Cancer immunotherapy prognosis prediction method, system and equipment based on CD8 + T cell density in necrotic area and medium

Abstract

The invention relates to the technical field of medical data processing, and discloses a cancer immunotherapy prognosis prediction method, system and equipment based on CD8 + T cell density in a necrotic area and a medium. The method comprises the following steps: firstly, extracting pathological features on a CD8 immunohistochemical staining (IHC) slice image of tumor tissue of a to-be-detected person after neoadjuvant immunotherapy by using a multi-scale feature extraction model, and then carrying out regional division on the CD8 IHC slice image by using a regional classification model to obtain a necrotic region; then, the CD8 + T cells in the necrotic area are recognized and counted through the target recognition model, the cell density of the CD8 + T cells in the necrotic area is rapidly obtained, and finally, the EFS of the person to be detected is accurately predicted based on the cell density of the CD8 + T cells in the necrotic area. According to the method, efficient and high-precision EFS prediction can be carried out only through the CD8IHC slice image of the tumor tissue, and a doctor can make a proper diagnosis and treatment plan for a patient according to a prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, and particularly relates to a cancer immunotherapy prognosis prediction method and system based on CD8 + T cell density in a necrotic area, a cancer immunotherapy prognosis prediction system, a cancer immunotherapy prognosis prediction device and a cancer immunotherapy prognosis prediction medium. BACKGROUND

[0002] Globally, there are nearly 10 million malignant tumor death cases every year, which is one of the main diseases causing human death. In the past few years, immunotherapy has completely changed the treatment method of tumors, and neoadjuvant immunotherapy has become the main treatment choice for patients with lung cancer, melanoma, liver cancer, etc. Unlike the traditional method of first surgery and then immunotherapy, neoadjuvant immunotherapy is to kill tumors through immunotherapy combined with chemotherapy and then perform surgery. Studies have found that neoadjuvant immunotherapy can help the immune system recognize tumors and enhance systemic anti-tumor immunity, and it is the most promising tumor treatment plan at present. It usually takes 5-10 years to judge new anti-tumor drugs or treatment methods, and follow-up of patients is needed to determine the tumor recurrence probability and the progression-free survival of patients. Since neoadjuvant immunotherapy occurs before surgery, it means that the tumor sample removed by surgery can already reflect the efficacy of the drug. Therefore, by developing a scientific method for accurately predicting the prognosis of patients through the surgical sample after neoadjuvant therapy, the time and cost of new drug clinical trials can be greatly shortened, and it is more likely to provide individualized intervention suggestions based on the recurrence risk of patients, which has very important clinical significance.

[0003] The current international standard for clinical solid tumor efficacy evaluation is RECIST 1.1 (Response Evaluation Criteria in Solid Tumors) based on imaging and irPRC (Immune-related Pathologic Response Criteria) based on pathology. The imaging response of RECIST criteria has been used as a surrogate indicator of long-term clinical benefit for patients with advanced unresectable tumors, and this standard also provides treatment guidance for clinical oncologists. However, the imaging-pathology inconsistency phenomenon that occurs in the field of neoadjuvant therapy highlights its application limitations. In some studies, it was found that the current RECIST criteria were not related to the event-free survival (EFS) of patients and could not predict the risk of recurrence of patients. The pathology criteria irPRC can predict efficacy by the proportion of residual tumor volume (RVT) in tumor resection samples within 2 months after treatment initiation, and is the most promising prognostic evaluation indicator in neoadjuvant therapy, which is expected to replace RECIST criteria as an alternative endpoint indicator for regulatory approval, and can also provide the basis for adjuvant therapy decision-making. irPRC determines drug efficacy by detecting the proportion of residual tumor (RVT) in the tumor bed through pathology, and usually uses no tumor cell residue (complete pathological remission, pCR), 10% RVT (major pathological remission, MPR) and more than 10% RVT (non-major pathological remission, non-MPR) as the judgment standard. However, the accuracy of irPRC is still not satisfactory: the proportion of residual tumor cannot further distinguish patients with high recurrence risk from those with low recurrence risk; more than 1 / 3 of patients achieve pathological complete remission, and more than 20% of patients still relapse within three years; and in the non-MPR group of patients, more than 40% of patients have not relapsed. The current threshold cannot provide more accurate prognostic judgment and potential adaptive treatment strategies for patients.

[0004] In addition, the assessment of tumor-infiltrating lymphocytes (TIL) in tumor tissue before treatment can predict the sensitivity of patients to immunotherapy drugs to some extent, but due to the activation of the immune system caused by neoadjuvant immunotherapy, TILs are abundant in most post-treatment patient tissues, and TILs have lost their prognostic value. In scientific research, new detection methods such as proteomics and single-cell sequencing are used to judge the prognosis of patients. However, due to high cost and complex analysis procedures, it is almost impossible to have clinical applications in a short period of time.

[0005] Therefore, there is an urgent need in the clinic for a method that can more accurately predict the prognosis of neoadjuvant immunotherapy, and the method needs to be low-cost, standardized and easy to operate in order to promote clinical application. SUMMARY

[0006] This invention provides a method based on CD8 within necrotic regions. + Methods, systems, devices, and media for predicting the prognosis of cancer immunotherapy based on T-cell density, aiming to address the current lack of methods capable of accurately predicting the prognosis of neoadjuvant immunotherapy.

