag aggregation module. This module uses attention weights to aggregate and obtain the bag representation. Specifically, it uses attention weights... α i Obtain a slice-level representation, and finally output the slice-level representation, which is the cladding representation. h bag ∈ R d The weighted aggregation module uses slice-level label supervision, which is also a weak supervision method.

[0079] The slice layer represents the final input classification output layer, which contains classifiers for three categories. Each classifier consists of a fully connected layer plus a softmax layer. The final output is the probability of three classes, and the sum of the three probabilities is 1. The probabilities are P(IE), P(Desert), and P(Stromal).

[0080] The entire model's image data processing logic is based on weakly supervised learning. It uses WSI as a bag and image patches as instances, and performs weakly supervised training through slice-level label supervision. Finally, it outputs the probabilities of three tumor microenvironment subtypes: IE, desert, and stromal. In some embodiments, the constructed clustering-constrained attention multi-instance learning model includes the following modules: (1) Instance feature input layer: Receives the set of tile features corresponding to each WSI. ; (2) Feature transformation layer: Nonlinear mapping is performed on the patch features to obtain the transformed instance features. ; (3) Attention scoring module: Calculate the contribution weight of each tile instance to the slice-level microenvironment classification results; (4) Package-level aggregation module: The instance features are weighted and summarized according to the attention weights to form slice-level representations; (5) Instance-level clustering constraint module: performs instance-level discrimination learning for high-attention instances and low-attention instances; (6) Classification output layer: Output the class probabilities of IE, Desert and Stromal types.

[0081] The following section will elaborate on the specific data processing procedures of the six modules mentioned above in the clustering-constrained attention multi-instance learning model.

[0082] In the feature transformation layer, for the input patch feature vector First, feature transformation is performed using fully connected layers and non-linear activation functions: in, and These are the learnable parameters of the feature transformation layer. For ReLU, GELU, or other nonlinear activation functions, These are the transformed instance features.

[0083] In the attention scoring module, for each instance feature Calculate attention score In a preferred embodiment, a gated attention mechanism is employed: in, , and For the learnable parameters of the attention scoring module, This represents element-wise multiplication. The hyperbolic tangent activation function is used. For activation function, This is the matrix transpose.

[0084] Subsequently, the attention scores of all instances are normalized using the softmax function to obtain the attention weights. : in, This represents the attention score of the k-th tile. For the softmax function, This represents the attention weight, specifically the first... The contribution weight of each patch to the WSI microenvironment typing result, and satisfying the following: In the package-level aggregation module, all instance features are weighted and aggregated according to attention weights to obtain slice-level representations. : in, This represents the slice-level characterization, specifically the overall characterization of the H&E full-view digital pathology image, which is used to output the predicted probability corresponding to the IHC microenvironment triad results in subsequent outputs.

[0085] In the classification output layer, slice-level representations are used. Input the classification layer and output the three-class logits: in, , , , These are the logits values ​​for IE, Desert, and Stormal types, respectively. and Learnable parameters for the classification output layer The predicted probabilities of the three microenvironment subtypes are obtained using the softmax function: in, This indicates that the full-view digital pathology image belongs to the category. The predicted probability, Set of microenvironment subtypes elements in Final prediction category for: In the instance-level clustering constraint module, since the IHC microenvironment triad labels are only provided at the slice level, and the H&E morphological cues related to the microenvironment state often only exist in some key patches, this invention further adopts instance-level clustering constraints to enhance the key patch recognition capability.

[0086] During training, for each training package According to attention weight Select Top-K high-attention instances as the candidate positive instance set. Bottom-K low-attention instances are selected as the candidate negative instance set. : in, The preset number of instances is used. High-attention instances are considered more likely to contain morphological cues related to the current slice's IHC microenvironment typing, while low-attention instances are considered less relevant to the current typing.

[0087] The selected high-attention and low-attention instances are then input into the instance-level discriminator. Instance-level prediction results are obtained: in, This is an instance-level discriminator (typically a single-layer fully connected network with a sigmoid activation function). For the learnable parameters of the instance-level discriminator, For the first The predicted probability that an instance belongs to a positive instance.

[0088] Instance-level clustering loss It can be represented as: In implementations employing instance-level support vector machine constraints, the discrimination boundary between high-attention and low-attention instances can also be enhanced through support vector machine loss or margin loss. The instance-level clustering constraint enables the model to learn representative H&E morphological regions associated with IE, Desert, and Stromal types even in the absence of tile-level manual annotation.

[0089] For the training set For each sample, the slice-level cross-entropy loss is: in, One-hot encoding of the real label. The model predicts the category to which the slice belongs. The probability of.

[0090] The overall loss function of the model is: in, For slice-level classification loss, For instance-level clustering constraint loss, For instance-level loss weights, The L2 regularization coefficient is... This represents the set of learnable parameters of the model.

