A non-invasive prediction method for pd-1 and pd-l1 expression in liver cancer patients

Abstract

This invention relates to a non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients, comprising: utilizing self-supervised contrastive learning to help the model extract deep representations of PD-1 and PD-L1 expression from CT images of liver cancer patients; using patch shuffling to enhance the model's ability to learn local features of PD-1 and PD-L1 expression in CT images; and introducing unlabeled training images for self-supervised training. The CT images of liver cancer patients are input into the prediction model, and the probabilities of PD-1 and PD-L1 expression are obtained simultaneously through a transfer learning strategy. The top k images with the highest probabilities are selected, and their average value is calculated as the final probability. This invention improves the ability of deep learning models to extract deep representations of CT images through a self-supervised contrastive learning strategy and introduces a patch shuffling data augmentation method to enrich the feature patterns of training CT images and enhance the model's ability to extract local feature representations.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to a non-invasive method for predicting PD-1 and PD-L1 expression in liver cancer patients. Background Technology

[0002] Hepatocellular carcinoma (HCC) is the third leading cause of cancer death worldwide, and its incidence is steadily increasing. Surgery and percutaneous treatments, such as liver transplantation and radiofrequency ablation (RFA), can be used for early-stage HCC, but recurrence rates are high. However, most patients are diagnosed at an advanced stage, resulting in a poor prognosis. Immune checkpoint inhibitors (ICIs) are currently a potentially effective treatment for patients with advanced HCC and have yielded encouraging clinical results. Programmed death receptor-1 (PD-1) and programmed death ligand-1 (PD-L1) are members of the extensively studied immune checkpoint pathway. PD-1 is primarily expressed on the T cell membrane and is a negative regulator of antigen responses. PD-L1 inhibits the cytotoxic activity of T cells by binding to PD-1 expressed on the T cell surface and allows tumor cells to escape immune responses. ICIs (immune checkpoint inhibitors) work by blocking PD-1 and PD-L1 checkpoints, showing potential to improve the treatment of HCC patients. However, only a small percentage of HCC patients benefit from this immunotherapy, with a durable response rate of only 15–20% for anti-PD-1 therapy.

[0003] Current research indicates that the expression status of PD-1 and PD-L1 is associated with clinical outcomes and treatment response to PD-1 / PD-L1 pathway inhibition. Therefore, assessing PD-1 and PD-L1 expression is crucial for identifying individuals who respond to checkpoint blockade, precise treatment decisions, and improved patient outcomes. Histopathological examination is the gold standard for assessing PD-1 and PD-L1 expression; however, histopathological biopsy is an invasive procedure and is associated with sampling errors and increased risk of disease progression.

[0004] The rapid development of deep learning (DL) technology has yielded significant achievements in many practical applications, such as image recognition and classification, and natural language processing. Based on sufficient training data, deep learning can effectively manage large amounts of high-dimensional and noisy data by capturing typical complex features and nonlinear relationships. Therefore, DL technology is increasingly being applied to medical research. For example, deep learning networks are used to screen for common, treatable blinding retinal diseases; convolutional neural networks (CNNs) are used to predict Covid-19 using X-rays; and lymph node and tumor regions are identified based on lymph node images. Therefore, how to apply deep learning technology to patient CT images to predict PD-1 / PD-L1 expression to achieve effective, non-invasive prediction and promote personalized and precise treatment decisions for liver cancer patients is a problem that needs to be solved at this stage.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients. This method applies deep learning technology to patient CT images to predict PD-1 / PD-L1 expression, thereby achieving effective and non-invasive prediction and promoting personalized and precise treatment decisions for liver cancer patients.

[0007] The objective of this invention is achieved through the following technical solution: a non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients, the non-invasive prediction method comprising:

[0008] Self-supervised contrastive learning training steps: Self-supervised contrastive learning helps the model extract deep representations of PD-1 and PD-L1 expression in CT images of liver cancer patients. Patch shuffling data augmentation method is used to enhance the model's ability to learn local features of PD-1 and PD-L1 expression in CT images. At the same time, unlabeled training images are introduced for self-supervised training to improve data utilization efficiency and the pattern of training images.

