Classification method and classification device for pancreatic neuroendocrine tumors

Abstract

The application discloses a classification method and device for pancreatic neuroendocrine tumors. The classification method comprises the following steps: segmenting tumor masks from enhanced CT images; for each tumor mask, generating a first peritumoral region and a second peritumoral region; dividing the tumor into multiple habitats for each tumor mask; extracting 3D radiomics features, multiple peritumoral microenvironment features, multiple tumor habitat features, multiple local pathological features and multiple global context features; selecting multiple target features for each enhanced CT image and calculating a radiomics score; inputting the radiomics score and clinical variables into a logistic regression model for training to obtain a trained logistic regression model; and using the trained logistic regression model to classify the enhanced CT images to be classified. The application can quickly and accurately realize G-grade classification of pancreatic neuroendocrine tumors, and provides a reliable non-invasive evaluation tool for clinical decision-making.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to classification methods, classification devices, media, electronic devices, and computer program products for pancreatic neuroendocrine tumors. Background Technology

[0002] Pancreatic neuroendocrine tumors (PNETs) are a rare type of tumor originating from peptidergic neurons and neuroendocrine cells in the pancreas. Their incidence has increased significantly in recent years, posing a serious threat to patients' health.

[0003] Accurate preoperative grading is crucial for treatment decisions in patients with PNETs. Current guidelines recommend active surveillance for G1 grade tumors smaller than 2 cm, while G2 and G3 grade tumors require immediate surgical resection. Therefore, accurate preoperative grading is essential to avoid overtreatment or treatment delay.

[0004] Endoscopic ultrasound-guided fine-needle aspiration biopsy (EUS-FNA) is currently the standard method for preoperative grading, but it has significant clinical limitations, including sampling errors due to intratumoral heterogeneity, insufficient tissue volume in up to 40% of cases, procedural complications such as bleeding and pancreatitis, and contraindications in patients with coagulation disorders or unfavorable anatomical structures. These limitations create an urgent clinical need for accurate, non-invasive grading alternatives. Summary of the Invention

[0005] This application provides a classification method, a classification device, a medium, an electronic device, and a computer program product for pancreatic neuroendocrine tumors.

[0006] In a first aspect, embodiments of this application provide a classification method for pancreatic neuroendocrine tumors, used in an electronic device, the classification method comprising:

[0007] The segmentation step uses a pre-trained segmentation model to segment multiple enhanced CT images from multiple patients to extract the tumor mask in each enhanced CT image.

[0008] In the generation step, for each tumor mask, a spherical structural element with a first radius is expanded to generate a first peritumoral region, and a spherical structural element with a second radius is expanded to generate a second peritumoral region, wherein the first radius is smaller than the second radius;

[0009] The segmentation step involves dividing the tumor into multiple habitats for each tumor mask, based on each voxel within the tumor.

[0010] The extraction steps include extracting 3D radiomics features from each enhanced CT image, extracting multiple local pathological features and multiple global context features from each tumor mask, extracting multiple peritumoral microenvironment features from the first and second peritumoral regions, and extracting multiple tumor habitat features from the multiple habitats.

[0011] In the selection step, for each enhanced CT image, multiple target features are selected from the multiple peritumoral microenvironment features, multiple tumor habitat features, multiple 3D radiomics features, multiple local pathological features, and multiple global context features, and a radiomics score is calculated for each enhanced CT image.

[0012] The training steps involve inputting the radiomics scores of multiple enhanced CT images and the clinical variables of multiple enhanced CT images into the logistic regression model for training, resulting in the trained logistic regression model.

[0013] The classification step involves using the trained logistic regression model to perform G-level classification on the enhanced CT images to be classified.

[0014] Secondly, the present invention provides a classification device for pancreatic neuroendocrine tumors, the classification device comprising:

[0015] The segmentation unit uses a pre-trained segmentation model to segment multiple enhanced CT images from multiple patients to extract the tumor mask in each enhanced CT image.

[0016] The generation unit, for each tumor mask, expands a spherical structural element with a first radius to generate a first peritumoral region, and expands a spherical structural element with a second radius to generate a second peritumoral region, wherein the first radius is smaller than the second radius;

[0017] The tumor is divided into multiple habitats for each tumor mask based on each voxel within the tumor.

[0018] The extraction unit extracts 3D radiomics features from each enhanced CT image, extracts multiple local pathological features and multiple global context features from each tumor mask, extracts multiple peritumoral microenvironment features from the first peritumoral region and the second peritumoral region, and extracts multiple tumor habitat features from the multiple habitats.

[0019] The selection unit selects multiple target features from the multiple peritumoral microenvironment features, multiple tumor habitat features, multiple 3D radiomics features, multiple local pathological features, and multiple global context features for each enhanced CT image, and calculates the radiomics score for each enhanced CT image.

[0020] The training unit inputs the radiomics scores of multiple enhanced CT images and the clinical variables of multiple enhanced CT images into the logistic regression model for training, and obtains the trained logistic regression model.

[0021] The classification unit uses the trained logistic regression model to perform G-level classification on the enhanced CT image to be classified.

[0022] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the classification method for pancreatic neuroendocrine tumors as described in any of the first aspects.

[0023] Fourthly, embodiments of the present invention provide an electronic device, including: one or more processors; one or more memories; the one or more memories storing one or more programs, which, when executed by the one or more processors, cause the electronic device to perform the classification method for pancreatic neuroendocrine tumors described in the first aspect.