[0007] This invention provides a method based on CD8 in necrotic regions + Methods for predicting the prognosis of cancer immunotherapy based on T-cell density include:

[0008] Obtain CD8 IHC (immunohistochemical staining) slice images of tumor tissue from subjects who have undergone neoadjuvant immunotherapy, where the subjects are cancer patients;

[0009] Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using a multi-scale feature extraction model.

[0010] Based on the pathological features on CD8 IHC slice images, the necrotic areas in CD8 IHC slice images are obtained through a region classification model.

[0011] Using a target recognition model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0012] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density predicts the EFS (event-free survival) of the subject.

[0013] The present invention provides a method based on CD8 in necrotic regions + A method for predicting the prognosis of cancer immunotherapy based on T-cell density, wherein the pathological features on CD8 IHC slice images of tumor tissue after neoadjuvant immunotherapy are extracted using a multi-scale feature extraction model, including:

[0014] Based on CD8 IHC slice images of tumor tissue from subjects after neoadjuvant immunotherapy, pathological features on CD8 IHC slice images were extracted using the HVIT model (HierarchicalVision Transformer is a deep learning model that combines the visual Transformer (ViT) architecture with the characteristics of the tumor hypoxic microenvironment).

[0015] In an embodiment, the HVIT model is first used to learn features of CD8 IHC slice images of tumor tissues of a cancer patient population after neoadjuvant immunotherapy, and then the HVIT model is fine-tuned using low-rank adaptive (LORA) to obtain a trained HVIT model for extracting pathological features on a CD8 IHC slice image of a to-be-tested person.

[0016] According to the application, a cancer immunotherapy prognosis prediction method based on CD8 + T cell density in a necrotic area is provided, which comprises:

[0017] Unsupervised clustering is performed on the pathological features on the CD8 IHC slice image to divide the CD8 IHC slice image into a plurality of sub-regions.

[0018] The plurality of sub-regions on the CD8 IHC slice image are divided into tumor regions, necrotic regions and interstitial regions by using a region classification model.

[0019] According to the application, a cancer immunotherapy prognosis prediction method based on CD8 + T cell density in a necrotic area is provided, which comprises:

[0020] The KNN clustering method is used to perform unsupervised clustering on the pathological features on the CD8 IHC slice image to divide the CD8 IHC slice image into 10 sub-regions.

[0021] According to the application, a cancer immunotherapy prognosis prediction method based on CD8 + T cell density in a necrotic area is provided, which comprises:

[0022] The ResNet101 model (a deep convolutional neural network with residual connection) is used to divide the 10 sub-regions on the CD8 IHC slice image into tumor regions, necrotic regions and interstitial regions, and obtain a mask of the necrotic region on the CD8 IHC slice image.

[0023] According to the mask of the necrotic region on the CD8 IHC slice image, the area of the necrotic region is obtained.

[0024] In one embodiment, the ResNet101 model is first trained to learn the features of IHC slice sample images labeled with tumor regions, necrotic regions, and stromal regions. In terms of model architecture, a transfer learning strategy is adopted, the network parameters are initialized with ImageNet pre-trained weights, and the network layers are fine-tuned through progressive unfreezing technique.

[0025] The present invention provides a method based on CD8 in necrotic regions + A method for predicting the prognosis of cancer immunotherapy based on T-cell density, which utilizes a target recognition model to analyze necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density includes:

[0026] Using the YOLOv11 model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used to identify CD8+ in the necrotic region of the CD8+ IHC slice image. + The number of T cells;

[0027] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + The number of T cells and the area of ​​necrosis were used to obtain the CD8+ in the necrotic area of ​​the CD8+ IHC slice image. + T cell density.

[0028] The present invention provides a method based on CD8 in necrotic regions + A method for predicting the prognosis of cancer immunotherapy based on T-cell density, wherein the CD8+ density is determined by the CD8+ necrotic region in CD8+ IHC slice images. + T cell density, used to predict the EFS of a subject, includes:

[0029] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + Based on T cell density and a preset threshold, the subject was identified as a CD8+ individual. + T cell hyperexpression or CD8+ + Individuals with low T-cell expression;

[0030] When the test subject is identified as CD8 + When individuals with high T-cell expression are identified, the EFS prediction result of the test subject is determined to be low risk of relapse.

[0031] When the test subject is identified as CD8 + When T cell counts are high or low, the EFS prediction result of the test subject is determined to be of high risk of relapse.

[0032] The present invention also provides a method based on CD8 in necrotic regions. + A cancer immunotherapy prognostic prediction system based on T-cell density includes:

[0033] The CD8 IHC slice image acquisition module is used to acquire CD8 IHC slice images of tumor tissue from a subject after neoadjuvant immunotherapy, wherein the subject is a cancer patient.

[0034] The pathological feature extraction module is used to extract pathological features from CD8 IHC slice images of tumor tissue after neoadjuvant immunotherapy in the subject, using a multi-scale feature extraction model.

[0035] The necrotic region acquisition module is used to: obtain the necrotic region in the CD8 IHC slice image based on the pathological features on the CD8 IHC slice image through a region classification model;

[0036] The cell recognition module is used to: utilize a target recognition model to perform CD8 identification on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0037] The EFS prediction module is used to: predict the CD8+ of necrotic regions in CD8 IHC slice images. + T cell density predicts the EFS of the subject.