[0009] Multi-task prediction steps: After obtaining the trained prediction model, the CT images of liver cancer patients are input into the prediction model. The probability of PD-1 and PD-L1 expression is obtained simultaneously through the transfer learning strategy. The top k images with high probabilities are selected from all images, and their average value is calculated as the final probability of PD-1 and PD-L1 expression in the patient.

[0010] The non-invasive prediction method further includes an image preprocessing step performed before the self-supervised contrastive learning training step, the image preprocessing step including: acquiring CT images and obtaining immunohistochemistry of PD-1 and PD-L1 expression.

[0011] The acquisition of CT images includes: acquiring three-stage CT images of liver cancer patients through a detector CT; obtaining arterial phase images and portal venous phase images 25 seconds and 60 seconds after contrast agent injection, respectively; selecting the tumor region as the input of the network for prediction; performing tumor segmentation on the initial CT image through SEVB-Net; centering the center point of the tumor region; and using an N×N region as the final input.

[0012] The immunohistochemistry for obtaining PD-1 and PD-L1 expression includes:

[0013] The parafibrous tissue of the surgically removed tumor specimen was cut into 4μm thick sections, dewaxed, hydrated, and then the antigen was extracted.

[0014] Tissue sections and primary antibodies were cultured overnight at 4°C with rabbit anti-human PD-1 polyclonal antibody and anti-human PD-L1 monoclonal antibody, and then cultured with secondary antibodies.

[0015] PD-1 was stained with 3,3'-diaminobenzidine and counterstained with hematoxylin. PD-1 expression results were expressed as the proportion of PD-1+ tumor-infiltrating immune cells. The threshold for PD-1 overexpression was determined by x-tile software. Cases with expression greater than 5% were considered PD-1 positive. The number of PD-L1 cells was quantified as ×400. PD-L1 expression results were expressed as the proportion of PD-L1+ tumor cells. The threshold for PD-L1 overexpression was determined to be 3%.

[0016] The self-supervised contrastive learning training steps specifically include the following:

[0017] For a given CT input image x, view x1 is obtained by randomly resizing the input image x, cropping it, randomly flipping it horizontally, and applying Gaussian blur.

[0018] The input image x is divided into four blocks by patching and shuffling, and the four blocks are randomly pieced together into a complete image with probability P to obtain view x2.

[0019] Different views x1 and x2 are input into an encoder f(x) consisting of a ResNet-50 model and a projective multilayer perceptron. View x1 first outputs c through encoder f(x), and then enters the predictive multilayer perceptron. The predictive multilayer perceptron adds a header h(x) operation to the input to obtain the output vector. View x2 enters another branch, obtaining the output vector only through encoder f(x).

[0020] After obtaining the two output vectors, the similarity between views x1 and x2 is maximized using negative cosine similarity. Where ||·||2 represents L2 norm normalization. To prevent model collapse, gradient backpropagation needs to be stopped for z2. The similarity is represented by D.

[0021] Finally, the loss function for model training is obtained as follows: Where D(p2, stopgrad(z1)) represents the similarity of the output vectors obtained after swapping views x1 and x2, and stopgrad() indicates that the gradient backpropagation operation is stopped.

[0022] The multi-task prediction step specifically includes:

[0023] The ResNet50 model trained by self-supervised contrastive learning is used as the backbone for multi-task prediction of PD-1 and PD-L1 expressions using transfer learning. The output is then passed through two prediction multilayer perceptrons, and two new fully connected layers are used to predict PD-1 and PD-L1 expressions separately and simultaneously.

[0024] For the labeled dataset DS = {x i y i 1 y i 2}, x i For training images, y i 1 For PD-1 expression labels, y i 2 The label is the PD-L1 expression, and the output a of the nth fully connected network for the i-th class is calculated using Softmax. n Predicted probability Where C represents the number of truth labels, and the probability p 1 and p 2 The PD-1 and PD-L1 expressions are calculated using the actual labels on CT images;

[0025] The loss function of the prediction model is derived from the cross-entropy of the two classifiers, PD-1 and PD-L1. The training efficiency of the network model is improved by updating the network model weights based on the similarity between training images and PD-1 and PD-L1 labels.