[0024] Fifthly, embodiments of this application provide a computer program product including computer-executable instructions that are executed by a processor to implement the classification method for pancreatic neuroendocrine tumors described in the first aspect.

[0025] In this invention, multiple peritumoral microenvironment features, tumor habitat features, local pathological features, global contextual features, and 3D radiomics features are extracted from each enhanced CT image, achieving extremely rich multi-dimensional feature extraction. Then, a hierarchical feature selection algorithm is used to select multiple target features from these extracted multi-dimensional features. Based on these target features, the radiomics score M-DLR for each enhanced CT image is calculated. Finally, a logistic regression model is trained (constructed) based on the radiomics score M-DLR of each enhanced CT image and the corresponding patient's clinical information to obtain a classification model. Thus, based on the output of the classification model, the G-grade classification of pancreatic neuroendocrine tumors can be achieved quickly and accurately, providing a reliable non-invasive assessment tool for clinical decision-making.

[0026] This invention, through multi-center clinical validation on a dataset of, for example, 407 patients (244 in the training set, 106 in the validation set, and 57 in the external test set), achieved excellent classification performance: training set AUC = 0.92, validation set AUC = 0.89, and external test set AUC = 0.87, significantly outperforming models using only clinical features (ΔAUC = 0.15–0.22, p < 0.001). The model achieved 100% sensitivity in external validation and provided accurate grading diagnoses for 13 out of 25 patients (52%) with insufficient EUS-FNA tissue. Attached Figure Description

[0027] Figure 1 According to an embodiment of this application, a flowchart of a classification method for pancreatic neuroendocrine tumors is shown;

[0028] Figure 2 According to an embodiment of this application, a structural diagram of a classification device for pancreatic neuroendocrine tumors is shown;

[0029] Figure 3 According to an embodiment of this application, a block diagram of an electronic device is shown. Detailed Implementation

[0030] The illustrative embodiments of this application include, but are not limited to, methods for classifying pancreatic neuroendocrine tumors, devices for classifying pancreatic neuroendocrine tumors, media, electronic devices, and computer program products.

[0031] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0032] Figure 1 This application illustrates a classification method for pancreatic neuroendocrine tumors, used in an electronic device. Specifically, in segmentation step S11, a pre-trained segmentation model is used to segment multiple enhanced CT images from multiple patients to extract the tumor mask in each enhanced CT image.

[0033] In this embodiment, for example, the pre-trained segmentation model is based on the nnMamba model. The nnMamba model employs an encoder-decoder structure. The encoder contains five downsampling layers, each containing a combination of Mamba blocks and convolutional blocks. The Mamba blocks are responsible for capturing long-range dependencies, while the convolutional blocks are responsible for extracting local features. The decoder employs a symmetrical structure, fusing feature information at different scales through skip connections. The network input is a multi-layer CT image of 512×512×N (N is the number of layers, typically 64-128 layers), and the output is a binary segmentation mask of the same size.

[0034] The nnMamba model is an optimized and improved version of the nnUNet framework, integrating the Mamba sequence modeling mechanism and the U-Net encoder-decoder structure. Compared with traditional CNN- or Transformer-based segmentation methods, nnMamba can more effectively capture long-range dependencies while maintaining lower computational complexity. This architecture is particularly suitable for handling the complex morphology and ambiguous boundaries of pancreatic tumors.

[0035] In generation step S12, for each tumor mask, a spherical structural element with a first radius is expanded to generate a first peritumoral region, and a spherical structural element with a second radius is expanded to generate a second peritumoral region, wherein the first radius is smaller than the second radius. In this embodiment, for example, the first radius is 2 mm and the second radius is 5 mm.

[0036] In this invention, for each segmented tumor mask, a morphological dilation operation is used to generate a peritumoral region. Specifically, the process is as follows: (1) the tumor mask is dilated by a spherical structural element with a radius of 2 mm to obtain the first peritumoral region (also known as the near-peritumoral region); (2) the tumor mask is dilated by a spherical structural element with a radius of 5 mm to obtain the second peritumoral region (also known as the far-peritumoral region); (3) a 2 mm peritumoral ring and a 5 mm peritumoral ring are extracted by mask subtraction operation; (4) the region outside the pancreas is excluded to ensure that the peritumoral analysis is limited to the pancreatic parenchyma.

[0037] In this invention, based on the tumor boundary, regions extending outward by 2 mm and 5 mm are systematically defined as the near-tumor periphery and the far-tumor periphery, respectively. This multi-scale design is based on tumor biology theory; the near-tumor periphery region mainly reflects the direct invasion and microangiogenesis of the tumor, while the far-tumor periphery region reflects the indirect effects of the tumor on surrounding tissues and inflammatory responses.

[0038] In segmentation step S13, for each tumor mask, the tumor is segmented into multiple habitats based on each voxel within the tumor.

[0039] For each tumor mask, the local entropy and CT intensity values ​​of each voxel within the tumor are calculated, and K-means clustering is used to perform cluster analysis on the entropy and CT intensity values, thereby dividing the tumor into multiple habitats.

[0040] Specifically, for each voxel within the tumor, a local entropy value (reflecting local texture complexity) and a CT intensity value (reflecting tissue density) are calculated. The local entropy value is calculated using a 3×3×3 neighborhood window, with the formula: Entropy = -Σp(i)×log2(p(i)), where p(i) is the probability of gray level i. The CT intensity value is directly obtained from the original image and then standardized.