[0091] In one specific implementation, the model uses a learning rate. Regularization coefficient The training was performed with a dropout ratio of 0.25, instance-level support vector machine constraints enabled, and a bandwidth parameter of 0.7. These parameters are merely preferred embodiments and do not constitute a limitation on the scope of protection of this invention.

[0092] In one specific implementation, the training process of the clustering-constrained attention multi-instance learning model includes the following steps: (1) Obtain the H&E full-view digital pathology image corresponding to each patient or each slide in the training queue; (2) Obtain the immunohistochemical microenvironment trigenization tags corresponding to the H&E slices, the tags including IE type, Desert type and Stromal type; (3) Perform organizational region detection, tile segmentation, tile quality control, and optional color standardization on H&EWSI; (4) Encode the image patches using a pre-trained pathological image feature extraction network to obtain the image patch feature matrix. ; (5) Construct each WSI as a multi-instance learning package And its corresponding IHC microenvironment label Forming training samples; (6) Input the training samples into the clustering-constrained attention multi-instance learning model and calculate the slice-level classification loss and instance-level clustering loss; (7) Update the model parameters using the backpropagation algorithm; (8) Five-fold cross-validation was used in the training set for model training, parameter selection and robustness assessment; (9) Select the model with the best validation performance as the final prediction model.

[0093] In another specific implementation, the CLAM-SB model is used as an example of a clustering-constrained attention multi-instance learning model. Its H&E digital pathology microenvironment triad training process based on IHC label transfer includes the following steps 1-15.

[0094] Step 1: Collect paraffin-embedded tissue samples from patients with small cell lung cancer, perform routine H&E staining and immunohistochemical detection on the corresponding tissue sections, and use a full-field pathology scanner to digitally scan the H&E sections to obtain H&E full-field digital pathological images.

[0095] Step 2: Establish a three-tiered labeling system for the tumor microenvironment based on the quantitative results of immunohistochemical markers. The immunohistochemical markers include one or more of T cells, B cells, myeloid cells, macrophages, endothelial cells, and matrix-related markers. Based on microenvironmental characteristics such as immune cell infiltration, matrix enrichment, and immune desertification, samples are labeled as IE, Desert, or Stromal types.

[0096] Step 3: Establish a pairing relationship between H&E full-view digital pathology images and IHC microenvironment triadic labels, and use the IHC microenvironment triadic labels as slice-level supervision labels for training the H&E image model.

[0097] Step 4: Read the low-magnification thumbnail of H&EWSI, convert the RGB color space to the HSV color space, construct an initial tissue mask based on the saturation and lightness channels, and obtain the final tissue mask through morphological opening, closing, hole filling and small connected component removal.

[0098] Step 5: Within the area defined by the tissue mask, divide the image into blocks according to a window size of 256×256 pixels or 512×512 pixels, calculate the tissue proportion of each block, and remove blocks whose tissue proportion is lower than a preset threshold.

[0099] Step 6: Perform quality control on candidate tiles by calculating the mean brightness, variance brightness, mean saturation, and Laplacian variance, and remove tiles that are blank, blurred, folded, contaminated, have an excessive proportion of dead regions, or lack sufficient information.

[0100] Step 7: Optionally, the retained patches are color-normalized to reduce color differences caused by different dyeing batches and scanning equipment.

[0101] Step 8: Input the retained map patches into the UNI feature extractor for encoding to obtain the 1024-dimensional high-dimensional feature vectors corresponding to each map patch.

[0102] Step 9: Construct a package using all the tile features corresponding to each H&EWSI image, and use the tile features as multiple instances in the package to form a multi-instance learning input.

[0103] Step 10: Feed the multi-instance learning input into the CLAM-SB model. The model first obtains instance features through a feature transformation layer, then calculates the attention weight of each tile instance through a gated attention mechanism, and performs weighted aggregation of the tile features to obtain a slice-level representation.

[0104] Step 11: The classification output layer outputs the class probabilities of IE, Desert, and Stromal types based on the slice-level representation, and calculates the cross-entropy loss using the IHC microenvironment triad labels as the true labels.

[0105] Step 12: Select Top-K high-attention patches and Bottom-K low-attention patches based on attention weights, construct instance-level clustering constraints, and enhance the discrimination ability of key patches related to IHC microenvironment typing.

[0106] Step 13: The total model loss consists of slice-level cross-entropy loss, instance-level clustering loss, and L2 regularization term. The model parameters are updated through backpropagation algorithm.

[0107] Step 14: Use 5-fold cross-validation to complete model training, parameter selection, and robustness evaluation on the training set, and characterize the model performance using one-vs-restAUC, macro-AUC, accuracy, sensitivity, specificity, F1 score, and confusion matrix.