[0026] After obtaining the trained prediction model, the CT images of liver cancer patients are input into the prediction model to obtain the probability of PD-1 and PD-L1 expression. The top k images with the highest probability are selected from all images, and their average value is used as the final probability of PD-1 or PD-L1 expression in the patient.

[0027] The present invention has the following advantages:

[0028] 1. By employing a self-supervised contrastive learning strategy, the ability of deep learning models to extract deep representations of CT images is improved. A patch shuffling data augmentation method is introduced to enrich the feature patterns of training CT images, increasing the difficulty of model training and thus enhancing the model's ability to extract local feature representations.

[0029] 2. The ability to simultaneously predict PD-1 and PD-L1 expression is crucial for personalized medicine and immunosuppressant selection in liver cancer patients. Predicting PD-1 and PD-L1 expression simultaneously from a patient's CT images is a meaningful but challenging task. Since the training phase of contrastive learning does not require label information, the deep expression extracted from convolutional layers can be used to predict both PD-1 and PD-L1 expression. Compared to previous work predicting the expression of a single protein (PD-1 or PD-L1), the method proposed in this invention can obtain more comprehensive expression data from liver cancer patient CT images, which is more conducive to developing personalized treatment plans.

[0030] 3. Additional unlabeled CT images are introduced during the self-supervised contrastive learning training phase. Generally, although unlabeled medical images meet the standards for network training, they are often discarded due to the lack of label information, resulting in a waste of data resources. The contrastive learning used in this invention is a self-supervised training method that can train the model without label information; therefore, this invention uses these unlabeled images that would otherwise be discarded to increase the number of training images during the self-supervised training phase, thereby enriching the data patterns and effectively improving data utilization efficiency. Attached Figure Description

[0031] Figure 1 This is a schematic flowchart of the method of the present invention;

[0032] Figure 2 A diagram illustrating different data augmentation methods;

[0033] Figure 3 A heatmap showing the t-test results for different deep learning models. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application provided below with reference to the accompanying drawings is not intended to limit the scope of protection of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The present invention will be further described below with reference to the accompanying drawings.

[0035] This invention relates to a non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients. This method is based on the contrastive learning network CLNet, utilizing self-supervised contrastive learning to help the model better extract deep representations of PD-1 / PD-L1 expression from CT images of liver cancer patients. A patch-shuffle data augmentation method is used to enhance the model's ability to learn local features of PD-1 / PD-L1 expression in CT images. Simultaneously, based on the self-supervised training characteristics of label-free image models, additional unlabeled training images that would otherwise be discarded are introduced to improve data utilization efficiency and the patterns of training images. Therefore, the non-invasive prediction method for PD-1 / PD-L1 expression in liver cancer patients proposed in this invention can effectively extract deep representations of PD-1 / PD-L1 expression from CT images of liver cancer patients, improve data utilization efficiency, and simultaneously predict PD-1 and PD-L1 expression, which is crucial for guiding personalized treatment of liver cancer patients.

[0036] like Figure 1 As shown, it includes three stages: image preprocessing, self-supervised contrastive learning training, and multi-task prediction.

[0037] 1. Image preprocessing;

[0038] (1) CT Image Acquisition: CT images of liver cancer patients used were acquired from multi-detector CT. All patients underwent three-phase CT (non-contrast phase, arterial phase, and portal venous phase), with the arterial and portal venous phases acquired at 25 and 60 seconds after contrast agent injection, respectively. After acquiring the CT images, the tumor region was selected as the input to the network for prediction. To obtain accurate tumor regions, the initial CT images were segmented using SEVB-Net, an improved V-network developed by United Imaging Intelligence. The center point of the tumor region was centered, and a 128×128 region was used as the final input.

