Pancreatic cancer diagnosis method, system, medium, electronic device

Abstract

The application provides a pancreatic cancer diagnosis method, system, medium and electronic equipment, and the method comprises the following steps: acquiring an abdominal plain CT image; dividing a pancreatic region in the abdominal plain CT image into image blocks, and extracting image block features and prediction probability values of the image blocks; respectively constructing corresponding case-level graphs based on image block features and prediction probability values of different case regions, wherein the different case regions comprise a non-pancreatic cancer case pancreatic region, a pancreatic cancer case tumor region and a pancreatic cancer case non-tumor region; training a graph neural network for pancreatic cancer prediction based on the graphs, and acquiring a pancreatic cancer prediction result based on the trained graph neural network. The pancreatic cancer diagnosis method, system, medium and electronic equipment of the application can realize intelligent early diagnosis of pancreatic cancer based on only a plain CT image, have high accuracy and strong applicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a pancreatic cancer diagnosis method, system, medium and electronic device. BACKGROUND

[0002] Pancreatic cancer is a highly malignant digestive system tumor with a very high mortality rate. This is because pancreatic cancer has no obvious specific symptoms in the early stage, and patients are usually diagnosed in the advanced stage. At present, surgical resection is the only effective method to cure pancreatic cancer, and traditional surgery, chemotherapy and radiotherapy can alleviate the symptoms of patients and prolong the survival rate of early patients. However, for patients in the advanced stage, there is still no effective cure. There is evidence that patients who are found in the early stage have a longer median survival time than those in the advanced stage. Therefore, early diagnosis of pancreatic cancer is of great significance and necessity. At present, there is no effective early diagnosis method for pancreatic cancer in clinical practice, and the early diagnosis rate is less than 5%. The only clinically approved blood biomarker CA19-9 for pancreatic cancer has low sensitivity and specificity for early patients, and is commonly used to assess the progression of pancreatic cancer, and cannot be used for general screening. The non-invasive imaging method for evaluating pancreatic cancer, enhanced CT and enhanced MRI, requires contrast agents, which limits their large-scale use and cannot provide a solution for early diagnosis and screening of pancreatic cancer. Therefore, an accurate and easily accessible early diagnosis method for pancreatic cancer is needed.

[0003] Non-contrast CT is a widely used imaging method that does not require contrast agents and is widely used in physical examinations and clinical routine examinations. This readily available medical detection method provides a cheap, simple and safe method for early diagnosis of pancreatic cancer and can be used as a potential tool for pancreatic cancer screening. However, non-contrast CT has low contrast for soft tissue, and it is difficult to identify tumors by visual observation, even for radiologists. In fact, many features of medical images are hidden from human observers. Deep learning technology can automatically analyze medical images and extract features such as local texture and shape changes that are difficult for the human eye to observe, making it possible to use non-contrast CT for automatic diagnosis of pancreatic cancer. Therefore, an accurate and reliable intelligent early diagnosis method for pancreatic cancer based on non-contrast CT images has become an urgent need in clinical practice.

[0004] The imaging methods commonly used in clinical practice for the evaluation of pancreatic cancer include endoscopic techniques (e.g., endoscopic ultrasonography), contrast-enhanced magnetic resonance imaging, and contrast-enhanced CT. Many studies have designed deep networks based on these images to develop intelligent diagnostic techniques to assist clinical diagnosis. For example, Tonokuza et al. developed a computer-aided diagnosis system using deep learning analysis of endoscopic ultrasonography images and verified the effectiveness of the system in detecting pancreatic cancer. Xie et al. developed a radiomics model based on multi-parametric MRI to evaluate the pathological outcome of pancreatic cancer. Chen et al. developed an intelligent pancreatic cancer detection tool based on contrast-enhanced CT images, which integrated five segmented convolutional neural networks (CNNs) to achieve sensitivity and specificity comparable to radiologists. Liu et al. used CNNs to extract patch features from contrast-enhanced CT images to achieve pancreatic tumor recognition, and the model was proven to be acceptable in generalization for patients of different ethnicities. These studies are good examples of intelligent pancreatic cancer diagnosis, but due to the invasiveness of the above-mentioned imaging modalities or the need for contrast agents, these methods are not suitable for large-scale use and cannot provide a solution for early diagnosis and screening of pancreatic cancer.

[0005] Currently, there are two studies exploring the feasibility of non-contrast CT, a widely used medical imaging modality, as a clinical tool for pancreatic cancer diagnosis. Qiu et al. proposed a multi-resolution statistical analysis framework that constructs a mathematical description of local subtle gray level changes in images, indicating that texture features related to pancreatic cancer histopathology can be extracted from non-contrast CT images. However, this work had a limited number of subjects, and the model constructed by manually extracting features had low accuracy. Xia et al. proposed a new Transformer method that integrates anatomical information, improving classification accuracy compared to the mainstream paradigm, but the binary classification task implemented in this work is the classification of abnormal pancreas (including pancreatic cancer and other pancreatic abnormalities) and normal pancreas, not the classification of pancreatic cancer and non-pancreatic cancer. These two works show good performance, but the data collected in both works is from a single center, and actual pancreatic cancer screening and early diagnosis methods need to be generalizable across multiple centers, so their true clinical value remains to be explored.