[0038] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor, when executing the computer program, implements any of the above-described methods based on CD8 within necrotic regions. + A method for predicting the prognosis of cancer immunotherapy based on T cell density.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the CD8 based on the necrotic region described above. + A method for predicting the prognosis of cancer immunotherapy based on T cell density.

[0040] The present invention also provides a computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is capable of executing any of the above-described methods based on CD8 within necrotic regions. + A method for predicting the prognosis of cancer immunotherapy based on T cell density.

[0041] This invention provides a method based on CD8 in necrotic regions + This paper describes a method, system, equipment, and medium for predicting the prognosis of cancer immunotherapy based on T cell density. First, a multi-scale feature extraction model is used to extract pathological features from CD8 IHC slices of tumor tissue after neoadjuvant immunotherapy. Then, a region classification model is used to cluster the pathological features on the CD8 IHC slices to achieve accurate region segmentation and identify necrotic regions. Finally, a target recognition model is used to analyze the CD8 cells in the necrotic regions. + T cells recognize and count the CD8+ cells in the necrotic area, quickly obtaining the CD8+. + T cell density, ultimately based on CD8+ in necrotic areas. + T cell density accurately predicts the EFS of the subject. This invention only requires CD8 IHC slices of tumor tissue after neoadjuvant immunotherapy for efficient and high-precision EFS prediction, which can greatly improve the work efficiency of physicians and help them formulate appropriate treatment plans for patients based on the prediction results. Moreover, this method is low-cost, easy to operate, and facilitates its rapid promotion in clinical application. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This invention provides a method based on CD8 in necrotic regions. + One of the flowcharts for a method to predict the prognosis of cancer immunotherapy based on T cell density.

[0044] Figure 2 This invention provides a method based on CD8 in necrotic regions. + The second flowchart illustrates the prognostic prediction method for cancer immunotherapy based on T cell density.

[0045] Figure 3 This invention provides a method based on CD8 in necrotic regions. + A schematic diagram of the structure of a cancer immunotherapy prognosis prediction system based on T cell density.

[0046] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0048] The histological characteristics of tumor necrosis include the absence of viable cells within the necrotic area, loss of tissue structural details, and the presence of amorphous cellular debris. Traditionally, spontaneous tumor necrosis is associated with tumor progression, suppressed immune responses, and poor prognosis. However, neoadjuvant immunotherapy, by enhancing the immune system and initiating an anti-tumor response, induces immune-mediated death of tumor cells. Unlike spontaneous tumor death, immunotherapy-induced tumor death may indicate a good therapeutic effect. TILs (tumor lymphocytes) are key to tumor killing in immune-mediated tumor death; however, as described in the background section, most patient tissues after treatment are filled with TILs, thus TILs cannot serve as a prognostic marker. Interestingly, after tumor cell elimination, T lymphocytes undergo apoptosis or death. Although lymphocytes in necrotic tumor areas (nTILs, TILs in necrotic tumor areas) lose their cellular structure and are not visible in routine histological examination (HE), they still retain common T-cell specific antigens, allowing them to be detected by immunohistochemistry (IHC). This invention suggests that these T cells that actually kill tumor cells and then die may represent the effectiveness of immunotherapy drugs, tumor killing efficiency, and predict patient prognosis.

[0049] This invention, by integrating prognostic data from 200 NSCLC patients at a medical center who received neoadjuvant PD-1 immunotherapy combined with chemotherapy, found that traditional methods of detecting the number of surviving TILs after neoadjuvant immunotherapy cannot predict patient prognosis, while CD8+ in necrotic areas... + nTIL density was strongly associated with event-free survival (HR = 0.08, p < 0.01). More importantly, the study also showed that nTIL density was significantly improved based on CD8+. + The prognostic methods for nTIL cancer immunotherapy were compared with the currently widely used clinical imaging RECIST 1.1 criteria and pathological irPRC criteria, CD8 +nTIL's prognostic prediction method for cancer immunotherapy has the highest accuracy, even accurately assessing the risk of recurrence in patient groups for whom prognosis cannot be predicted according to the pathological irPRC criteria. Therefore, it demonstrates that CD8... + nTIL can be used alone as a core biomarker for assessing the prognosis of neoadjuvant immunotherapy, or in combination with the pathological irPRC criteria. Meanwhile, routine CD8 IHC staining is a standardized clinical testing method; it only requires selecting necrotic area sections for staining and reading to achieve standardized CD8-based biomarkers. + Predicting the relapse risk in nTIL patients can be rapidly adopted in clinical practice.

[0050] Figure 1 and Figure 2 This invention provides a method based on CD8 in necrotic regions. + A flowchart illustrating a method for predicting the prognosis of cancer immunotherapy based on T cell density. This invention provides a method based on CD8+ within necrotic regions. + The method for predicting the prognosis of cancer immunotherapy based on T-cell density can be implemented by any applicable terminal-side device or network-side device, such as based on CD8+ within necrotic regions. + Devices for predicting the prognosis of cancer immunotherapy based on T-cell density, etc.