[0039] (2) Immunohistochemistry of PD-1 and PD-L expression:

[0040] Parafibrous tissue from surgically removed tumor specimens was cut into 4 μm thick sections, dewaxed, hydrated, and then antigens were extracted. The tissue sections and primary antibodies were then cultured overnight at 4°C with rabbit anti-human PD-1 polyclonal antibody and anti-human PD-L1 monoclonal antibody, followed by culture with secondary antibodies. PD-1 staining was performed with 3,3'-diaminobenzidine and counterstained with hematoxylin. PD-1 expression results were expressed as the proportion of PD-1+ tumor-infiltrating immune cells (PD-1+ immune cells / total immune cells). The cutoff value for PD-1 overexpression was determined using x-tile software; cases with expression greater than 5% were considered PD-1 positive. The number of PD-L1 cells was quantified as ×400 (0.0484 mm^2), and PD-L1 expression results were expressed as the proportion of PD-L1+ tumor cells (PD-L1+ tumor cells / total tumor cells). The cutoff value for PD-L1 overexpression was determined to be 3%.

[0041] 2. Self-supervised comparative learning training;

[0042] This stage uses self-supervised contrastive learning to train the deep neural network. The ability to extract deep representations from CT images is one of the most important functions of a deep learning network model, and the quality of the deep representations generated by the model directly affects the final prediction performance. Contrastive learning is a novel self-supervised training strategy that can effectively mine deep image representations. It collects two viewpoints of the same image (positive pairs) and rejects different images (negative pairs). This invention employs an advanced contrastive learning scheme, SimSiam, which uses only positive pairs.

[0043] like Figure 1 As shown, for a given CT input image x, two different views x1 and x2 are generated using different data augmentation methods. Constructing two different views of the same image is an important part of extracting the deep representation of CT images. View x1 is obtained using common data augmentation methods of contrastive learning, such as random resizing cropping, random horizontal flipping, and Gaussian blur. Based on this, this invention adds a novel data augmentation method—patch shuffle—to obtain view x2. Different data augmentation methods are as follows... Figure 2 As shown. The random resizing cropping method crops any part of the image and then resizes it to a uniform size. The random horizontal flipping method randomly flips the image horizontally. The Gaussian blur method uses Gaussian noise to randomly blur the image. The patch shuffle data augmentation method divides the image into four blocks and randomly stitches them together into a complete image with probability P. Contrastive learning training improves the similarity between the deep representation of the patch shuffled image and the deep representation of the original image. This novel data augmentation method aims to deepen the model's understanding of the local representation of the training image.

[0044] After obtaining different views x1 and x2 of the input image x, the image is fed into the encoder f(x). Figure 1 As shown, the encoder f(x) consists of a ResNet-50 network and a projective multilayer perceptron (MLP). x1 is output c by the encoder and then enters the predictive MLP. The predictive MLP adds a header h(x) operation to the input, resulting in the output vector. View x2 enters another branch, obtaining the output vector only through encoder f(x).

[0045] After obtaining the two output vectors, the similarity between views x1 and x2 is maximized using negative cosine similarity. Where ||·||2 represents L2 norm normalization. To prevent model collapse, gradient backpropagation needs to be stopped for z2. The similarity is represented by D.

[0046] Finally, the loss function for model training is obtained as follows: Where D(p2, stopgrad(z1)) represents the similarity of the output vectors obtained after swapping views x1 and x2, and stopgrad() indicates that the gradient backpropagation operation is stopped.

[0047] 3. Multi-task prediction;

[0048] In the multi-task prediction stage, a transfer learning strategy was used to train the prediction model. Multi-task prediction refers to simultaneously predicting two tasks: PD-1 and PD-L1 expression. In multi-task prediction, the self-supervised contrastive learning-trained model is well-suited for simultaneously predicting PD-1 and PD-L1 because it doesn't use PD-1 or PD-L1 labels but extracts tumor-related features from CT images, which are applicable to predicting PD-1 and PD-L1 expression. Therefore, using transfer learning, the self-supervised contrastive learning-trained ResNet50 model was used as the backbone for multi-task prediction of PD-1 and PD-L1 expression. The output was then passed through two predictive multilayer perceptrons, utilizing two new fully connected layers to predict PD-1 and PD-L1 expression separately and simultaneously, as shown below. Figure 1 As shown, in Figure 1 In this network, the two fully connected layers have the same network structure, consisting of three linear layers, including a batch normalization (BN) layer and a ReLU activation function. For a labeled dataset, DS = {x} i ,y i 1 ,y i 2}, x i For training images, yi 1 For PD-1 expression labels, y i 2 Labels for PD-L1 expressions. Softmax is used to compute the a generated by the nth fully connected layer for class i. n The predicted probability is:

[0049]

[0050] Where C represents the number of truth labels, and the probability p 1 and p 2 The PD-1 and PD-L1 expressions are calculated using the actual labels from CT images; the loss function of the prediction model is composed of the cross-entropy of the two classifiers for the PD-1 and PD-L1 expressions:

[0051]

[0052] Here, CrossEntropy represents cross-entropy, which updates the network model weights by measuring the similarity between training images and PD-1 and PD-L1 labels, thus achieving multi-task prediction training. This not only improves the training efficiency of the network but also helps the model explore the relationship between PD-1 and PD-L1 expressions.

[0053] As shown above, the first stage of network model training is self-supervised training, using contrastive learning to better extract deep representations of CT images. The second stage, multi-task prediction training, is supervised training, using transfer learning to simultaneously predict PD-1 and PD-L1 expressions. In self-supervised contrastive learning training, this invention uses unlabeled images to pre-train the model. Many unlabeled medical images are added in actual image processing. Generally, although these images meet the criteria for network training, they are discarded due to the lack of label information, leading to a waste of data resources. However, in the self-supervised training stage, this invention still uses these unlabeled images, adding these unlabeled CT images to the training set for the self-supervised training stage. This not only improves data utilization efficiency but also increases data patterns and improves the model's generalization ability.

[0054] After obtaining the trained prediction model, CT images of liver cancer patients are input into the prediction model to obtain the probability of PD-1 and PD-L1 expression. For each patient, only the top k images with the highest probability are selected from all images, and their average value is obtained as the final probability of PD-1 or PD-L1 expression in that patient.

[0055] The data used in this invention comprised a cohort of 121 histologically confirmed hepatocellular carcinoma patients who received systemic sorafenib treatment between July 2012 and April 2017. Of the 121 patients, 87 met the inclusion criteria, and 34 were excluded. Patients who underwent surgery between July 2012 and March 2016 were assigned to the training set (number: 63; mean age: 53.13 ± 11.79 years; 57 males and 6 females), and patients who underwent surgery between April 2016 and April 2017 were assigned to the validation set (number: 24; mean age: 47.96 ± 13.76 years; 22 males and 2 females). The validation set included younger patients compared to the training set. There were no gender differences between the training and validation sets. The PD-1 positivity rates in the training and validation sets were 41.27% and 41.67%, respectively, and the PD-L1 positivity rates were 22.22% and 29.17%, respectively. Clinical characteristics of hepatocellular carcinoma patients in the training and validation cohorts are shown in Table 1.

[0056]

[0057] Experiment 1: Performance Comparison with Other Deep Learning Models

[0058] First, the CLNet model used in the proposed method of this invention was compared with other deep learning (DL) models to predict PD-1 / PD-L1 expression in the validation set. Other DL models included ResNet-50, VGGNet-19, DenseNet-100, and PyramidNet-101. Experimental results are shown in Table 2. The CLNet model proposed in this invention achieved AUCs of 86.56% and 83.93% for predicting PD-1 and PD-L1 expression, respectively; ACCs of 84.38% and 83.33%; sensitivity (Sen) of 92.86% and 85.00%; specificity (Spec) of 80.88% and 82.14%; and MCCs of 0.688 and 0.671, respectively. Compared with ResNet-50, which has the same convolutional structure as the proposed method, the AUCs for PD-1 and PD-L1 expression increased by 8.41% and 8.57%, respectively, and the ACCs increased by 4.17% and 6.25%, respectively. This demonstrates that contrastive learning is helpful in extracting features associated with PD-1 and PD-L1 expression. Furthermore, the proposed method outperforms other models such as DensNet-100 and PrymidNet-100, illustrating the effectiveness of our model.