[0006] Multiple instance learning (MIL) is a supervised learning framework widely used in object detection, segmentation, classification, and other tasks, demonstrating strong representation capabilities. In this framework, data is organized into bags, each containing a set of instances, and the goal is to learn a model that predicts the label of the bag. In image analysis tasks, the entire image is usually taken as a bag, and image patches obtained by dividing the entire image are taken as instances.

[0007] In medical image analysis, multiple-instance learning has two main advantages: first, dividing images into small patches (or instances) within this framework can help the model pay more attention to relatively small lesion areas; second, learning multiple instances can reduce the sensitivity of the model to image quality changes. Therefore, many medical image-related works use a multiple-instance learning scheme and design various fusion strategies for multiple-instance features. Traditional schemes usually use non-learnable aggregators such as average pooling and max pooling. Recently, some studies have proposed weighting the instances in the bag, achieving joint classification through instance-level attention. Wang et al. introduced uncertainty quantification, using the confidence of the instance to filter out noisy instances, and then combined a recurrent neural network to obtain a bag-level prediction result. These studies show that parameterized aggregation methods have an advantage over simple non-learnable aggregation methods. However, the above studies mostly ignore the relationship between different instances in the same bag, which can lead to the loss of information. Although some studies have combined inter-instance dependencies for aggregation, they do not explicitly construct and represent the prior relationship between instances. Therefore, the multiple-instance learning strategy still has deficiencies and challenges in medical image analysis tasks, which often result in the loss of neighborhood information and the lack of completeness of the detection target, leading to reduced diagnostic performance of the model. SUMMARY

[0008] In view of the above-mentioned disadvantages of the prior art, the purpose of the present application is to provide a pancreatic cancer diagnosis method, system, medium and electronic device, which can realize intelligent early diagnosis of pancreatic cancer based on only plain CT images, has high accuracy and strong applicability.

[0009] In a first aspect, the present application provides a pancreatic cancer diagnosis method, which comprises the following steps: obtaining an abdominal plain CT image; dividing a pancreatic region in the abdominal plain CT image into image blocks and extracting image block features and prediction probability values of the image blocks; constructing corresponding case-level graphs based on image block features and prediction probability values of different case regions, the different case regions including a non-pancreatic cancer case pancreatic region, a pancreatic cancer case tumor region and a pancreatic cancer case non-tumor region; training a graph neural network for pancreatic cancer prediction based on the graphs to obtain a pancreatic cancer prediction result based on the trained graph neural network.

[0010] In an implementation form of the first aspect, dividing the pancreatic region in the abdominal plain CT image into image blocks comprises the following steps:

[0011] Obtaining the pancreatic region in the abdominal plain CT image;

[0012] Taking a circumscribed cube of the pancreatic region as a region of interest;

[0013] Segmenting the region of interest into image blocks of a preset size.

[0014] In an implementation form of the first aspect, the extracting the image block features and the prediction probability values of the image blocks comprises the following steps:

[0015] training a convolutional neural network for extracting image block features and probability values of image blocks;

[0016] inputting the image blocks into the trained convolutional neural network to obtain the image block features and the probability values corresponding to the image blocks.

[0017] In an implementation form of the first aspect, the constructing the case-level graph based on the image block features and the prediction probability values of different case regions respectively comprises the following steps:

[0018] for each case region, concatenating the corresponding image block features and prediction probability values to form a node feature matrix of the case-level graph;

[0019] calculating the spatial clustering measure and the texture similarity measure between each pair of corresponding image blocks, and combining the spatial clustering measure and the texture similarity measure to construct an adjacency matrix of the case-level graph.

[0020] In an implementation form of the first aspect, the graph neural network sets a first loss function for the graphs corresponding to the non-pancreatic cancer case pancreatic region and the pancreatic cancer case tumor region; sets a second loss function for the graphs corresponding to the pancreatic cancer case tumor region and the pancreatic cancer case non-tumor region, and constructs a total loss function of the graph neural network based on the first loss function and the second loss function.

[0021] In an implementation form of the first aspect, the total loss function L = (L patient + L con × w con ) / (1 + w con ), wherein L patient is the first loss function, L con is the second loss function, and w con is a weighting parameter.

[0022] In an implementation form of the first aspect, the method further comprises evaluating the performance of the trained graph neural network; wherein the evaluation parameters comprise one or more combinations of accuracy, sensitivity, specificity, positive predictive rate, negative predictive rate, receiver operating characteristic curve and area under the curve.