[0041] Next, the K-means clustering algorithm was used to perform cluster analysis on the two-dimensional feature space (local entropy value and CT intensity value). The number of clusters K was determined by the elbow rule and was usually set to 3-5 habitats. The clustering results divided the tumor into different habitats (also known as subregions), such as low-entropy, low-intensity areas (which may correspond to necrotic areas) and high-entropy, high-intensity areas (which may correspond to active proliferative areas).

[0042] This invention proposes a tumor habitat analysis technique based on a dual index of local entropy and CT intensity, achieving precise quantification of intratumoral heterogeneity. Each habitat represents a tumor region with similar biological characteristics; that is, each habitat locates a tumor region with different functions in three-dimensional space, such as a high-proliferation area, a necrotic area, or a highly vascularized area.

[0043] In extraction step S14, 3D radiomics features are extracted from each enhanced CT image, multiple local pathological features and multiple global context features are extracted from each tumor mask, multiple peritumoral microenvironment features are extracted from the first and second peritumoral regions, and multiple tumor habitat features are extracted from multiple habitats.

[0044] This invention uses a pre-trained ResNet-50 network as a feature extractor, and the input layer is modified to accept multi-channel images (typically 5-7 channels, including the target slice and its adjacent slices). The last fully connected layer of the ResNet-50 network is removed, and the 2048-dimensional feature vector of the penultimate layer is used as the deep learning feature.

[0045] In this embodiment, the ResNet-50 network is used to extract 3D radiomics features from each enhanced CT image, extract multiple peritumoral microenvironment features from the first and second peritumoral regions, and extract multiple tumor habitat features from multiple habitats.

[0046] Specifically, the multiple peritumoral microenvironment features include the peritumoral features of the first and second peritumoral regions, the correlation between the peritumoral features and the intratumoral features, and the multiple tumor habitat features include the morphological features, texture features, and spatial relationship features between each habitat.

[0047] Multiple peritumoral microenvironment characteristics

[0048] Specifically, for each peritumor region in the first and second peritumor regions, the following peritumor features are extracted: (1) First-order statistical features: mean, standard deviation, skewness, kurtosis, minimum, maximum, median, quartiles, etc.; (2) Second-order texture features: contrast, correlation, energy, homogeneity, etc. based on the gray-level co-occurrence matrix (GLCM); (3) Higher-order features: short run emphasis, long run emphasis, gray-level non-uniformity, etc. based on the gray-level run matrix (GLRLM); (4) Shape features: volume, surface area, sphericity, density, etc. of the peritumor region.

[0049] The correlation characteristics between peritumoral and intratumoral features include: (1) Correlation analysis: Calculate the Pearson correlation coefficient between the peritumoral region and the corresponding intratumoral features (Pearson correlation coefficient is a statistical indicator that measures the degree and direction of linear correlation between two continuous variables); (2) Difference analysis: Calculate the percentage difference between the peritumoral region and the intratumoral features; (3) Gradient analysis: Calculate the feature gradient change from the tumor center to the boundary of the peritumoral region; (4) Ratio characteristics: Calculate the ratio between the peritumoral region and the intratumoral features, reflecting the degree of influence of the tumor on the surrounding tissues.

[0050] It is understandable that the first peritumoral region (i.e., the near peritumoral region) mainly reflects the tumor's invasiveness, while the second peritumoral region (i.e., the distant peritumoral region) mainly reflects the tumor's indirect impact on surrounding tissues. Therefore, this invention employs a differentiated feature extraction strategy. Specifically, features reflecting tumor invasiveness, such as texture heterogeneity, blood vessel density, and enhancement patterns, are extracted primarily from the near peritumoral region, while features reflecting the tumor's extent of influence, such as changes in tissue density, inflammatory markers, and the degree of structural distortion, are extracted primarily from the distant peritumoral region.

[0051] Understandably, by calculating the correlation, differences, gradient changes, and ratio characteristics between the peritumoral region and intratumoral features, the association characteristics between peritumoral and intratumoral features can be obtained, thereby constructing a composite feature reflecting the interaction between the tumor and the microenvironment.

[0052] Existing radiomics studies primarily focus on internal tumor features, with insufficient consideration given to the peritumoral microenvironment or limited to single-scale analysis. This invention systematically introduces multi-scale peritumoral analysis, filling this technological gap and providing a new technical means for the comprehensive assessment of tumor biological behavior.

[0053] Multiple tumor habitat characteristics

[0054] For each habitat, morphological and textural features were extracted independently. Morphological features included habitat volume, surface area, centroid location, and principal axis orientation. Textural features included statistical parameters based on matrices such as GLCM (Gray-Level Co-occurrence Matrix), GLRLM (Gray-Level Run-Length Matrix), and GLSZM (Gray-Level Size Zone Matrix).

[0055] The spatial relationship characteristics between various habitats include: (1) Adjacency analysis: calculate the contact area and boundary length between different habitats; (2) Distribution pattern: calculate the spatial distribution entropy and aggregation degree of habitats; (3) Geometric relationship: calculate the distance between the centroids of each habitat and their relative positions; (4) Volume ratio: calculate the proportion of each habitat in the total tumor volume.

[0056] In this invention, morphological and textural features are independently extracted for each habitat to construct a habitat-specific feature set. This method can capture the spatial heterogeneous distribution patterns within tumors, providing more refined biological information for tumor grading.