[0059] Table 2. Comparison of prediction performance of different models for PD-1 and PD-L1 expression

[0060]

[0061] AUC=Area under the receiver operating characteristic curve; Acc=Accuracy; Sen=Sensitivity; Spec=Specificity; Mcc=Matthews correlationcoefficient.

[0062] Experiment 2: Performance Comparison with Machine Learning Methods

[0063] The proposed method was also compared with other machine learning (ML) methods, including K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Decision Trees. The experimental results are shown in Table 3. As can be seen from Table 3, the prediction performance of the proposed method is significantly better than that of traditional machine learning methods. When predicting PD-1 and PD-L1, the AUC of the proposed method is 18.8% and 21.07% higher than that of the decision tree, respectively. The results indicate that, compared with traditional machine learning methods, the proposed method has a stronger ability to predict PD-1 and PD-L1 expressions.

[0064] Table 3. Comparison of predictive performance with machine learning methods

[0065]

[0066] AUC=Area under the receiver operating characteristic curve; Acc=Accuracy; Sen=Sensitivity; Spec=Specificity; Mcc=Matthews correlationcoefficient.

[0067] Experiment 3: Ablation Experiment

[0068] This experiment describes several ablation experiments conducted to analyze the effectiveness of the training strategy. First, the performance of a model trained using only general images during the self-supervised contrastive learning training phase was compared with that trained using additional unlabeled but high-quality CT images. The results are shown in Table 4. Table 4 shows that including more images during self-supervised contrastive learning pre-training significantly improves the network's performance in PD-1 and PD-L1 expression prediction (PD-1: p-value = 0.015; PD-L1: p-value = 0.0113). Although these unlabeled images are usually discarded in traditional data processing, they are used for self-supervised pre-training because contrastive learning does not require label information. Experimental results demonstrate that additional image information can expand the image patterns learned by the model, thereby improving the model's ability to extract image features.

[0069] Table 4. Comparison of prediction performance with and without training using unlabeled images

[0070]

[0071] To verify the effectiveness of the patch shuffle data augmentation method used in contrastive learning, a series of comparative experiments were conducted. In these experiments, networks were trained with and without the patch shuffle strategy, and the results are shown in Table 5. The results show that, compared with conventional contrastive learning, adding the patch shuffle strategy improves the AUC and ACC of PD-1 expression by 6.50% and 8.34%, respectively, and the AUC and ACC of PD-L1 expression by 5.36% and 6.25%, respectively. However, the training effect using only the patch shuffle strategy is not ideal, with an AUC of only 72.06% and ACC of only 73.96% for PD-1 expression, and an AUC of 71.25% and ACC of 72.92% for PD-L1 expression. This indicates that patch shuffle is a suitable method for adding more image patterns in contrastive learning, but traditional data augmentation methods, such as random size cropping and random horizontal flipping, are still necessary.

[0072] Table 5. Comparison of prediction performance using different data augmentation methods

[0073]