[0023] In a second aspect, the present application provides a pancreatic cancer diagnosis system, the system comprising an acquisition module, an extraction module, a construction module and a training module;

[0024] The acquisition module is configured to acquire an abdominal plain scan CT image;

[0025] The extraction module is configured to divide a pancreas region in the abdominal plain scan CT image into image blocks, and extract image block features and prediction probability values of the image blocks;

[0026] The construction module is configured to construct a corresponding case level graph based on the image block features and the prediction probability values of different case regions, the different case regions including a non-pancreatic cancer case pancreas region, a pancreatic cancer case tumor region, and a pancreatic cancer case non-tumor region;

[0027] The training module is configured to train a graph neural network for pancreatic cancer prediction based on the graph, so as to obtain a pancreatic cancer prediction result based on the trained graph neural network.

[0028] In a third aspect, the present application provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the pancreatic cancer diagnosis method described above.

[0029] In a fourth aspect, the present application provides an electronic device, comprising a processor and a memory.

[0030] The memory is configured to store a computer program.

[0031] The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the pancreatic cancer diagnosis method described above.

[0032] As described above, the pancreatic cancer diagnosis method, system, medium and electronic device of the present application have the following beneficial effects:

[0033] (1) The intelligent early detection of pancreatic cancer can be realized based on only the plain scan CT image, which is of great significance for the realization and popularization of pancreatic cancer screening and early diagnosis;

[0034] (2) By dividing the image blocks to extract features, and designing a graph neural network to fuse the image block features according to the tumor characteristics, the great difficulty brought by small target regions is overcome, and relatively fine and complete discriminative features can be extracted; and once the model is trained, no tumor label is needed in the use process, which is time-saving and convenient to use;

[0035] (3) Local fine features are extracted using a multi-instance framework, and then a graph structure is modeled using the spatial position relationship and texture feature relationship of different instances in the same bag, and adaptive fusion of case level overall features is realized through a graph neural network;

[0036] (4) By embedding the causal graph neural network, the mining ability of the overall characteristics of the tumor and the strengthening ability of the causal discriminative features are improved, so that the pancreatic cancer detection accuracy is significantly improved, and stable generalization can be realized on a multi-center data set. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A flowchart showing the pancreatic cancer diagnosis method of the present application in an embodiment;

[0038] Figure 2 A schematic diagram showing the framework of the pancreatic cancer diagnosis method of the present application in an embodiment;

[0039] Figure 3 A schematic diagram showing the division of the pancreatic region into image blocks in an embodiment of the present application;

[0040] Figure 4 A schematic diagram showing the construction of the case-level graph in an embodiment of the present application;

[0041] Figure 5 A schematic diagram showing the structure of the pancreatic cancer diagnosis system of the present application in an embodiment;

[0042] Figure 6 A schematic diagram showing the structure of the electronic device of the present application in an embodiment. DETAILED DESCRIPTION

[0043] The embodiments of the present application will be described in detail below with specific reference to the drawings. Those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied in different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component can be a random change in number, shape and size, and the layout of the components can be more complex.

[0045] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0046] As shown in Figure 1 and Figure 2 In an embodiment, the pancreatic cancer diagnosis method of the present application includes the following steps:

[0047] Step S1: Obtain abdominal plain CT images.

[0048] Specifically, the present invention only requires acquiring plain CT images of the abdomen. In one embodiment, the thickness of the plain CT images ranges from 0.8 to 5 mm, and the imaging equipment manufacturers include SIEMENS, GE, TOSHIBA, Philips, and UIS. Resampling is used to standardize the slice thickness of all plain CT images to 5 mm.

[0049] Step S2: Divide the pancreatic region in the abdominal plain CT image into image blocks, and extract the image block features and prediction probability values ​​of the image blocks.

[0050] Specifically, in order to focus on the relatively small tumor region from the plain CT image, this invention designs a basic framework similar to multi-instance learning to extract fine local features of the tumor. Specifically, the pancreatic region image is cropped into image patches, and local features are extracted from each image patch to address the difficulty of extracting effective features from a small tumor region in the image.

[0051] In one embodiment, dividing the pancreatic region in the abdominal plain CT image into image blocks includes the following steps:

[0052] 21) Obtain the pancreatic region from the abdominal plain CT image.

[0053] Specifically, the pancreatic region in the plain CT image is determined using methods such as pancreatic segmentation models or manual annotation.

[0054] 22) The outer cube of the pancreatic region is taken as the region of interest, thereby removing the non-pancreatic portion that accounts for a large proportion.

[0055] 23) Divide the region of interest into image blocks of a preset size.

[0056] Among them, such as Figure 3 As shown, within the region of interest, a window-moving method is used slice by slice, starting from the top left corner and translating with a fixed step size to crop the image into sub-regions, i.e., image blocks. Preferably, the step size and the image block size were determined to be 25 and 50×50 (in pixels) after multiple experiments.