[0108] Step 15: Select the model that performs best during the training phase, validate it on the internal test set and in three independent external center validation queues, and output the classification probabilities, final classification results, attention heatmaps and model performance metrics for IE, Desert and Stromal types.

[0109] It should be noted that, in this embodiment, the supervision labels used for model training are derived from the tumor microenvironment trigenization results obtained based on the quantitative results of immunohistochemical markers, rather than labels directly derived from human visual interpretation of H&E images. The immunohistochemical marker combination preferably includes one or more of T cell, B cell, myeloid cell, macrophage, endothelial cell, and matrix-related markers. Further, the immunohistochemical markers may include one or more of CD3, CD4, CD8, CD20, CD45RO, CD14, CD68, CD163, FOXP3, CD31, α-SMA, Collagen, or other tumor microenvironment-related markers.

[0110] In one specific implementation, the training cohort samples are first subjected to immunohistochemical detection, and based on the degree of immune cell infiltration, the distribution of myeloid cells or macrophages, and quantitative characteristics of vascular and matrix-related components, the samples are classified into IE, Desert, and Stromal types. Subsequently, the IE, Desert, and Stromal types are used as slice-level supervisory labels, and a pairing relationship is established with the corresponding patient or corresponding slice's H&E full-view digital pathology image for training the H&E image prediction model.

[0111] This embodiment achieves cross-modal label transfer from the IHC microenvironment typing system to H&E digital pathology images using the aforementioned method. Specifically, IHC typing results provide biologically interpretable supervisory signals, while H&E images provide morphological input that can be obtained at low cost in routine pathology workflows. The weakly supervised multi-instance learning model learns the implicit correspondence between the two during training, that is, it identifies histological morphological patterns related to immune enrichment, immune desertification, and matrix enrichment states from H&E images.

[0112] Unlike conventional pathological image classification tasks, the triadic labels described in this invention are not simple pathological diagnosis categories, benign / malignant categories, or high / low prognostic risk categories. Instead, they originate from the tumor microenvironment state defined by immunohistochemical quantitative analysis. These labels lack clear pixel-level boundaries in H&E images, and related morphological cues are often scattered across the tumor parenchyma, immune cell infiltration areas, stroma, perivascular areas, and perinecrotic areas. Therefore, this invention employs a weakly supervised learning approach combining package-level supervision, instance-level attention scoring, and instance-level clustering constraints to automatically select the key regions most relevant to IHC microenvironment classification from a large number of image patches.

[0113] During training, the IHC microenvironment triadic labels corresponding to each H&E full-view digital pathology image are input into the model to supervised clustering-constrained attention multi-instance learning model to learn slice-level prediction capabilities. Preferably, the training set uses 5-fold cross-validation for model training and performance evaluation. After model training, independent validation is further performed in at least one independent external validation queue, preferably in three independent external center queues, to examine the model's generalization ability under different sample sources, centers, staining batches, and data distribution conditions.

[0114] S60: Generate attention heatmaps, key tile ranking results, or probability reports; where attention heatmaps are obtained based on attention weights, key tile ranking results are obtained by sorting in descending order based on the magnitude of attention weights, and probability reports are generated based on category probabilities.

[0115] In some embodiments, an exemplary attention heatmap generation method is provided. Specifically, to improve model interpretability, this invention uses tile attention weights... Generate an attention heatmap. Specifically, map the attention weight of each patch back to its spatial coordinates in WSI. And normalize it: in, For the first Normalized attention weights for each patch, Attention weights for all tiles The minimum value, The maximum value of the attention weights for all tiles. To prevent constants with a denominator of zero, normalize the attention weights. By overlaying the corresponding spatial locations onto the WSI thumbnail, an attention heatmap is obtained. This attention heatmap is used to highlight pathological areas that contribute significantly to the prediction of IE, Desert, or Stromal types.

[0116] In some embodiments, an exemplary probability report generation method is provided. Specifically, when the same patient has multiple H&EWSI slices, predictions are first performed for each slice separately. Let the first slice be... The probability of the class of slice is The slice weight is Then the patient-level (target-level) category probability is: in, The number of slides for this patient. It can be determined based on the effective tissue area of ​​the slice, image quality score, maximum prediction confidence, or attention region intensity, and must satisfy: The final patient-level classification results are as follows: in, The final comprehensive classification result of the tumor microenvironment is the target.

[0117] In some embodiments, the model output is divided into three pathological microenvironment subtypes, preferably including but not limited to the following three categories: (1) IE type (Immune-Enriched): characterized by abundant lymphocyte infiltration, obvious immune activation features, and active interaction between tumor and immune cells. (2) Desert type (immune desert type): characterized by a pathological image pattern with a high proportion of tumor parenchyma, sparse infiltration of immune cells, and insufficient signs of immune activation in the microenvironment. (3) Stromal type (stromal enrichment type): It is characterized by a pathological image pattern with abundant interstitial components, significant fibrosis, increased myeloid cells or macrophage-related areas, and prominent immunosuppressive microenvironment features.