[0057] In this invention, spatial relationship features between various habitats are extracted, including geometric features such as relative location, size ratio, and boundary complexity of habitats, as well as topological features such as feature gradient changes and connectivity between habitats. These interaction features reflect the complexity of the internal tissue structure of the tumor.

[0058] Dynamic habitat evolution analysis: By comparing the changes in habitat distribution during the arterial and portal venous phases, the dynamic evolution pattern of habitat is analyzed, and time-series features reflecting tumor hemodynamic characteristics are extracted.

[0059] Existing methods for analyzing tumor heterogeneity often rely on simple partitioning based on a single indicator (such as intensity value or texture feature), lacking precise modeling of the complex biological structures within the tumor. The dual-indicator habitat analysis technology of this invention can more accurately identify and quantify intratumoral heterogeneity, providing crucial technical support for precision diagnosis.

[0060] 3D image omics features

[0061] In this invention, the PyRadiomics library is used to extract 3D radiomics features from each enhanced CT image, including seven different categories of features: morphological features (including 3D tumor diameter, volume, surface area, maximum diameter, and aspect ratio), first-order statistics, gray-level co-occurrence matrix (GLCM) features, gray-level dependency matrix (GLDM) features, gray-level size region matrix (GLSZM) features, gray-level run-length matrix (GLRLM) features, and nearest neighbor gray-level difference matrix (NGTDM) features. This multidimensional approach generates, for example, 1,409 comprehensive features, providing reliable quantitative characteristics of tumor morphology, intensity distribution, and texture complexity patterns.

[0062] The following details the extraction methods for multiple local pathological features and multiple global contextual features.

[0063] For each tumor mask, a lightweight ResNet18 network is used to extract multiple local pathological features (also known as “ResNet18 features” in Table 1) from multiple consecutive axial slices centered on the largest cross-section of the tumor, and a DINOv2 network is used to extract multiple global contextual features (also known as “large model features” in Table 1) from multiple consecutive axial slices. The multiple local pathological features and the multiple global contextual features are referred to as deep learning features.

[0064] To comprehensively analyze pancreatic pathological features, this invention employs a dual deep learning architecture: a 2.5D deep learning feature extraction method. This method extracts information from five key tomographic slices of the lesion (tumor). Specifically, the 2.5D deep learning feature extraction method involves selecting five consecutive slices centered on the largest cross-section of the lesion (in this embodiment, specifically layers -4, -2, 0, +2, and +4) to capture the local three-dimensional information (local pathological features) of the lesion. This architecture comprises two complementary models: a ResNet18 network and a DINOv2 network.

[0065] ResNet18: This model is pre-trained on ImageNet and excels at identifying minute lesions and classifying tissues, accurately capturing local details of lesions. 512 features are extracted from each slice, generating a total of 2560 feature vectors for each lesion.

[0066] DINOv2: This is a visual transformer model trained on large-scale natural image data, with the advantage of capturing global patterns and complex contextual information. 1024 features are extracted from each slice, generating a total of 5120 feature vectors for each lesion. Using DINOv2 for feature extraction on each of the five slices is suitable for characterizing a wide range of lesions and capturing complex spatial relationships.

[0067] By combining local insights with the global perspective of DINOv2, a rich and complementary set of deep learning features was constructed for pancreatic pathology analysis.

[0068] Existing deep learning methods for medical images mostly employ pure 2D or pure 3D architectures. 2D methods lack spatial context information, while 3D methods suffer from high computational complexity and are prone to overfitting. The 2.5D method of this invention cleverly balances computational efficiency and information integrity, providing a new technical path for medical image analysis and achieving efficient semantic feature extraction.

[0069] In this embodiment, for example, 13,542 features were extracted.

[0070] In step S15, for each enhanced CT image, multiple target features are selected from multiple peritumoral microenvironment features, multiple tumor habitat features, 3D radiomics features, multiple local pathological features, and multiple global context features, and a radiomics score for each enhanced CT image is calculated.

[0071] This invention employs a hierarchical feature selection algorithm, which can efficiently identify the optimal feature subset from a high-dimensional feature space. The specific process is as follows.

[0072] 1) Statistical significance screening: For each enhanced CT image, multiple peritumoral microenvironment features and multiple tumor habitat features were statistically significant, and features that were significantly correlated with tumor grade were retained.

[0073] For example, the Mann-Whitney U test was used to assess the statistical significance of all 13,542 extracted features. Patients were divided into two groups according to their pathological grade: G1 (low-grade) and G2 / G3 (high-grade). The statistical difference between the two groups was calculated for each feature, and features with a p-value less than 0.05 (i.e., significantly correlated with tumor grade) were retained. This effectively filters out noisy and irrelevant features. This step typically identifies approximately 3,000-4,000 significant features.

[0074] 2) Correlation analysis implementation control: Calculate the Spearman correlation coefficient between the retained features, identify highly correlated feature pairs based on the Spearman correlation coefficient, and retain the feature with greater information content in the feature pair by adopting the principle of maximizing variance.

[0075] Specifically, highly correlated feature pairs (correlation coefficient > 0.9) are identified through Spearman correlation coefficient analysis. The principle of maximizing variance is used to retain features with greater information content and remove redundant features. For feature pairs with an absolute Spearman correlation coefficient greater than 0.9, the following strategies are employed to select and retain features: (1) prioritizing features with larger variances; (2) prioritizing features with higher correlation to clinical variables; and (3) prioritizing features with stronger interpretability. This step typically reduces the number of features to 1500-2000. This step effectively reduces multicollinearity among features and improves model stability.