[0051] See Figure 1 and 2 This invention provides a method based on CD8 in necrotic regions. + Methods for predicting the prognostic effects of cancer immunotherapy based on T-cell density may include:

[0052] S110. Obtain IHC (immunohistochemical staining) slide images of tumor tissue from the subject after neoadjuvant immunotherapy. The subject is a cancer patient, which may be a non-small cell lung cancer (NSCLC) patient who has received PD-1 inhibitor combined with chemotherapy. The cancer type may also be various subtypes of NSCLC, such as squamous cell carcinoma of the lung (LUSC) and adenocarcinoma of the lung (LUAD).

[0053] CD8 IHC slide images can be obtained by physicians through fixation, paraffin embedding, sectioning (5μm thickness), immunohistochemical staining (using commercially available anti-CD8 monoclonal antibodies for standard IHC procedure staining (antigen retrieval, blocking, incubation with primary antibody, incubation with secondary antibody, DAB staining)), and scanning (40x magnification, 0.25μm / pixel resolution) of tumor tissue after neoadjuvant immunotherapy.

[0054] S120. Based on the CD8 IHC slice images of the tumor tissue of the subject after neoadjuvant immunotherapy, the pathological features on the CD8 IHC slice images are extracted using a multi-scale feature extraction model (in this embodiment, 4K-dimensional feature vectors).

[0055] In one embodiment, the multi-scale feature extraction model is a pre-trained HVIT model. It can be an HVIT model that has already undergone extensive feature learning on publicly available datasets. The HVIT model is then used to learn features from IHC slice images of tumor tissues from cancer patients who have undergone neoadjuvant immunotherapy, adapting it to this type of data. Low-rank adaptive (LORA) is then used to fine-tune the HVIT model to improve its running efficiency and reduce the computational resources required by the model, resulting in a trained HVIT model used to extract pathological features from CD8 IHC slice images of the subjects.

[0056] The core value of the HVIT model lies in its ability to learn and encapsulate rich and general-purpose pathological image feature representations. During training, it incorporates tumor regions, cell morphology, tissue structure, and key biomarkers, then encapsulates the prediction results into a high-dimensional vector feature file. This high-dimensional feature vector file is a quantitative compressed representation of the visual semantics of IHC images, containing semantic features such as the spatial distribution of immune / stromal / tumor cells, regional heterogeneity of biomarker expression, structural relationships between microenvironment components, and tissue characteristics. When pathologists read the pathological features contained in these vector representations, problems such as long reading times and difficulty in judging details arise.

[0057] In this embodiment, HVIT is used for model generalization. CD8 IHC slice images contain a large number of pathological features, and different IHC images may come from different parts of the patient or the patient may be in different pathological states. Therefore, the pathological features of each IHC image are also different, making the pathological features of CD8 IHC slice images extremely complex. Therefore, training the generalization model on a large amount of data makes the model highly applicable and can be used to process IHC images of different patients. The region classification model used next is a high-dimensional vector representation generated by HVIT, which is used to accurately identify residual tumor, stroma, and necrotic areas.

[0058] S130. Based on the pathological features on the CD8 IHC slice images, the necrotic areas in the CD8 IHC slice images are obtained through a region classification model.

[0059] In one embodiment, S130 may include:

[0060] KNN clustering was used to perform unsupervised clustering on pathological features on CD8 IHC slice images (method and parameters: Gaussian blur was used for noise reduction, then PCA was used for dimensionality reduction, n_components=0.99 to retain 99% of the original data variance, cluster=10), dividing the CD8 IHC slice images into 10 sub-regions.

[0061] Using the ResNet101 model, the 10 subdivision regions on the CD8 IHC slice image are divided into tumor region, necrotic region and stroma region, and the mask of the necrotic region on the CD8 IHC slice image is obtained.

[0062] The area of ​​the necrotic region is obtained by using the mask of the necrotic region on the CD8 IHC slice image.

[0063] In one embodiment, the region classification model is a pre-trained ResNet101 model. During training, the ResNet101 model first learns the features of IHC slice sample images labeled with tumor regions, necrotic regions, and stromal regions. In terms of model architecture, a transfer learning strategy is adopted, the network parameters are initialized with ImageNet pre-trained weights, and the network layers are fine-tuned through progressive unfreezing technology.

[0064] Because the aforementioned high-dimensional vector representation contains rich and complex pathological features of CD8 IHC slice images, and the pathological features contained in different IHC slices are also different, it is impossible to formulate a unified classification standard for each specific pathological feature. In order to reduce the complexity of subsequent use of the ResNet model, unsupervised clustering of this high-dimensional vector representation is required. Specifically, this embodiment uses the KNN algorithm to perform unsupervised clustering of these high-dimensional vector features. First, the size of the smoothing filter (Gaussian blur) is used for noise reduction, and then PCA is used for dimensionality reduction. Setting n_components = 0.99 retains 99% of the original data variance. Extensive experiments have shown that when the clustering is 10 clusters, it can very accurately segment the subtypes of tumor regions, necrotic regions, and stromal regions, while saving a lot of computational resources. The clustering in this embodiment reveals 10 unique patterns in the HVIT feature space, but these patterns essentially correspond to the morphological substructures within the three classic tissue partitions: tumor regions, stromal regions, and necrotic regions. These 10 unique patterns are essentially subtypes of each region. For example, cluster 1: solid type; cluster 7: glandular type (different clusters on each pathological image will vary depending on the IHC image, so the ResNet model will be used later to classify them into three categories (tumor region, stroma region and necrotic region).