[0118] like Figure 4 The image shown is a schematic diagram of representative pathological images of three pathological microenvironment subtypes provided in the embodiments of the present invention. Figure 4 The images also show H&E staining and CD3 immunohistochemical staining for each subtype, which are used to present the differences in tissue morphology and immune cell distribution among the different subtypes.

[0119] In H&E staining images of the Desert type (immune desert type), the tumor parenchyma accounts for a very high proportion, the overall tissue morphology is homogeneous, and no obvious lymphocyte aggregation areas are observed. The corresponding CD3 immunohistochemical staining images show an overall pale tone, with very few and sparsely distributed positive staining areas, indicating sparse T lymphocyte infiltration and insufficient signs of immune activation in the microenvironment.

[0120] In H&E staining images of the IE (immunoenriched) type, dense areas of lymphocyte infiltration are visible, appearing as dark purple clumps or cords, intertwined with tumor tissue, indicating an active interaction between the tumor and immune cells. Correspondingly, CD3 immunohistochemical staining images show numerous dark brown positive staining areas, widely distributed within and around the tumor tissue, indicating abundant T lymphocyte infiltration and significant immune activation.

[0121] In H&E staining images of the stromal type (stromal-enriched type), there is extensive proliferation of stromal components, exhibiting significant fibrotic changes. The tumor tissue is separated into irregular nest-like or cord-like structures by fibrous connective tissue. In the corresponding CD3 immunohistochemical staining images, the positive stained areas are mainly concentrated in the stromal region, showing a scattered or focal distribution. At the same time, a large number of matrix-related structures are visible, reflecting the characteristics of abundant stromal components and a prominent immunosuppressive microenvironment.

[0122] In some embodiments, an exemplary implementation flow of the method is provided, including the following steps: (1) Input the H&EWSI to be tested; (2) Perform tissue region detection, tile segmentation, and tile quality control; (3) Input the retained map tiles into the pre-trained feature extraction network to obtain the map tile feature set; (4) Input the set of map features into the trained clustering-constrained attention multi-instance learning model; (5) Output the predicted probabilities of IE, Desert and Stromal types. ; (6) The category with the highest predicted probability is taken as the final classification result; (7) Output attention heatmaps and high contribution patches to assist pathologists in review.

[0123] In one specific implementation, when this method is used for the triadic prediction of the pathological microenvironment on a single slide, the process includes: inputting the H&E full-view pathological slide of the patient to be tested; the system automatically completes tissue region extraction, patch segmentation, quality control, and UNI feature encoding; inputting the obtained patch feature set into the trained CLAM-SB triadic model to obtain probability scores for IE, Desert, and Stromal types; taking the category with the highest probability as the final triadic result of the pathological microenvironment for the slide; and simultaneously outputting attention heatmaps, high-contribution patches, or probability reports to assist pathologists in reviewing and interpreting the results.

[0124] In one specific implementation, when the method is used for patient-level multi-slice fusion prediction, the process includes: for the case of multiple pathological slices of the same patient, the three-type prediction can be performed on each slice separately, and then the final patient-level pathological microenvironment three-type result can be output by majority voting, probability weighted average, maximum confidence priority or other fusion strategies.

[0125] Preferably, when the prediction results of different slices are inconsistent, weighted fusion can be performed by combining the tumor area of ​​each slice, the intensity of the attention region, or the image quality score to improve the robustness of patient-level judgment.

[0126] In some embodiments, the model outputs three probability values ​​and a final predicted label for the input full-view pathological image. Model performance is primarily evaluated using the area under the receiver operating characteristic curve (ROC) for a pair of other classes, i.e., one-vs-restAUC, supplemented by macro-average AUC, accuracy, sensitivity, specificity, F1 score, and confusion matrix. Based on the model's prediction results, further outputs include attention heatmaps, key patch ranking results, patient-level comprehensive classification results, and explanatory results related to prognostic risk, recurrence risk, treatment response probability, or potential benefit of immunotherapy. The attention heatmap is used to display the pathological regions that contribute most to the three-part classification; the key patch ranking results assist pathologists in review; and the patient-level comprehensive classification results integrate information from multiple slides.

[0127] In some embodiments, in addition to being applied to the identification of pathological microenvironment subtypes in patients with small cell lung cancer, this method can also be used to assist in the assessment of overall survival, disease-free survival, recurrence risk or progression risk, screening of potential beneficiaries of immunotherapy, immunotherapy combined with chemotherapy or other systemic therapies, and as part of digital pathology-assisted diagnostic software to provide reference for clinical decision-making.