[0057] In one embodiment, extracting the image patch features and predicted probability values ​​of the image patch includes the following steps:

[0058] A) Train a convolutional neural network to extract image patch features and probability values ​​of image patches.

[0059] wherein the image block containing the tumor is labeled as 1 and the image block not containing the tumor is labeled as 0. A five-fold cross-validation strategy is adopted, and the plain CT images in the dataset are divided into 5 fixed independent folds. Four of the 5 folds are used for training, and the remaining 1 fold is used for testing. The image blocks in the training set are used to train a 6-layer convolutional neural network, and the initial learning rate is set to 0.001, the batch size is set to 2560 image blocks, and the Adam optimizer is used as the optimization strategy. During the training of the convolutional neural network, the error between the output score and the input score is calculated according to the loss function, and the parameters of the convolutional neural network are adjusted and optimized by using the back propagation algorithm. Preferably, the training set is trained for 50 rounds, and after the training is completed, the test set is input into the trained convolutional neural network to obtain the image block features and the prediction probability value of the image block. The prediction probability value of whether the image block contains a tumor is output through a fully connected layer.

[0060] The convolutional neural network adopts cross-entropy as a constraint term to calculate the prediction loss at the image block level wherein t represents the index of the image block, represents the true class of the image block t, is the prediction probability value of the image block t, and w0 and w1 are used to adjust the penalty weight of two types of image blocks (containing or not containing a tumor).

[0061] B) inputting the image block into the trained convolutional neural network to obtain the image block features and the probability value corresponding to the image block.

[0062] wherein the image block obtained by step S1 is input into the trained convolutional neural network, and the image block features and the probability value corresponding to the image block are obtained.

[0063] Step S3, based on the image block features and the prediction probability value of different case regions, a corresponding case-level graph is constructed, and the different case regions include the pancreatic region of the non-pancreatic cancer case, the tumor region of the pancreatic cancer case, and the non-tumor region of the pancreatic cancer case.

[0064] Specifically, the present application combines spatial aggregation measurement and texture similarity measurement to encode the association between image blocks, and uses a graph neural network to maintain the integrity of the case-level features and the reliability of the case cancer prediction. In addition, a causal mechanism is embedded in the graph neural network, and a causal feature is constructed using part of the tumor region of the pancreatic cancer patient, and a non-causal feature is constructed using part of the non-tumor region of the pancreatic cancer patient. The stability of model decision is improved by guiding the model to distinguish between causal features and non-causal features.

[0065] For a given case image patch, the degree of association between each pair of image patches is measured using spatial clustering and texture similarity metrics to obtain the adjacency matrix of the corresponding case-level graph. The features and probability values ​​of each image patch are concatenated to form the node feature matrix of the case-level graph. Specifically, partial image patches from the tumor region and partial image patches from the non-tumor region of pancreatic cancer cases are used to construct causal graphs and non-causal graphs, respectively. Preferably, the number of image patches used to construct each graph is determined to be 50 after multiple experiments.

[0066] In one embodiment, such as Figure 4 As shown, constructing corresponding case-level maps based on image patch features and predicted probability values ​​for different case regions includes the following steps:

[0067] 31) For each case region, the corresponding image patch features and predicted probability values ​​are concatenated to form the node feature matrix of the graph at the case level.

[0068] For each case, a fixed number τ image patches with a high probability of containing a tumor are obtained. The case index is denoted as p, and the τ image patches of case p constitute the corresponding image G. p =(V p E p The set of nodes V p The image patch features and predicted probability values ​​of τ image patches are concatenated to form the node feature matrix f. p .

[0069] 32) Calculate the spatial clustering measure and texture similarity measure between each pair of corresponding image patches, and combine the spatial clustering measure and the texture similarity measure to construct the adjacency matrix of the graph at the case level.

[0070] Wherein, the edge set E of the graph p From the adjacency matrix A p This is used to encode and represent the strength of connections between various image patches. The fusion of multiple local features (i.e., image patch features) for a case is crucial. Due to the spatial clustering and image texture feature similarity among image patches from the tumor region, this invention considers both feature similarity and spatial distance metrics to achieve adaptive fusion of local feature representations. The adjacency matrix A... p It includes two components: spatial clustering measure and the texture similarity measure Spatial clustering measure and the texture similarity measure By combining them, the corresponding adjacency matrix A can be obtained. p .

[0071] In the present application, the physical distance between the center points of each pair of image blocks is calculated based on the spatial clustering measurement module, and is mapped through a function that is monotonically decreasing with respect to a non-negative spatial independent variable to obtain the connection strength between the image blocks to form the spatial clustering measurement part of the adjacency matrix Therefore, the closer the spatial distance between the image blocks, the greater the corresponding mapping value, representing the stronger the connection between the two image blocks, which is more conducive to the effective transmission of features in subsequent graph convolution.