[0076] 3) Recursive feature elimination optimization: A recursive feature elimination algorithm based on random forest is adopted. Through iterative training and feature importance evaluation, the features that contribute the least to the prediction are gradually removed.

[0077] Specifically, the random forest algorithm is used as the basic evaluator, with the number of trees set to 100 and the maximum depth to 10. The algorithm process is as follows: (1) train the random forest model using all candidate features; (2) calculate the importance score of each feature; (3) remove the 10% of features with the lowest importance scores; (4) repeat steps 1-3 until the number of features drops to a preset threshold (usually 100-200).

[0078] Understandably, a recursive feature elimination algorithm based on random forests is employed, which iteratively removes features that contribute the least to prediction through training and feature importance evaluation. This algorithm can take into account the interactions between features and identify the feature combinations with the most predictive value.

[0079] 4) LASSO Regularized Selection: LASSO logistic regression is used to select the multiple target features.

[0080] Specifically, 10-fold cross-validation is used to determine the optimal regularization parameter λ. In each cross-validation fold, a LASSO logistic regression model is fitted using the training data, and its performance is evaluated on the validation data. The λ value that maximizes the validation AUC is selected as the optimal parameter. The final LASSO model typically selects the most important features. The LASSO method can select the most concise subset of features while ensuring predictive performance.

[0081] Existing feature selection methods often employ a single strategy, such as relying solely on statistical screening or machine learning methods, which can easily lead to insufficient feature selection or overfitting. The four-step hierarchical feature selection algorithm of this invention integrates the advantages of multiple methods, including statistics, correlation analysis, and machine learning, and can more reliably identify the optimal feature subset (target features).

[0082] In this embodiment, for example, the selected target features are shown in Table 1.

[0083] Table 1

[0084]

[0085]

[0086] In Table 1, ResNet18 features represent local pathological features, while big_model_feature represents global contextual features. Specifically, `resnet18_220.+02.` indicates the ResNet18 feature at layer 220, slice +02; `big_model_feature1017.+04` indicates the big_model (DINOv2) feature at layer 1017, slice +04. Other features follow the same pattern.

[0087] As can be understood, as shown in Table 1, the final selected target features include 2 local pathological features, 5 global contextual features, 2 peritumoral features, 1 3D radiomics feature, and 1 tumor habitat feature. In other words, the target features include local pathological features, global contextual features, peritumoral features, 3D radiomics features, and tumor habitat features.

[0088] Next, based on the selected target features, a radiomics score is calculated for each enhanced CT image.

[0089] Specifically, the radiomics score for each enhanced CT image is calculated based on the following formula (1):

[0090] M-DLR=1 / (1+exp(-(β0+Σβi×X i )))Formula (1);

[0091] Where M-DLR represents the radiomics score of an enhanced CT image, β0 is the intercept term (adjustment parameter), βi represents the weight of the i-th target feature, and X i Let M-DLR represent the standardized feature value of the i-th target feature, where the radiomics score ranges from 0 to 1. A higher M-DLR value indicates a higher degree of tumor malignancy.

[0092] Understandably, a corresponding radiomics score, M-DLR, can be calculated for each enhanced CT image.

[0093] In training step S16, the radiomics scores of multiple enhanced CT images and the clinical variables of multiple enhanced CT images are input into the logistic regression model for training, resulting in the trained logistic regression model (i.e., the classification model).

[0094] Clinical variables included patient clinical information for each enhanced CT image. This information included patient age, sex, tumor location (pancreatic head, body, tail), maximum tumor diameter, dilation of the main pancreatic duct, and dilation of the bile duct. Univariate and multivariate analyses were used to identify clinical variables with independent predictive value, which were then combined with the M-DLR score to construct the model.

[0095] Risk stratification threshold determination: The optimal classification threshold was determined through ROC curve analysis. The main threshold was determined using the Youden index (sensitivity + specificity - 1) maximization principle. Simultaneously, considering actual clinical needs, multiple thresholds were set: low-risk threshold (high sensitivity, avoiding missed diagnoses), high-risk threshold (high specificity, avoiding overtreatment), and medium-risk threshold (balancing sensitivity and specificity).

[0096] In this invention, the calculated radiomics score (M-DLR score) is integrated with clinical variables (age, gender, tumor location, size, etc.) to construct a radiomics-clinical joint model. Multivariate analysis is used to determine the independent predictive value of each variable, achieving complementary advantages between imaging and clinical information.

[0097] A three-tiered risk stratification system was established based on the M-DLR score: low-risk group (score <0.3, corresponding to a high probability of G1 grade tumors), intermediate-risk group (score 0.3-0.7, requiring further evaluation), and high-risk group (score ≥0.7, corresponding to a high probability of G2 / G3 grade tumors). Each risk group corresponds to a different clinical management strategy.

[0098] Dynamic threshold optimization mechanism: This invention establishes a dynamic threshold optimization mechanism based on clinical feedback, which can adjust the risk stratification threshold according to the patient characteristics and clinical preferences of different medical institutions, thereby achieving individualized risk assessment.