[0128] In some embodiments, an exemplary clinical application process based on typing results is provided, wherein the typing results can be obtained by the methods proposed in the above embodiments, including the following steps 1-3.

[0129] Step 1: Obtain the three-part prediction results of the patient's pathological microenvironment.

[0130] Step 2: Perform correlation analysis between the results and the patient's clinical data, staging information, treatment plan and follow-up outcome.

[0131] Step 3: Based on the different pathological microenvironment characteristics corresponding to IE, Desert, and Stromal types, patients are stratified for prognosis, relapse risk is assessed, and potential benefits of immunotherapy are indicated.

[0132] In a preferred embodiment, IE patients indicate a high background of immune activation and may be a potential beneficiary of immunotherapy; Desert patients indicate insufficient immune infiltration; and Stromal patients indicate a richer matrix or inhibitory microenvironment.

[0133] The feasibility and technological advancement of the method of the present invention will be further illustrated below with specific experimental embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0134] This experiment used formalin-fixed paraffin-embedded tissue samples from small cell lung cancer patients to obtain H&E-stained full-field digital pathological images and immunohistochemical microenvironment triadic labels for the corresponding samples. Each H&E full-field digital pathological image was paired with a unique immunohistochemical triadic label. The immunohistochemical triadic labels were obtained based on the quantitative analysis results of multiple tumor microenvironment-related immunohistochemical markers, such as CD3, CD4, CD8, CD68, and α-SMA, rather than from manual interpretation of H&E images.

[0135] According to the data organization rules of multi-instance learning, a single H&E full-view digital pathology image is constructed into a package, and each valid patch obtained from the image is constructed into an instance in the package for subsequent model training and validation.

[0136] This experiment was conducted on a Linux x86_64 operating system, using Python version 3.10 and the PyTorch deep learning framework. Key dependencies included torch, torchvision, timm, OpenSlide, openslide-python, h5py, opencv-python, scikit-learn, matplotlib, pandas, and numpy. All model training and inference processes were performed on a GPU server supporting CUDA acceleration.

[0137] This experiment describes the image preprocessing workflow and parameters. For all H&E full-view digital pathology images, the following operations were performed sequentially: tissue region detection, non-tissue background removal, patch segmentation, and patch quality control. Specific parameter settings are as follows: Patch segmentation parameters: patch_size=256 pixels, step_size=256 pixels (no overlapping segmentation), patch_level=0; Organizational region segmentation parameters: seg_level=-1, sthresh=8, mthresh=7, close=4, use_otsu=False; Quality filtering parameters: a_t=100, a_h=16, max_n_holes=8; The block outline determination method is four_pt, and the use_padding parameter is enabled.

[0138] This experiment describes the feature extraction process and parameters for image patches. A pre-trained pathological image encoder (UNI) is used to extract features from the pre-processed valid image patches. Specific parameter settings are as follows: input image patches are uniformly scaled to 224×224 pixels and normalized using the mean and standard deviation of the ImageNet dataset; feature extraction batch_size=256; each valid image patch outputs a 1024-dimensional feature vector; all patch features from each full-view digital pathological image are saved as h5 and pt format feature files respectively.

[0139] Explanation of the model training process and parameters. This experiment uses the clustering-constrained attention multi-instance learning model CLAM-SB for training, with specific parameter settings as follows: Model structure parameters: Model size is small, input feature dimension is 1024, hidden layer dimension is 512, attention layer dimension is 256, number of classification categories is 3, dropout=0.25; Loss function parameters: The packet-level loss uses cross-entropy loss, the instance-level loss uses support vector machine loss, the packet-level loss weight bag_weight=0.7, and the number of Top-K / Bottom-K instance samples B=8; Optimizer parameters: Adam optimizer is used, with an initial learning rate of 5×10⁻⁶. -5 Weight decay coefficient 1×10 -3 ; Training strategy parameters: maximum number of training rounds 200, early stopping mechanism enabled to prevent overfitting, and 5-fold cross-validation used for model training and performance evaluation.

[0140] Model performance validation results are as follows Figure 5 , Figure 6 and Figure 7 As shown.

[0141] like Figure 5 The diagram shows a schematic of a 5-fold cross-validation confusion matrix for the ternary subtypes of small cell lung cancer tumor microenvironment provided in an embodiment of the present invention. After each fold of the test set prediction is completed, the correspondence between the true labels and the predicted labels is statistically analyzed to generate a single-fold confusion matrix. The average AUC of the 5-fold test set is 0.8266 ± 0.0117.