[0072] In the present application, the image block features of each image block are mapped to a low-dimensional space through two learnable functions ψ and θ based on the texture similarity measurement module, and the similarity between each pair of image block features is calculated through matrix multiplication in the low-dimensional space, and the normalized similarity value is calculated using the softmax function, thereby constructing the texture similarity measurement part of the adjacency matrix The image blocks derived from the tumor region have more similar texture features than the image blocks derived from the non-tumor region, and the corresponding mapping value is greater, representing the stronger the connection between the two image blocks, which is more conducive to the effective transmission of features in subsequent graph convolution.

[0073] Step S4, training a graph neural network for pancreatic cancer prediction based on the graph, to obtain a pancreatic cancer prediction result based on the trained graph neural network.

[0074] Specifically, the present application proposes a causal feature enhancement mechanism in the framework of graph neural network, which is realized through multiple graph neural network branches, aiming to enable the model to make a generalizable decision based on causal discriminative features.

[0075] From a causal perspective, the factors that lead to the inference results of deep learning models include non-causal correlation and causality. Non-causal correlation factors are factors that interfere with the discrimination of observation targets (such as background, brightness, noise, etc.), which limit the generalization performance of the model and make the model learn biased results. Causal factors are key discriminative factors of observation targets, and correctly mining such features is a prerequisite for the model to achieve stable and generalizable prediction. In practical work, non-causal features mainly come from features outside the patient's lesion area and features of suspected lesion areas in healthy populations; causal features are clinically meaningful and interpretable diagnostic features, mainly coming from the lesion, i.e., the pancreatic tumor area.

[0076] Therefore, in order to make the graph neural network achieve accurate diagnosis by using causal features, the application respectively constructs two types of features, causal and non-causal, to strengthen causal features and suppress non-causal features, so as to make the graph neural network capture the essential information of the tumor and achieve causal diagnosis. Specifically, on the basis of using the graph neural network to classify cancer cases and non-cancer cases, a branch for causal reinforcement is added, and this branch is still implemented by using the graph neural network. In this branch, a causal graph is constructed by using part of the image blocks of the tumor region of the pancreatic cancer case, and a non-causal graph is constructed by using part of the image blocks of the non-tumor region of the pancreatic cancer case. The two network branches share parameters, and cross-entropy is used as the loss function for optimization.

[0077] According to the above causal reinforcement mechanism, the graph corresponding to the training set case obtained in step S3 is input into the two branches of the graph neural network embedded with the causal mechanism, and the graph neural network is composed of two graph convolution layers. The label of the graph constructed by the pancreatic cancer case tumor region image block is set to 1, and the label of the graph constructed by the non-pancreatic cancer case image block and the graph constructed by the non-tumor region image block of the pancreatic cancer case is set to 0. The initial learning rate is set to 0.01, the batch size is set to 64 graphs, and the Adam optimizer is used as the optimization strategy. In the training process of the graph neural network, the error between the output score and the input score is calculated according to the loss function, and the parameters of the network are adjusted and optimized by using the back propagation algorithm. In this embodiment, the training set is trained for 80 rounds, and the learning rate is adaptively decayed during the process. After the training is completed, the test set is input into the saved trained model to obtain the prediction result of the pancreatic cancer of the case.

[0078] wherein the graph neural network sets a first loss function for the graph corresponding to the non-pancreatic cancer case pancreatic region and the pancreatic cancer case tumor region sets a second loss function for the graph corresponding to the pancreatic cancer case tumor region and the non-tumor region of the pancreatic cancer case wherein, and respectively represent the true class and the prediction result of the case p; q represents the case index used for the causal reinforcement branch, y q and are the labels of the causal and non-causal graph samples, and are the prediction results of the causal and non-causal graph samples.

[0079] Finally, the total loss function L of the graph neural network is constructed based on the first loss function and the second loss function, L = (L patient + L con ) x w con ) / (1 + w con ), wherein L patient is the first loss function, Lcon is the second loss function, w con is a weighting parameter. By changing the size of w con , the influence degree of the two loss terms can be adjusted, and the specific value is obtained through multiple experiments.

[0080] When obtaining the pancreatic cancer prediction result based on the trained graph neural network, first, the image block of the plain scan CT image to be detected is extracted, then the graph is constructed based on the image block, and after the graph construction is completed, the information transmission between the image blocks is realized through the information propagation mechanism of the graph neural network. In the graph neural network, the node features are pooled and fused after several graph convolution layers to generate case-level features, and then the pancreatic cancer prediction result of the case is obtained by using a fully connected layer and an activation function.

[0081] In an embodiment, after the training of the graph neural network is completed, its performance also needs to be evaluated. The evaluation parameters include one or more combinations of accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive rate (PPV), negative predictive rate (NPV), receiver operating characteristic curve (ROC), and area under the curve (AUC). Among them, ACC=(TP+TN) / (TP+TN+FN+FP), SEN=TP / (TP+FN), SPE=TN / (TN+FP), PPV=TP / (TP+FP), NPV=TN / (TN+FN), TP, TN, FP and FN respectively represent the number of true positive, true negative, false positive and false negative classification samples.