[0065] The ResNet101 model (a deep convolutional neural network that incorporates residual connections) was used to classify the 10 classes of KNN unsupervised clustering into three categories: tumor region, stromal region, and necrotic region. As can be seen from the KNN clustering algorithm above, the high-dimensional vector representation generated by the HVIT model contains a large number of complex pathological features from IHC slice images. Moreover, the pathological features contained in different CD8 IHC slices are also different, making it impossible to formulate a unified classification standard for each specific pathological feature. This makes it impossible to directly use the high-dimensional vector representation generated by the HVIT model for ResNet. Therefore, the high-dimensional vector representation generated by the HVIT model must first be simplified into 10 clusters using the KNN unsupervised clustering algorithm. According to pathology and pathological images, there is a big difference between necrotic and tumor areas and stromal areas. This invention uses a clever binary tree approach here. First, the necrotic area is distinguished from the "other" area because the necrotic area has no cellular structure and can be directly distinguished. Then, the binary tree approach is used again for the "other" area because the tumor cell nuclei in the tumor area are very large and dense, and the tumor area can be directly distinguished. Then, the remaining areas are classified as stromal areas. In this way, the labels of tumor area, stromal area and necrotic area can be assigned to these 10 clusters. Because the stromal region contains complex structures such as blood vessels and lymph nodes, it cannot be directly quantified. Therefore, we use a binary tree elimination method to first distinguish the simplest necrotic region. Then, the remaining part can be easily distinguished as the tumor region (tumor cell nuclei are very large and dense). Finally, what remains is the stromal region. After the pathologist labels these regions, we use the ResNet model for training. Therefore, the ResNet101 model only needs to focus on the necrotic region and the tumor region. The remaining very complex stromal region does not need to be learned. This solves both the problem of dividing into three regions and the problem of the difficulty in labeling the stromal region.

[0066] S140. Using a target recognition model, perform CD8 analysis on the necrotic region in the CD8 IHC slice image. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density.

[0067] In one embodiment, S140 may include:

[0068] Using the YOLOv11 model (parameters: confidence threshold = 0.7, anchor frame size optimized), CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used to identify CD8+ in the necrotic region of the CD8+ IHC slice image. +The number of T cells is shown in this embodiment using confusion matrix results, demonstrating the effectiveness of the YOLOv11 model in recognizing CD8+. + The accuracy rate in T cells was as high as 97.2%, while the error rate was less than 0.83%.

[0069] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + The number of T cells and the area of ​​necrosis were used to obtain the CD8+ in the necrotic area of ​​the CD8+ IHC slice image. + T cell density.

[0070] After using the ResNet101 model for region segmentation, a mask corresponding to each region in the original image can be obtained. Using the mask, the actual scanning resolution of the corresponding region can be extracted separately, and the actual region area can be calculated based on the image information of the IHC pathological image (image resolution of the scanned area and slice size of the scanned area). Then, based on the CD8 of the necrotic region in the CD8 IHC slice image... + By removing the number of T cells to calculate the actual area, the CD8+ region in the necrotic area of ​​the CD8+ IHC slice image can be obtained. + T cell density can be calculated by masking the necrotic area using a microscope. For example, if the necrotic area in patient A has x pixels, and the side length of each pixel is 0.25 micrometers under 40x magnification, the actual size of the area is x * 0.25 * 0.25 square micrometers. Then, dividing the number of CD8+ T cells (n) obtained by this area yields the corresponding density: n / (x * 0.25 * 0.25 * 0.000001) cells / mm². 2 .

[0071] S150, Based on the CD8+ necrotic region in the CD8 IHC slice image... + T cell density predicts the EFS of the subject.

[0072] In one embodiment, S150 may include:

[0073] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density, combined with a preset threshold (0.41 cells / mm² in this example). 2 (Determined through ROC analysis) The subject was identified as CD8. + T cell hyperexpression (greater than 0.41 cells / mm²) 2 ) or CD8 + Individuals with low T cell expression (less than 0.41 cells / mm²) 2 );

[0074] When the test subject is identified as CD8 + When individuals with high T-cell expression are identified, the EFS prediction result of the test subject is determined to be low risk of relapse.

[0075] When the test subject is identified as CD8 + When T cell counts are high or low, the EFS prediction result of the test subject is determined to be of high risk of relapse.

[0076] This study followed up the prognostic outcomes of 200 non-small cell lung cancer patients after neoadjuvant immunotherapy, collecting recurrence / death information and obtaining event-free survival data. The ethical approval numbers were: Beijing Chest Hospital LW-2024-022, Shengjing Hospital of China Medical University 2024PS1588K, and Hebei University Affiliated Hospital HDFYLL-2024-188. Simultaneously, CD8 IHC staining was performed on 99 patients with necrotic areas, revealing CD8+ within the necrotic areas. + The density of tumor-infiltrating T cells (nTILs) in necrotic areas is closely related to the patient's event-free survival (EFS), and its predictive accuracy for prognosis is superior to that of currently used pathological and imaging responses.