[0099] Existing radiomics scoring methods are mostly simple linear combinations or based on single algorithms, lacking the ability to organically integrate with clinical variables and dynamically optimize. The M-DLR scoring system of this invention constructs a more complete and practical classification model for classifying pancreatic neuroendocrine tumors through multi-dimensional feature fusion, clinical variable integration, and dynamic optimization.

[0100] In classification step S17, the trained logistic regression model is used to perform G-level classification on the enhanced CT image to be classified.

[0101] Understandably, the trained logistic regression model, as a classification model, can classify the enhanced CT images to be classified. In this embodiment, the classification model divides the enhanced CT images into two categories: G1 level and G2 / G3 level.

[0102] The enhanced CT images to be classified are input into the trained logistic regression model, which outputs the M-DLR score. Then, the patient is classified into G1 or G2 / G3 based on the M-DLR score.

[0103] In addition, in this embodiment, the performance of the classification model was fully validated and its clinical applicability was evaluated through various methods such as ROC curves, Kaplan-Meier survival analysis, and decision curve analysis.

[0104] In this invention, multiple peritumoral microenvironment features, tumor habitat features, local pathological features, global contextual features, and 3D radiomics features are extracted from each enhanced CT image, achieving extremely rich multi-dimensional feature extraction. Then, a hierarchical feature selection algorithm is used to select multiple target features from these extracted multi-dimensional features. Based on these target features, the radiomics score M-DLR for each enhanced CT image is calculated. Finally, a logistic regression model is trained (constructed) based on the radiomics score M-DLR of each enhanced CT image and the corresponding patient's clinical information to obtain a classification model. Thus, based on the output of the classification model, the G-grade classification of pancreatic neuroendocrine tumors can be achieved quickly and accurately, providing a reliable non-invasive assessment tool for clinical decision-making.

[0105] The present invention also provides a classification device 20 for pancreatic neuroendocrine tumors, such as... Figure 2 As shown, the sorting device 20 includes:

[0106] Segmentation unit 201 uses a pre-trained segmentation model to segment multiple enhanced CT images from multiple patients to extract the tumor mask in each enhanced CT image.

[0107] The generation unit 202, for each tumor mask, expands a spherical structural element with a first radius to generate a first peritumoral region, and expands a spherical structural element with a second radius to generate a second peritumoral region, wherein the first radius is smaller than the second radius;

[0108] Division unit 203, for each tumor mask, divides the tumor into multiple habitats based on each voxel within the tumor;

[0109] Extraction unit 204 extracts 3D radiomics features from each enhanced CT image, extracts multiple local pathological features and multiple global context features from each tumor mask, extracts multiple peritumoral microenvironment features from the first peritumoral region and the second peritumoral region, and extracts multiple tumor habitat features from the multiple habitats.

[0110] Selection unit 205 selects multiple target features from the multiple peritumoral microenvironment features, multiple tumor habitat features, multiple 3D radiomics features, multiple local pathological features, and multiple global context features for each enhanced CT image, and calculates the radiomics score for each enhanced CT image;

[0111] Training unit 206 inputs the radiomics scores of multiple enhanced CT images and the clinical variables of multiple enhanced CT images into the logistic regression model for training, and obtains the trained logistic regression model.

[0112] Classification unit 207 uses the trained logistic regression model to perform G-level classification on the enhanced CT image to be classified.

[0113] It is understandable that the segmentation unit 201, generation unit 202, partitioning unit 203, extraction unit 204, selection unit 205, training unit 206, and classification unit 207 can be... Figure 3 The processor 102 in the electronic device 100 has the functions of these modules or units to implement them.

[0114] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform [operations]. Figure 1 The method shown.

[0115] The present invention also provides a computer program product, including computer-executable instructions, which are executed by processor 102 to implement [the program]. Figure 1 The method shown.

[0116] Now for reference Figure 3 , Figure 3 An example electronic device 1400 according to an embodiment of the present invention is illustrated schematically. In one embodiment, the electronic device 1400 may include one or more processors 1404, a system control logic unit 1408 connected to at least one of the processors 1404, a system memory 1412 connected to the system control logic unit 1408, a non-volatile memory (NVM) 1416 connected to the system control logic unit 1408, and a network interface 1420 connected to the system control logic unit 1408.

[0117] In some embodiments, processor 1404 may include one or more single-core or multi-core processors. In some embodiments, processor 1404 may include any combination of general-purpose processors and special-purpose processors (e.g., graphics processors, application processors, baseband processors, etc.). In embodiments where electronic device 1400 employs an eNB (Evolved Node B) or RAN (Radio Access Network) controller, processor 1404 may be configured to perform various conforming embodiments, such as... Figure 1 The example shown.

[0118] In some embodiments, the system control logic unit 1408 may include any suitable interface controller to provide any suitable interface to at least one of the processors 1404 and / or any suitable device or component communicating with the system control logic unit 1408.

[0119] In some embodiments, the system control logic unit 1408 may include one or more memory controllers to provide an interface to the system memory 1412. The system memory 1412 may be used to load and store data and / or instructions. In some embodiments, the system memory 1412 of the electronic device 1400 may include any suitable volatile memory, such as suitable dynamic random access memory (DRAM).

[0120] The non-volatile memory 1416 may include one or more tangible, non-transitory computer-readable media for storing data and / or instructions. In some embodiments, the non-volatile memory 1416 may include any suitable non-volatile memory such as flash memory and / or any suitable non-volatile storage device, such as at least one of HDD (Hard Disk Drive), CD (Compact Disc) drive, and DVD (Digital Versatile Disc) drive.