[0082] Compared with other existing related researches in the field of early intelligent detection of pancreatic cancer images, the superiority of the pancreatic cancer diagnosis method of the present application can be shown through the experimental results on the experimental data set. The specific analysis of the experimental results is as follows:

[0083] (1) Classification result analysis

[0084] Ten repeated five-fold cross-validation experiments were conducted on the experimental data set to evaluate the pancreatic cancer detection accuracy and stability of the present application. The experimental results are shown in Table 1, which includes the analysis of the accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive rate (PPV), negative predictive rate (NPV), and AUC value of the model classification. The average performance of the six evaluation indicators is 89.3%, 91.0%, 87.1%, 90.8%, 87.5%, and 95.1%, respectively, and the numerical fluctuations are very small, indicating that the present application has good and stable classification performance.

[0085] Table 1, pancreatic cancer classification results of the present application on the experimental data set

[0086] ACC SEN SPE PPV NPV AUC Mean 89.3％ 91.0％ 87.1％ 90.8％ 87.5％ 95.1％ Standard deviation 1.6％ 1.5％ 2.5％ 1.7％ 1.9％ 0.8％ Maximum 93.1％ 94.2％ 91.6％ 94.1％ 91.7％ 96.9％

[0087] (2) Comparison with advanced medical image classification models

[0088] In the process of comparison with three advanced and robust medical image classification methods, the experimental data set used in the present application is used in the network structure in their official open source code respectively. The methods compared include VGG, deep learning model based on image block, Transformer, and the classification results of each model on the experimental data set are shown in Table 2. The present application is superior to these methods in accuracy and AUC, and achieves the optimal detection performance.

[0089] Table 2, comparison of the present application with advanced medical image classification methods

[0090] Advanced medical image classification model ACC AUC VGG 72.4％ 80.2％ Image patch-based deep learning model 83.8％ 89.5％ Transformer 84.1％ 91.6％ The model proposed by the present invention 89.3％ 95.1％

[0091] (3) Analysis of generalization performance

[0092] In order to verify the generalization performance of the present application, tests were carried out on three independent test sets from different centers, and the sample size and classification results of the three independent test sets are shown in Table 3. Among them, PC represents pancreatic cancer cases, and non-PC represents non-pancreatic cancer cases. Since independent test set II and independent test set III only contain pancreatic cancer cases, only accuracy and sensitivity are calculated. The present application shows good classification ability and stable generalization performance on the three independent test sets, respectively achieving an accuracy of 83.5%, 81.1% and 85.0%, which proves the stability and generalization of the present application in pancreatic cancer screening, which is mainly due to the design of the graph neural network embedded with causal mechanism in the present application.

[0093] Table 3, multi-center generalization performance of the present application

[0094]

[0095] (4) Analysis of ablation experiment

[0096] In order to verify the necessity and superiority of all strategies, the present application was subjected to item-by-item ablation experiment, and the results are shown in Table 4. Compared with the baseline model, the three multi-instance feature fusion schemes using graph convolution module significantly improve the classification effect, with the highest accuracy improvement of about 3%. Among them, the graph neural network combining spatial aggregation measurement module and texture feature similarity measurement module simultaneously shows the best results. The causal reinforcement mechanism can further improve the model accuracy by about 2% on the basis of measuring the graph neural network, and other evaluation indicators are also improved. The excellent performance of quantitative analysis results confirms the effectiveness of the graph neural network and causal reinforcement mechanism of the present application.

[0097] Table 4, ablation experiment results of the present application

[0098]

[0099] Therefore, it can be known that the pancreatic cancer diagnosis method effectively extracts key discriminative features in the plain CT image that lead to a diagnosis result, realizes automatic screening and early diagnosis of the pancreatic cancer that can be generalized, and overcomes limitations of conventional detection means based on an enhanced image mode, thereby providing a potential new tool with accuracy, stability and clinical feasibility for early diagnosis and screening of the pancreatic cancer.

[0100] The protection scope of the pancreatic cancer diagnosis method according to the embodiments of the present application is not limited to the execution order of the steps listed in the embodiments, and any scheme realized by adding, reducing or replacing steps of the prior art according to the principle of the present application is included in the protection scope of the present application.

[0101] The embodiments of the present application further provide a pancreatic cancer diagnosis system, which can realize the pancreatic cancer diagnosis method according to the present application, but the implementation device of the pancreatic cancer diagnosis system according to the present application includes but is not limited to the structure of the pancreatic cancer diagnosis system listed in the embodiments, and any structure deformation and replacement of the prior art according to the principle of the present application is included in the protection scope of the present application.