[0142] The AUC for the first fold test set was 0.824, with an accuracy of 69.0%, and included 95 samples. The AUC for the second fold test set was 0.835, with an accuracy of 64.2%, and included 95 samples. The AUC for the third fold test set was 0.809, with an accuracy of 56.8%, and included 95 samples. The AUC for the fourth fold test set was 0.840, with an accuracy of 66.0%, and included 94 samples. The AUC for the fifth fold test set was 0.826, with an accuracy of 66.0%, and included 97 samples. Summarizing the prediction results across all test sets, there were 178 cases of the Desert type, of which 128 were correctly predicted; 130 cases of the Stromal type, of which 73 were correctly predicted; and 168 cases of the IE type, of which 97 were correctly predicted.

[0143] like Figure 6 The figure shows the ROC curves for the pathological microenvironment subtypes in the test set provided in this embodiment of the invention. Based on the three-class probability values ​​output by the model, the receiver operating characteristic curve and the corresponding area under the curve are calculated for each subtype using a pairwise parallel method.

[0144] The AUC for the IE type was 0.853, for the Desert type it was 0.802, and for the Stromal type it was 0.786. The overall classification performance metrics for the test set were: accuracy 0.626, macro-average F1 score 0.618, sensitivity 0.619, and specificity 0.813.

[0145] like Figure 7 The figure shows Kaplan-Meier survival curves for patients with different tumor microenvironment subtypes provided in this embodiment of the invention. Patients were divided into IE, Desert, and Stromal groups based on the tumor microenvironment subtypes predicted by the model. Survival curves were plotted using the Kaplan-Meier method, taking into account the patients' overall survival follow-up time and outcome status. The log-rank test was then used to compare the survival differences among the three groups.

[0146] The results showed a statistically significant difference in overall survival among the three groups (p-value of 0.0044 for log-rank test). The IE group had the best overall survival, the Desert group had the worst, and the Stromal group had an overall survival in between. This confirms that the tumor microenvironment subtype predicted by the method of this invention is significantly correlated with patient prognosis and can be used for prognostic stratification in clinical practice.

[0147] This invention also provides a tumor microenvironment triad prediction device for small cell lung cancer, used to implement the methods described in any of the above embodiments, such as... Figure 8 As shown, the device includes: The data acquisition module 801 is configured to acquire stained full-field digital pathological images of the target small cell lung cancer tumor tissue sample; The data preprocessing module 802 is configured to perform tissue region detection, non-tissue background removal, patch segmentation, patch quality control and / or color standardization on the full-view digital pathology image to obtain an effective patch set. Feature extraction module 803 is configured to input the effective set of map patches into a pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors; The instance feature construction module 804 is configured to construct multi-instance learning features by taking the full-view digital pathology image as a package and the multiple patch feature vectors as multiple instances in the package; The model prediction module 805 is configured to construct a clustering-constrained attention multi-instance learning model. This model responds to input multi-instance learning features by obtaining attention weights for each instance learning feature. Based on these attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on these slice-level representations, the model outputs the category probabilities and classification results for three tumor microenvironment subtypes: IE, Desert, and Stromal. The clustering-constrained attention multi-instance learning model is trained using the three-level classification results of small cell lung cancer tumor microenvironment obtained based on immunohistochemical marker quantification as a full-view digital pathology image-level supervised label. The prediction result generation module 806 is configured to generate an attention heatmap, a key tile ranking result, or a probability report; wherein the attention heatmap is obtained based on the attention weights, the key tile ranking result is obtained by sorting the key tiles in descending order based on the magnitude of the attention weights, and the probability report is generated based on the category probability.

[0148] 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.

[0149] 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.

[0150] 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.

[0151] The electronic device provided in this embodiment of the invention can be the terminal device described in the above embodiments.

[0152] 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 above-described embodiment for the three-level prediction method of tumor microenvironment in small cell lung cancer.

[0153] 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 three-type prediction method for tumor microenvironment in small cell lung cancer described in the above embodiments.

[0154] 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.

[0155] 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.

[0156] 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.

[0157] 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.

[0158] 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.

[0159] The memory may include high-speed RAM, and may also include non-volatile storage (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.

[0160] 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.

[0161] 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.

[0162] 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.

[0163] 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.

[0164] 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 three types of tumor microenvironment in small cell lung cancer, characterized in that, The method includes: Acquire stained full-field digital pathological images of target small cell lung cancer tumor tissue samples; The full-view digital pathological image is subjected to tissue region detection, non-tissue background removal, patch segmentation, patch quality control and / or color standardization to obtain an effective patch set. The effective set of map patches is input into a pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors; Using the full-view digital pathological image as a package, and the multiple image patch feature vectors as multiple instances in the package, a multi-instance learning feature is constructed. A clustering-constrained attention multi-instance learning model is constructed. This model responds to input multi-instance learning features and obtains attention weights for each instance learning feature. Based on these attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on these slice-level representations, the model outputs the category probabilities and classification results for three tumor microenvironment subtypes: IE, Desert, and Stromal. The clustering-constrained attention multi-instance learning model is trained using the three-level classification results of small cell lung cancer tumor microenvironment obtained based on immunohistochemical marker quantification as supervised labels for full-view digital pathology images. Generate an attention heatmap, a key tile ranking result, or a probability report; wherein the attention heatmap is obtained based on the attention weights, the key tile ranking result is obtained by sorting the key tiles in descending order based on the magnitude of the attention weights, and the probability report is generated based on the category probability.