[0121] The non-volatile memory 1416 may include a portion of the storage resources on the device on which the electronic device 1400 is installed, or it may be accessible by the electronic device, but is not necessarily part of the electronic device. For example, the non-volatile memory 1416 may be accessed over a network via network interface 1420.

[0122] Specifically, system memory 1412 and non-volatile memory 1416 may each include a temporary copy and a permanent copy of instruction 1424. Instruction 1424 may include, when executed by at least one of processors 1404, causing electronic device 1400 to perform, as Figure 1The instructions for the method shown. In some embodiments, the instructions 1424, hardware, firmware and / or software components thereof may additionally / alternatively be located in the system control logic unit 1408, the network interface 1420 and / or the processor 1404.

[0123] Network interface 1420 may include a transceiver for providing a radio interface to electronic device 1400, thereby enabling communication with any other suitable device (such as a front-end module, antenna, etc.) via one or more networks. In some embodiments, network interface 1420 may be integrated into other components of electronic device 1400. For example, network interface 1420 may be integrated into at least one of processor 1404, system memory 1412, non-volatile memory 1416, and firmware device (not shown) with instructions, which, when at least one of processor 1404 executes the instructions, enable electronic device 1400 to perform as follows: Figure 1 The method shown.

[0124] The network interface 1420 may further include any suitable hardware and / or firmware to provide a multiple-input multiple-output radio interface. For example, the network interface 1420 may be a network adapter, a wireless network adapter, a telephone modem, and / or a wireless modem.

[0125] In one embodiment, at least one of the processors 1404 may be packaged together with the logic of one or more controllers for the system control logic unit 1408 to form a system package (SiP). In another embodiment, at least one of the processors 1404 may be integrated on the same die with the logic of one or more controllers for the system control logic unit 1408 to form a system on chip (SoC).

[0126] The electronic device 1400 may further include an input / output (I / O) device 1432. The I / O device 1432 may include a user interface enabling a user to interact with the electronic device 1400; the peripheral component interface is designed to allow peripheral components to also interact with the electronic device 1400. In some embodiments, the electronic device 1400 may also include sensors for determining at least one type of environmental condition and location information related to the electronic device 1400.

[0127] In some embodiments, the user interface may include, but is not limited to, a display (e.g., a liquid crystal display, a touch screen display, etc.), a speaker, a microphone, one or more cameras (e.g., a still image camera and / or a video camera), a flashlight (e.g., a light-emitting diode flash), and a keyboard.

[0128] The various embodiments of the mechanisms disclosed in this application can be implemented in hardware, software, firmware, or a combination of these implementation methods. Embodiments of this application can be implemented as computer programs or program code executable on a programmable system, the programmable system including at least one processor, a storage system (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device.

[0129] Program code can be applied to input instructions to execute the functions described in this application and generate output information. The output information can be applied to one or more output devices in a known manner. For the purposes of this application, the processing system includes any system having a processor such as, for example, a digital signal processor (DSP), a microcontroller, an application-specific integrated circuit (ASIC), or a microprocessor.

[0130] The program code can be implemented using a high-level procedural language or an object-oriented programming language to communicate with the processing system. Assembly language or machine language can also be used when needed. In fact, the mechanisms described in this application are not limited to any particular programming language. In either case, the language can be a compiled language or an interpreted language.

[0131] In some cases, the disclosed embodiments may be implemented in hardware, firmware, software, or any combination thereof. The disclosed embodiments may also be implemented as instructions carried or stored thereon on one or more temporary or non-temporary machine-readable (e.g., computer-readable) storage media, which may be read and executed by one or more processors. For example, the instructions may be distributed via a network or through other computer-readable media. Therefore, machine-readable media may include any mechanism for storing or transmitting information in a machine-readable (e.g., computer-readable) form, including but not limited to floppy disks, optical disks, CD-ROMs, magneto-optical disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic cards or optical cards, flash memory, or tangible machine-readable storage for transmitting information (e.g., carrier waves, infrared signals, digital signals, etc.) using the Internet in the form of electrical, optical, acoustic, or other propagation signals. Therefore, machine-readable media include any type of machine-readable medium suitable for storing or transmitting electronic instructions or information in a machine-readable (e.g., computer-readable) form.

[0132] In the accompanying drawings, some structural or methodological features may be shown in a specific arrangement and / or order. However, it should be understood that such a specific arrangement and / or order may not be necessary. Rather, in some embodiments, these features may be arranged in a manner and / or order different from that shown in the illustrative drawings. Furthermore, the inclusion of structural or methodological features in a particular figure does not imply that such features are required in all embodiments, and in some embodiments, these features may be omitted or may be combined with other features.

[0133] It should be noted that all units / modules mentioned in the device embodiments of this application are logical units / modules. Physically, a logical unit / module can be a physical unit / module, a part of a physical unit / module, or a combination of multiple physical units / modules. The physical implementation of these logical units / modules themselves is not the most important factor; the combination of functions implemented by these logical units / modules is the key to solving the technical problems proposed in this application. Furthermore, to highlight the innovative aspects of this application, the above-described device embodiments of this application have not introduced units / modules that are not closely related to solving the technical problems proposed in this application. This does not mean that the above-described device embodiments do not contain other units / modules.