[0102] As shown in FIG. 1, Figure 5 In an embodiment, the pancreatic cancer diagnosis system according to the present application includes an acquisition module 51, an extraction module 52, a construction module 53 and a training module 54.

[0103] The acquisition module 51 is configured to acquire an abdominal plain CT image.

[0104] The extraction module 52 is connected to the acquisition module 51 and configured to divide a pancreatic region in the abdominal plain CT image into image blocks and extract image block features and prediction probability values of the image blocks.

[0105] The construction module 53 is connected to the extraction module 52 and configured to construct a graph at a case level based on the image block features and the prediction probability values of different case regions, the different case regions including a non-pancreatic cancer case pancreatic region, a pancreatic cancer case tumor region and a pancreatic cancer case non-tumor region.

[0106] The training module 54 is connected to the construction module 53 and configured to train a graph neural network for pancreatic cancer prediction based on the graph, so as to acquire a pancreatic cancer prediction result based on the trained graph neural network.

[0107] The structures and principles of the acquisition module 51, the extraction module 52, the construction module 53 and the training module 54 correspond to the steps of the pancreatic cancer diagnosis method one by one, and thus will not be described here again.

[0108] In several embodiments provided in the present application, it should be understood that the disclosed system, device or method can be implemented in other ways. For example, the device embodiments described above are merely illustrative, for example, the division of the modules / units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or units can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be through some interface, indirect coupling or communication connection between the modules or units, which can be electrical, mechanical or other forms.

[0109] The modules / units described as separate components can or can not be physically separated, and the components shown as modules / units can or can not be physical modules, that is, they can be located in one place or distributed on multiple network units. Part or all of the modules / units can be selected according to actual needs to achieve the purpose of the embodiments of the present application. For example, the functional modules / units in each embodiment of the present application can be integrated in a processing module, or each module / unit can be physically present separately, or two or more modules / units can be integrated in one module / unit.

[0110] Those of ordinary skill in the art should further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0111] The embodiment of the present application further provides a computer readable storage medium. A person skilled in the art can understand that all or part of the steps of the pancreatic cancer diagnosis method described above can be completed by a processor instructed by a program, and the program can be stored in a computer readable storage medium, and the storage medium is a non-transitory medium, for example, a random access memory, a read only memory, a flash memory, a hard disk, a solid state disk, a magnetic tape, a floppy disk, an optical disc and any combination thereof. The storage medium can be any available medium accessible by a computer or a data storage device such as a server, a data center and the like integrated with one or more available medium sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a digital video disc (DVD)) or a semiconductor medium (for example, a solid state disk (SSD)) and the like.

[0112] The embodiment of the present application further provides an electronic device. The electronic device comprises a processor and a memory.

[0113] The memory is used for storing a computer program.

[0114] The memory comprises a ROM, a RAM, a disk, a U disk, a memory card or an optical disc and the like various medium capable of storing program codes.

[0115] The processor is connected with the memory, and is used for executing the computer program stored in the memory, so that the electronic device executes the pancreatic cancer diagnosis method described above.

[0116] Preferably, the processor can be a general processor, including a central processing unit (CPU), a graphic processing unit (GPU), a network processor (NP) and the like; and can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0117] As Figure 6As shown, the electronic device of the present application is in the form of a general- purpose computing device. The components of the electronic device can include, but are not limited to, one or more processors or processing units 61, a main memory 62, and a bus 63 that couples various system components including the memory 62 to the processing unit 61.

[0118] The bus 63 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics bus (e.g., an Accelerated Graphics Port, or AGP bus) and a processor or local bus using any of a variety of bus architectures. By way of example, these bus architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0119] The electronic device typically includes a variety of computer system readable media. Such media can be any available media that is located either internally or externally to the electronic device. It includes storage of removable and non-removable, volatile and non-volatile computer system storage media.

[0120] The memory 62 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 621 and / or cache memory 622. The electronic device can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 623 can be provided for reading from and writing to a non-removable, non-volatile magnetic media (e.g., a "hard drive"). Figure 6 Not shown, a removable / non-removable, volatile / non-volatile computer system storage medium can be used to store data including one or more programs and / or program modules. Figure 6 Not shown, a removable / non-removable, volatile / non-volatile computer system storage medium can be used to store data including one or more programs and / or program modules.

[0121] A program / utility 624 having a set (at least one) of program modules 6241 can be stored in, for example, memory 62 by way of example, such program modules 6241 include an operating system, one or more application programs, other program modules, and program data, each of which or a combination can include implementation of a networking environment. Program modules 6241 generally carry out the functions and / or methodologies of embodiments of the present application as described herein.

[0122] The electronic device can also communicate with one or more external devices such as a keyboard or a pointing device, displays, etc.; other devices associated with the electronic device; and / or one or more devices that enable Figure 6 The network adapter 65 can communicate with the other modules of the electronic device through the bus 63. As will be appreciated, various other hardware and / or software modules are typically implemented in electronic devices, e.g., microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc. which are not further shown for the sake of brevity and clarity.