2. The method according to claim 1, characterized in that, The full-view digital pathological image is subjected to tissue region detection, non-tissue background removal, patch segmentation, patch quality control, and / or color normalization to obtain an effective patch set, including: The low-magnification thumbnail of the full-view digital pathology image is converted from the RGB color space to the HSV color space. Based on the set saturation and brightness thresholds, the initial tissue mask is calculated using the following formula. : in, Indicates the position of image pixels. The saturation threshold, The brightness threshold. Indicates saturation. Indicates brightness. Indicates the initial tissue mask At image pixel location The value at; Based on the initial tissue mask The final tissue mask is calculated using the following formula. : in, For the removal of small connected components. For filling the holes, This is a morphological closing operation. Open operations for morphology. In the final tissue mask Within the corresponding tissue region, the full-view digital pathology image is segmented according to a preset tile size and step size, and a candidate tile set is obtained using the following formula. ,in, Indicates the first i One candidate patch, Indicates the total number of candidate tiles; For each candidate patch The proportion of its tissue area is calculated using the following formula. : in, Representing candidate plots The total number of pixels; when Remove the candidate tile at that time. ,when The candidate tile is retained at that time. , Preset a threshold for the proportion of organizational areas; For the retained candidate patches, the average brightness value is calculated using the following formula. Luminance variance and Laplace variance : in, Indicates the brightness channel value. Represents the Laplace operator. Used to evaluate tile sharpness This is the variance calculation function; Based on the average brightness of the retained candidate patches Luminance variance and Laplace variance Each tile is compared with its corresponding preset threshold, and tiles that do not meet the threshold requirements are removed to obtain the intermediate set of valid tiles. in, Indicates the first One valid tile, , Indicates the total number of valid tiles; For the valid set of tiles Color standardization is performed, and the resulting set of color-standardized tiles is used as the final valid set of tiles. ,in, Indicates the color-normalized first... If a valid tile is not color-normalized, then .

3. The method according to claim 2, characterized in that, The effective set of image patches is input into a pre-trained pathological image feature extraction network to obtain multiple image patch feature vectors, including: Each valid tile Input the pre-trained pathological image feature extraction network and calculate the patch feature vector using the following formula. : in, For pre-training the pathological image feature extraction network, UNI, CONCH, CTransPath, ResNet, ConvNeXt, or VisionTransformer were selected. For feature dimension, It is the space of real numbers; For the full-view digital pathology image, the feature vectors of all valid patches constitute the feature set of the image.

4. The method according to claim 3, characterized in that, The clustering-constrained attention multi-instance learning model responds to the input multi-instance learning features and outputs the category probabilities and classification results of three tumor microenvironment subtypes: IE, Desert, and Stromal, in the following manner: Patch feature vectors from input multi-instance learning features The transformed instance features are calculated using the following formula. : in, and These are the learnable parameters of the feature transformation layer. It is a non-linear activation function; Transformed instance features A gating attention mechanism is used to calculate the attention score using the following formula. : in, , and These are the learnable parameters for the attention scoring module. This represents element-wise multiplication. The hyperbolic tangent activation function is used. For activation functions; Based on the attention score Attention weights are calculated using the following formula. : in, , This represents the attention score of the k-th tile. It is the softmax function; Based on the attention weight The slice-level representation is calculated using the following formula. : Slice-level representation The input classification output layer calculates the three-class logical values ​​using the following formula. : in, , , , These are the logits values ​​for IE, Desert, and Stormal types, respectively. and Learnable parameters for the classification output layer; The predicted probabilities of the three microenvironment subtypes are calculated using the following formula: in, This indicates that the full-view digital pathology image belongs to the category. The predicted probability, Set of microenvironment subtypes Elements in; Based on the predicted probability, the final predicted category is obtained using the following formula. : Here, arg max is a function that takes the index corresponding to the maximum value.