[0077] This embodiment uses the patient's CD8... + The median nTIL was 0.41 cells / mm. 2 Patients were divided into CD8 groups. + nTIL high-density and low-density groups. CD8 + Patients in the high-density nTIL group had significantly better prognoses compared to the low-density group. More importantly, currently used pathological response prediction criteria based on residual outcome transit (RVT) categorize patients into PCR, MPR, and non-MPR groups, but RVT has low predictive value within each group, failing to further differentiate patients' recurrence risk. Using CD8... + nTIL density can be used in conjunction with pathological response prediction criteria to further improve prediction accuracy; it can also be used alone, using only CD8. + When using nTIL density to predict patient prognosis, it has achieved very high predictive accuracy.

[0078] This invention provides a method based on CD8 in necrotic regions +A method for predicting cancer immunotherapy prognosis based on T cell density first uses a multi-scale feature extraction model to extract pathological features from CD8 IHC slices of tumor tissue after neoadjuvant immunotherapy. Then, a region classification model is used to cluster the pathological features on the CD8 IHC slices to achieve precise region segmentation and identify necrotic areas. Finally, a target recognition model is used to analyze the CD8 cells in the necrotic areas. + T cells recognize and count the CD8+ cells in the necrotic area, quickly obtaining the CD8+. + T cell density, ultimately based on CD8+ in necrotic areas. + T cell density accurately predicts the EFS of the subject. This invention only requires CD8 IHC slice images of tumor tissue after neoadjuvant immunotherapy for efficient and high-precision EFS prediction, which can greatly improve the work efficiency of physicians and help them formulate appropriate treatment plans for patients based on the prediction results. Moreover, this method is low-cost, easy to operate, and facilitates rapid promotion of clinical application.

[0079] This embodiment employs a multi-stage deep learning framework to achieve accurate region segmentation of tumor tissue pathological images. Its core technical approach involves the cascaded processing of three key modules. First, a pre-trained HVIT (Hierarchical Vision Transformer) model acquired from the Hugging Face platform is used as the feature extractor. During the training phase, this model integrates a large-scale multi-source pathological image dataset, including the TCGA database, covering over 50,000 immunohistochemical (IHC) staining images from different anatomical sites (such as the breast, lung, and colon). Through a self-supervised learning strategy, the model establishes a powerful feature representation capability during training, enabling it to extract high-order pathological feature vectors containing 4K dimensions from the input images. These features not only encode conventional cell morphology information but, more importantly, capture the spatial association patterns between hypoxic regions (HIF-1α high-expression areas) and key biological processes such as immune cell infiltration and angiogenesis within the tumor microenvironment. Notably, due to significant tissue structural differences in IHC images from different anatomical sites, this embodiment specifically strengthens cross-domain adaptability during model training. By using a pre-trained model trained on a large amount of data, it ensures that the extracted features have cross-tissue generalization performance.

[0080] In the feature processing stage, this embodiment innovatively introduces the KNN clustering algorithm as an intermediate processing module. This design is mainly based on two important findings: First, direct end-to-end classification of high-dimensional features extracted by HVIT suffers from uneven feature space distribution, with images from different tissue sources exhibiting obvious regional clustering in the feature space; second, through systematic hyperparameter optimization experiments, it was determined that dividing the features into 10 subtypes achieves the optimal clustering effect while maintaining biological significance. Specifically, when the number of clusters K=10, the segmented regions can be correctly classified as tumor regions, stromal regions, and necrotic regions. Increasing the K value to 15 further improves performance by less than 0.5% but increases computation time by 50%. These 10 feature subtypes were confirmed in subsequent analysis to correspond to different pathological entities. For example, Cluster-3 represents "hypoxic tumor nests with high CD8+ T cell infiltration," while Cluster-7 corresponds to "vascularized fibrotic stroma," and other patterns with clear biological interpretations.

[0081] The final classification task was performed by a specially trained ResNet-101 model. The model's training data came from 200 IHC image regions independently annotated by three senior pathologists and 1000 IHC images (including 400 tumor regions, 400 stromal regions, and 400 necrotic regions) annotated in a public database. The annotation consistency reached 0.82 (p<0.001) according to the Kappa test. In terms of model architecture design, a transfer learning strategy was adopted, initializing network parameters with ImageNet pre-trained weights and fine-tuning the network layers using a progressive unfreezing technique. It is particularly important to emphasize that the innovation of this embodiment lies in its three-stage processing flow of "feature extraction - subtype clustering - fine classification": compared to the baseline scheme of directly simplifying HVIT high-dimensional features and inputting them into ResNet for three-class classification (accuracy 86.3%), the cascaded method in this embodiment improves the classification accuracy to 95.5% while reducing inference time by 80%. This performance improvement mainly stems from the effective regularization of the feature space by the KNN clustering module, which allows the ResNet classifier to focus more on feature subspaces with clear pathological significance.

[0082] From a system-wide perspective, the HVIT model and the ResNet classifier form the core framework of the method, respectively undertaking the key functions of feature extraction and final classification. The introduction of the KNN clustering module reflects an important engineering optimization idea: by performing pre-clustering in the feature space, it not only preserves the subtle differences in tissue subtypes (such as tumor regions with different degrees of hypoxia), but also significantly reduces the computational complexity of subsequent classification.