[0123] The foregoing merely illustrates the principles of the application and applies example buffering techniques only to illustrate the principles of the application. Furthermore, those skilled in the art will recognize that the foregoing description and examples have been presented for the purposes of illustration and example only. The description as set forth is not intended to be exhaustive or to limit the application to the precise form disclosed. Many modifications and variations are possible in light of the above teaching without departing from the spirit and scope of the forthcoming claims. Accordingly, it should be understood that the application can be practiced otherwise than as specifically set forth herein.

Claims

1. A method for diagnosing pancreatic cancer, characterized in that: The method includes the following steps: Obtain plain CT images of the abdomen; The pancreatic region in the abdominal plain CT image is divided into image blocks, and the image block features and prediction probability values ​​of the image blocks are extracted. Based on the image patch features and predicted probability values ​​of different case regions, corresponding case-level maps are constructed respectively. The different case regions include the pancreatic region of non-pancreatic cancer cases, the tumor region of pancreatic cancer cases, and the non-tumor region of pancreatic cancer cases. A graph neural network for pancreatic cancer prediction is trained based on the graph, so as to obtain pancreatic cancer prediction results based on the trained graph neural network. Extracting the image patch features and predicted probability values ​​of the image patch includes the following steps: Train a convolutional neural network, which is used to extract image patch features and probability values ​​of image patches; The image patch is input into a trained convolutional neural network to obtain the image patch features and probability values ​​corresponding to the image patch; The steps involved in constructing corresponding case-level maps based on image patch features and predicted probability values ​​for different case regions are as follows: For each case region, the corresponding image patch features and predicted probability values ​​are concatenated to form the node feature matrix of the graph at the case level; Calculate the spatial clustering measure and texture similarity measure between each pair of corresponding image patches, and combine the spatial clustering measure and the texture similarity measure to construct the adjacency matrix of the graph at the case level.

2. The method for diagnosing pancreatic cancer according to claim 1, characterized in that: Dividing the pancreatic region in the abdominal plain CT image into image blocks includes the following steps: Obtain the pancreatic region from the abdominal plain CT image; The outer cube of the pancreatic region is taken as the region of interest; The region of interest is divided into image blocks of a preset size.

3. The method for diagnosing pancreatic cancer according to claim 1, characterized in that: The graph neural network sets a first loss function for the graphs corresponding to the pancreatic region of the non-pancreatic cancer case and the tumor region of the pancreatic cancer case; sets a second loss function for the graphs corresponding to the tumor region of the pancreatic cancer case and the non-tumor region of the pancreatic cancer case; and constructs the total loss function of the graph neural network based on the first loss function and the second loss function.

4. The method for diagnosing pancreatic cancer according to claim 3, characterized in that: The total loss function L = (L patient +L con ×w con ) / (1+w con ), where L patient It is the first loss function, L con It is the second loss function, w con It is a trade-off parameter.

5. The method for diagnosing pancreatic cancer according to claim 1, characterized in that: It also includes evaluating the performance of the trained graph neural network; the evaluation parameters include one or more combinations of accuracy, sensitivity, specificity, positive predictive value, negative predictive value, receiver operating characteristic curve and area under the curve.

6. A pancreatic cancer diagnostic system, characterized in that: The system includes an acquisition module, an extraction module, a construction module, and a training module; The acquisition module is used to acquire abdominal plain CT images; The extraction module is used to divide the pancreatic region in the abdominal plain CT image into image blocks, and extract the image block features and prediction probability values ​​of the image blocks; The construction module is used to construct corresponding case-level maps based on image patch features and predicted probability values ​​of different case regions. The different case regions include pancreatic regions in non-pancreatic cancer cases, tumor regions in pancreatic cancer cases, and non-tumor regions in pancreatic cancer cases. The training module is used to train a graph neural network for pancreatic cancer prediction based on the graph, so as to obtain pancreatic cancer prediction results based on the trained graph neural network. Extracting the image patch features and predicted probability values ​​of the image patch includes the following steps: Train a convolutional neural network, which is used to extract image patch features and probability values ​​of image patches; The image patch is input into a trained convolutional neural network to obtain the image patch features and probability values ​​corresponding to the image patch; The steps involved in constructing corresponding case-level maps based on image patch features and predicted probability values ​​for different case regions are as follows: For each case region, the corresponding image patch features and predicted probability values ​​are concatenated to form the node feature matrix of the graph at the case level; Calculate the spatial clustering measure and texture similarity measure between each pair of corresponding image patches, and combine the spatial clustering measure and the texture similarity measure to construct the adjacency matrix of the graph at the case level.

7. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the pancreatic cancer diagnosis method as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory to cause the electronic device to perform the pancreatic cancer diagnosis method according to any one of claims 1 to 5.

Images