5. The method according to claim 4, characterized in that, Generate attention heatmaps, key tile ranking results, or probability reports, including: The attention weight of each patch is mapped to its spatial coordinates in the full-view digital pathology image. The attention weights are normalized using the following formula: in, For the first Normalized attention weights for each patch, Attention weights for all tiles The minimum value, The maximum value of the attention weights for all tiles. To prevent constants with a denominator of zero; The normalized attention weights are superimposed on the thumbnail of the full-view digital pathology image according to their corresponding spatial locations to obtain an attention heatmap, which is used to indicate the pathological region that contributes the most to the triad determination. When the same target has multiple full-view digital pathology images, prediction is performed for each full-view digital pathology image separately, and the prediction is made on the first... The category probability of Zhang Quan's digital pathology image is Image weights are In this case, the target-level category probability is calculated using the following formula: in, The total number of full-view digital pathology images targeting [the target]. According to the Determining the effective tissue area, image quality score, maximum prediction confidence, or attention region intensity of Zhang Quan's digital pathology images; The target-level classification result is obtained using the following formula: in, The final comprehensive classification result of the tumor microenvironment is the target.

6. The method according to any one of claims 1 to 5, characterized in that, The clustering-constrained attention multi-instance learning model is trained in the following manner: Acquire the full-view digital pathology image corresponding to each target in the training queue; Acquire immunohistochemical microenvironment triad labels corresponding to the full-view digital pathology images used for training, the labels including IE type, Desert type and Stromal type; Tissue region detection, patch segmentation, patch quality control, and color standardization were performed on the full-view digital pathology images used for training. A pre-trained pathological image feature extraction network was used to encode the image patches to obtain the patch feature matrix; Each training full-view digital pathology image is constructed as a multi-instance learning package and combined with the corresponding immunohistochemical microenvironment label to form training samples. The training samples are input into the clustering-constrained attention multi-instance learning model, and the overall model loss, including slice-level cross-entropy loss and instance-level clustering loss, is calculated. The model parameters are updated using the backpropagation algorithm; Five-fold cross-validation was used for model training, parameter selection, and robustness evaluation in the training set. The model with the best validation performance was selected as the trained clustering-constrained attention multi-instance learning model.

7. The method according to claim 6, characterized in that, The training samples are input into the clustering-constrained attention multi-instance learning model, and the overall model loss, including slice-level cross-entropy loss and instance-level clustering loss, is calculated, including: For each training session, use a multi-instance learning package. According to attention weight The candidate positive instance sets are obtained using the following formulas. and candidate negative instance set : in, Before selection Operations for finding the maximum value, After selection Operations to find the minimum value The preset number of instances; Input candidate positive instances and candidate negative instances into the instance-level discriminator. The instance-level prediction results are obtained using the following formula; in, For instance-level discriminators, For the learnable parameters of the instance-level discriminator, For the first The predicted probability that an instance is a positive instance; The instance-level clustering loss is calculated using the following formula. : in, It is the natural logarithm function; For the training set j For each sample, the slice-level cross-entropy loss is calculated using the following formula. : in, The one-hot encoding of the true label of the j-th sample. The full-view digital pathology images used for model prediction training belong to the category. c The probability of; The overall loss of the model is calculated using the following formula. : in, For instance-level loss weights, The L2 regularization coefficient is... This represents the set of learnable parameters of the model.

8. A tumor microenvironment triad prediction device for small cell lung cancer, characterized in that, The device includes: The data acquisition module is configured to acquire stained full-field digital pathological images of target small cell lung cancer tumor tissue samples; The data preprocessing module is configured to perform tissue region detection, non-tissue background removal, patch segmentation, patch quality control and / or color standardization on the full-view digital pathology image to obtain an effective patch set. The feature extraction module is configured to input the effective set of map patches into a pre-trained pathological image feature extraction network to obtain multiple map patch feature vectors; The instance feature construction module is configured to construct multi-instance learning features by taking the full-view digital pathology image as a package and the multiple patch feature vectors as multiple instances in the package; The model prediction module is configured to construct a clustering-constrained attention multi-instance learning model. This model responds to input multi-instance learning features by obtaining attention weights for each instance learning feature. Based on these attention weights, multiple patch feature vectors are weighted and aggregated to obtain slice-level representations. Based on these slice-level representations, the model outputs the category probabilities and classification results for three tumor microenvironment subtypes: IE, Desert, and Stromal. The clustering-constrained attention multi-instance learning model is trained using the three-level classification results of small cell lung cancer tumor microenvironment obtained based on immunohistochemical marker quantification as a full-view digital pathology image-level supervised label. The prediction result generation module is configured to generate an attention heatmap, a key tile ranking result, or a probability report; wherein the attention heatmap is obtained based on the attention weights, the key tile ranking result is obtained by sorting the key tiles in descending order based on the magnitude of the attention weights, and the probability report is generated based on the category 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 tumor microenvironment triad prediction method for small cell lung cancer 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 tumor microenvironment triad prediction method for small cell lung cancer as described in any one of claims 1-7.

Images