[0083] Practical application testing showed that this scheme demonstrated stable performance in the pathological image analysis of lung cancer (from 99 cases at Beijing Chest Hospital), achieving F1-scores of 0.97, 0.95, and 0.98 for tumor regions, stromal regions, and necrotic regions, respectively. This stratified processing strategy not only aligns with the diagnostic logic of pathologists (identifying tissue subtypes first and then determining region categories), but more importantly, it provides interpretable intermediate results for clinicopathological analysis, contributing to a deeper understanding of the heterogeneous characteristics of the tumor microenvironment. EFS analysis was then performed on this data; these patients all had CD8... + T cell expression was obtained by analyzing the CD8 expression in each region. + After T cell density (the core of this model is to replace pathologists in manually counting and calculating CD8) + T cell density), CD8 in these patients + The density of T cells was statistically analyzed, and then a threshold of 0.41 cells / mm² was set. Cells above the threshold were considered CD8+. + Patients with high T cell expression, but below the threshold for CD8 expression... + Patients with low T cell expression. Results showed that CD8... + High expression of T cells (CD8) + Patients with high T-cell density have better prognoses, while those with high CD8 density have better prognoses. + Low expression of T cells (CD8) + Patients with low T-cell density have a poorer prognosis, consistent with hospital prognostic data. Following this, pathologists performed random sampling tests on the patients, selecting five fields of view from each region of each pathological image and using the mean to represent the CD8 count for that region. + The number of T cells is then divided by the area of ​​the region to obtain the corresponding CD8. + The density of T cells, after testing, showed that the results of the model in this embodiment were consistent with the sampling results of the pathologist. The pathologist then randomly selected 20 IHC cells for counting and statistical analysis, and found that CD8... +The number of T cells corresponds to the number identified by the model. The model accuracy in this embodiment is over 95.5%. During this process, the model in this embodiment takes 60 minutes to process each IHC image (using standard hospital computer equipment), while pathologists require at least 3 hours to count and statistically analyze each IHC image. Because each part of the model in this embodiment is separable, the reading process can be switched from fully automated to semi-automated nTIL density reading. After pathologists manually delineate the region and then use the model for identification and statistics, the average processing time per image can be reduced to 10 minutes. This avoids errors in nTIL density readings by different pathologists and significantly saves pathologists' reading time. Table 1 below illustrates a method based on CD8+ in necrotic regions provided by this invention. + The cancer immunotherapy prognostic prediction method based on T cell density is superior to existing clinical protocols. Table 2 illustrates the advantages of this invention based on CD8+ in necrotic regions. + Advantages of T-cell density as a method for predicting prognosis in cancer immunotherapy.

[0084] Table 1

[0085]

[0086] Table 2

[0087] Indicator Traditional AI recognition method This method Improved effect Analysis time / case 2.5 hours 60±8 minutes Speed up 2.2 times Counting difference rate 12.7％ 3.8％ Accuracy improved 3.34 times Necrotic area recognition accuracy 84.2％ 95.5％ Increased by 11.3%

[0088] The following describes the CD8-based necrotic region based invention. + This paper describes a cancer immunotherapy prognostic prediction system based on T cell density, specifically CD8+ within necrotic regions. + The cancer immunotherapy prognostic prediction system based on T cell density, as described above, is similar to the CD8+ system within necrotic regions. + Methods for predicting the prognosis of cancer immunotherapy based on T-cell density can be cross-referenced.

[0089] See Figure 3 This invention provides a method based on CD8 in necrotic regions. + A cancer immunotherapy prognostic prediction system based on T-cell density may include:

[0090] The CD8 IHC slice image acquisition module is used to acquire CD8 IHC slice images of tumor tissue from a subject after neoadjuvant immunotherapy, wherein the subject is a cancer patient.

[0091] The pathological feature extraction module is used to extract pathological features from CD8 IHC slice images of tumor tissue after neoadjuvant immunotherapy in the subject, using a multi-scale feature extraction model.

[0092] The necrotic region acquisition module is used to: obtain the necrotic region in the CD8 IHC slice image based on the pathological features on the CD8 IHC slice image through a region classification model;

[0093] The cell recognition module is used to: utilize a target recognition model to perform CD8 identification on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0094] The EFS prediction module is used to: predict the CD8+ of necrotic regions in CD8 IHC slice images. + T cell density predicts the EFS of the subject.

[0095] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the following steps:

[0096] IHC (immunohistochemical staining) slice images of tumor tissue from subjects who have undergone neoadjuvant immunotherapy were obtained. The subjects were cancer patients.

[0097] Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using a multi-scale feature extraction model.

[0098] Based on the pathological features on CD8 IHC slice images, the necrotic areas in CD8 IHC slice images are obtained through a region classification model.

[0099] Using a target recognition model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0100] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density predicts the EFS of the subject.

[0101] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0102] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps:

[0103] IHC (immunohistochemical staining) slice images of tumor tissue from subjects who have undergone neoadjuvant immunotherapy were obtained. The subjects were cancer patients.

[0104] Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using a multi-scale feature extraction model.

[0105] Based on the pathological features on CD8 IHC slice images, the necrotic areas in CD8 IHC slice images are obtained through a region classification model.

[0106] Using a target recognition model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0107] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density predicts the EFS of the subject.

[0108] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0109] IHC (immunohistochemical staining) slice images of tumor tissue from subjects who have undergone neoadjuvant immunotherapy were obtained. The subjects were cancer patients.

[0110] Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using a multi-scale feature extraction model.