[0074] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A non-invasive method for predicting PD-1 and PD-L1 expression in liver cancer patients, characterized in that: The non-invasive prediction method includes: Self-supervised contrastive learning training steps: Self-supervised contrastive learning helps the model extract deep representations of PD-1 and PD-L1 expression in CT images of liver cancer patients. Patch shuffling data augmentation method is used to enhance the model's ability to learn local features of PD-1 and PD-L1 expression in CT images. At the same time, unlabeled training images are introduced for self-supervised training to improve data utilization efficiency and training image patterns. Multi-task prediction steps: After obtaining the trained prediction model, the CT images of liver cancer patients are input into the prediction model. The probability of PD-1 and PD-L1 expression is obtained simultaneously through the transfer learning strategy. The top k images with high probabilities are selected from all images, and their average value is calculated as the final probability of PD-1 and PD-L1 expression in the patient. The multi-task prediction step specifically includes: The ResNet50 model trained by self-supervised contrastive learning is used as the backbone for multi-task prediction of PD-1 and PD-L1 expressions using transfer learning. The output is then passed through two prediction multilayer perceptrons, and two new fully connected layers are used to predict PD-1 and PD-L1 expressions separately and simultaneously. For the labeled dataset DS = {x i y i 1 y i 2 }, x i For training images, y i 1 For PD-1 expression labels, y i 2 The label is the PD-L1 expression, and the output a of the nth fully connected network for the i-th class is calculated using Softmax. n Predicted probability Where C represents the number of truth labels and the probability p 1 and p 2 The PD-1 and PD-L1 expressions are calculated using the actual labels on CT images; The loss function of the prediction model is derived from the cross-entropy of the two classifiers, PD-1 and PD-L1. By updating the network model weights based on the similarity between training images and PD-1 and PD-L1 labels, the training efficiency of the network model can be improved. After obtaining the trained prediction model, the CT images of liver cancer patients are input into the prediction model to obtain the probability of PD-1 and PD-L1 expression. The top k images with the highest probability are selected from all images, and their average value is used as the final probability of PD-1 or PD-L1 expression in the patient.

2. The non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients according to claim 1, characterized in that: The non-invasive prediction method further includes an image preprocessing step performed before the self-supervised contrastive learning training step, the image preprocessing step including: acquiring CT images and obtaining immunohistochemistry of PD-1 and PD-L1 expression.

3. The non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients according to claim 2, characterized in that: The acquisition of CT images includes: acquiring three-stage CT images of liver cancer patients through a detector CT; obtaining arterial phase images and portal venous phase images 25 seconds and 60 seconds after contrast agent injection, respectively; selecting the tumor region as the input of the network for prediction; performing tumor segmentation on the initial CT image through SEVB-Net; centering the center point of the tumor region; and using an N×N region as the final input.

4. The non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients according to claim 2, characterized in that: The immunohistochemistry for obtaining PD-1 and PD-L1 expression includes: The parafibrous tissue of the surgically removed tumor specimen was cut into 4μm thick sections, dewaxed, hydrated, and then the antigen was extracted. Tissue sections and primary antibodies were cultured overnight at 4°C with rabbit anti-human PD-1 polyclonal antibody and anti-human PD-L1 monoclonal antibody, and then cultured with secondary antibodies. PD-1 was stained with 3,3'-diaminobenzidine and counterstained with hematoxylin. PD-1 expression results were expressed as the proportion of PD-1+ tumor-infiltrating immune cells. The threshold for PD-1 overexpression was determined by x-tile software. Cases with expression greater than 5% were considered PD-1 positive. The number of PD-L1 cells was quantified as ×400. PD-L1 expression results were expressed as the proportion of PD-L1+ tumor cells. The threshold for PD-L1 overexpression was determined to be 3%.

5. The non-invasive prediction method for PD-1 and PD-L1 expression in liver cancer patients according to claim 1, characterized in that: The self-supervised contrastive learning training steps specifically include the following: For a given CT input image x, view x1 is obtained by randomly resizing the input image x, cropping it, randomly flipping it horizontally, and applying Gaussian blur. The input image x is divided into four blocks by patching and shuffling, and the four blocks are randomly pieced together into a complete image with probability P to obtain view x2. Different views x1 and x2 are input into the encoder f(x), which is composed of a ResNet-50 model and a projection multilayer perceptron. View x1 first outputs c through the encoder f(x) and then enters the prediction multilayer perceptron. The prediction multilayer perceptron adds a head h(x) operation to the input and obtains the output vector p1≜ h(f(x1)). View x2 enters another branch and only passes through the encoder f(x) to obtain the output vector z2≜ f(x2). After obtaining the two output vectors, the similarity between views x1 and x2 is maximized using negative cosine similarity. ,in, Let z2 be the L2 norm normalization. To prevent the model from collapsing, gradient backpropagation needs to be stopped for z2. Then the similarity is represented by D. Finally, the loss function for model training is obtained as follows: ,in, This indicates the similarity of the output vectors obtained after swapping views x1 and x2. stopgrad() indicates that the gradient backpropagation operation is stopped.

Images