[0111] Based on the pathological features on CD8 IHC slice images, the necrotic areas in CD8 IHC slice images are obtained through a region classification model.

[0112] Using a target recognition model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density;

[0113] Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density predicts the EFS of the subject.

[0114] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units 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 achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0115] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0116] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method based on CD8+ in necrotic regions + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... include: Obtain CD8 IHC slice images of tumor tissue from subjects who have undergone neoadjuvant immunotherapy. The subjects are cancer patients. Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using a multi-scale feature extraction model. Based on the pathological features on CD8 IHC slice images, the necrotic areas in CD8 IHC slice images are obtained through a region classification model. Using a target recognition model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density; Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + T cell density predicts the EFS of the subject.

2. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The method involves extracting pathological features from CD8 IHC slice images of tumor tissue from the subject after neoadjuvant immunotherapy using a multi-scale feature extraction model, including: Based on the CD8 IHC slice images of tumor tissue from the subjects after neoadjuvant immunotherapy, pathological features on the CD8 IHC slice images were extracted using the HVIT model. The HVIT model is a pre-trained model. During the training process, the HVIT model first learns features from IHC slice images of tumor tissues of cancer patients who have undergone neoadjuvant immunotherapy. Then, low-rank adaptive tuning is used to fine-tune the HVIT model to obtain the trained HVIT model.

3. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The process of identifying necrotic regions in CD8 IHC slice images based on pathological features using a region classification model includes: Unsupervised clustering was performed on the pathological features on CD8 IHC slice images to divide the CD8 IHC slice images into multiple sub-regions; Using a region classification model, multiple sub-regions on CD8 IHC slice images were divided into tumor regions, necrotic regions, and stromal regions.

4. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The unsupervised clustering of pathological features on CD8 IHC slide images divides the CD8 IHC slide images into multiple subdivided regions, including: The KNN clustering method was used to perform unsupervised clustering on the pathological features of CD8 IHC slice images, dividing the CD8 IHC slice images into 10 sub-regions.

5. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The region classification model is used to divide multiple sub-regions on the CD8 IHC slice image into tumor regions, necrotic regions, and stroma regions, including: Using the ResNet101 model, the 10 subdivision regions on the CD8 IHC slice image are divided into tumor region, necrotic region and stroma region, and the mask of the necrotic region on the CD8 IHC slice image is obtained. The area of ​​the necrotic region is obtained by using the mask of the necrotic region on the CD8 IHC slice image; Among them, the ResNet101 model is a pre-trained model. During the training process, the ResNet101 model first learns the features of IHC slice sample images labeled with tumor areas, necrotic areas, and stromal areas. In terms of model architecture, a transfer learning strategy is adopted, the network parameters are initialized with ImageNet pre-trained weights, and the network layers are fine-tuned through a progressive unfreezing method.

6. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The target recognition model is used to perform CD8 analysis on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density includes: Using the YOLOv11 model, CD8 spectroscopy was performed on necrotic regions in CD8 IHC slice images. + T cells were used to identify the necrotic CD8+ region in the CD8+ IHC slice image. + The number of T cells; Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + The number of T cells and the area of ​​necrosis were used to obtain the CD8+ in the necrotic area of ​​the CD8+ IHC slice image. + T cell density.

7. The method based on CD8+ within the necrotic region as described in claim 1 + A method for predicting the prognosis of cancer immunotherapy based on T cell density, characterized in that... The CD8 based on the necrotic region in the CD8 IHC slice image + T cell density, used to predict the EFS of a subject, includes: Based on the CD8 IHC slice image, the necrotic area of ​​CD8 + Based on T cell density and a preset threshold, the subject was identified as a CD8+ individual. + T cell hyperexpression or CD8+ + Individuals with low T-cell expression; When the test subject is identified as CD8 + When individuals with high T-cell expression are identified, the EFS prediction result of the test subject is determined to be low risk of relapse. When the test subject is identified as CD8 + When T cell counts are high or low, the EFS prediction result of the test subject is determined to be of high risk of relapse.

8. A method based on CD8+ in necrotic regions + A cancer immunotherapy prognostic prediction system based on T cell density, characterized in that... include: The CD8 IHC slice image acquisition module is used to acquire CD8 IHC slice images of tumor tissue from a subject after neoadjuvant immunotherapy, wherein the subject is a cancer patient. The pathological feature extraction module is used to extract pathological features from CD8 IHC slice images of tumor tissue after neoadjuvant immunotherapy in the subject, using a multi-scale feature extraction model. The necrotic region acquisition module is used to: obtain the necrotic region in the CD8 IHC slice image based on the pathological features on the CD8 IHC slice image through a region classification model; The cell recognition module is used to: utilize a target recognition model to perform CD8 identification on necrotic regions in CD8 IHC slice images. + T cells were used for identification, and CD8+ was obtained from the necrotic area in the CD8+ IHC slice image. + T cell density; The EFS prediction module is used to: predict the CD8+ of necrotic regions in CD8 IHC slice images. + T cell density predicts the EFS of the subject.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the CD8-based necrotic region as described in any one of claims 1 to 7. + A method for predicting the prognosis of cancer immunotherapy based on T cell density.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the CD8-based method within the necrotic region as described in any one of claims 1 to 7. + A method for predicting the prognosis of cancer immunotherapy based on T cell density.

Images