[0134] It should be noted that in the examples and description of this patent, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0135] Although this application has been illustrated and described with reference to certain preferred embodiments thereof, those skilled in the art should understand that various changes in form and detail may be made thereto without departing from the spirit and scope of this application.

Claims

1. A method for classifying a pancreatic neuroendocrine tumor for an electronic device, the method comprising: receiving a plurality of features of the pancreatic neuroendocrine tumor; and determining a class of the pancreatic neuroendocrine tumor based on the plurality of features. The classification method comprises: a segmentation step of segmenting each of a plurality of enhanced CT images from a plurality of patients using a pre-trained segmentation model to segment a tumor mask in each enhanced CT image; a generation step of, for each tumor mask, performing a spherical structure element dilation of a first radius to generate a first peritumoral region and performing a spherical structure element dilation of a second radius to generate a second peritumoral region, wherein the first radius is smaller than the second radius, the first peritumoral region reflects the invasion of the tumor, and the second peritumoral region reflects the influence of the tumor on the surrounding tissue; a division step of, for each tumor mask, dividing the tumor into a plurality of habitats based on each voxel within the tumor, wherein each habitat locates a tumor region of different function in three-dimensional space; an extraction step of extracting 3D radiomics features from each enhanced CT image, extracting a plurality of local pathological features and a plurality of global context features from each tumor mask, extracting a plurality of peritumoral microenvironment features from the first peritumoral region and the second peritumoral region, and extracting a plurality of tumor habitat features from the plurality of habitats; a selection step of, for each enhanced CT image, selecting a plurality of target features from the plurality of peritumoral microenvironment features, the plurality of tumor habitat features, the 3D radiomics features, the plurality of local pathological features, and the plurality of global context features, and calculating a radiomics score of each enhanced CT image; a training step of inputting the radiomics score of each of the plurality of enhanced CT images and the clinical variables of each of the plurality of enhanced CT images into a logistic regression model to train the logistic regression model to obtain a trained logistic regression model; a classification step of using the trained logistic regression model to perform G-grade classification on a to-be-classified enhanced CT image.

2. The classification method of claim 1, wherein, The plurality of peritumoral microenvironment features comprises peritumoral features of the first peritumoral region and the second peritumoral region, and associated features of the peritumoral features and intratumoral features, and the plurality of tumor habitat features comprises morphological features, texture features of each habitat, and spatial relationship features between the habitats.

3. The classification method of claim 1, wherein, In the division step, for each tumor mask, a local entropy value and a CT intensity value are calculated for each voxel within the tumor, and a K-means clustering algorithm is used to cluster analyze the entropy values and the CT intensity values, thereby dividing the tumor into the plurality of habitats.

4. The classification method of claim 1, wherein, The selection step comprises: for each enhanced CT image, performing statistical significance evaluation on the plurality of peritumoral microenvironment features and the plurality of tumor habitat features, and retaining features that are significantly related to tumor grading; calculating Spearman correlation coefficients between the retained features, identifying highly correlated feature pairs according to the Spearman correlation coefficients, and retaining features with greater information content in the feature pairs using a variance maximization principle; using a recursive feature elimination algorithm based on random forest, iteratively training and evaluating feature importance to gradually remove features with the smallest prediction contribution; using LASSO logistic regression for feature selection to select the plurality of target features.

5. The classification method of claim 1, wherein, The radiomics score of each enhanced CT image is calculated based on the following formula (1): M-DLR = 1 / (1 + exp(-(β0+∑βi×Xi)) Formula (1), i ))) Formula (1), wherein M-DLR represents the radiomics score, β0is an intercept term, βi represents a weight of the i-th target feature, X i represents a standardized feature value of the i-th target feature, wherein the radiomics score ranges from 0 to 1.

6. The classification method of claim 1, wherein, The clinical variables include clinical information of a patient corresponding to each enhanced CT image.

7. The classification method of claim 1, wherein, For each tumor mask, the ResNet18 network is used to extract the plurality of local pathological features from a plurality of continuous axial slices centered on the maximum cross-section of the tumor, and wherein the DINOv2 network is used to extract the plurality of global context features from the plurality of continuous axial slices.

8. A device for classification of pancreatic neuroendocrine tumors, characterized in that, The classification device comprises: A segmentation unit uses a pre-trained segmentation model to segment a plurality of enhanced CT images from a plurality of patients respectively to segment out a tumor mask in each enhanced CT image; A generation unit, for each tumor mask, performs a spherical structure element dilation of a first radius to generate a first peritumoral region, and performs a spherical structure element dilation of a second radius to generate a second peritumoral region, wherein the first radius is smaller than the second radius, the first peritumoral region reflects the invasion of the tumor, and the second peritumoral region reflects the influence of the tumor on the surrounding tissue; A division unit, for each tumor mask, divides the tumor into a plurality of habitats based on each voxel within the tumor, wherein each habitat locates a tumor region of different function in three-dimensional space; An extraction unit extracts 3D radiomics features from each enhanced CT image, extracts a plurality of local pathological features and a plurality of global context features from each tumor mask, extracts a plurality of peritumoral microenvironment features from the first peritumoral region and the second peritumoral region, and extracts a plurality of tumor habitat features from the plurality of habitats; A selection unit, for each enhanced CT image, selects a plurality of target features from the plurality of peritumoral microenvironment features, the plurality of tumor habitat features, the 3D radiomics features, the plurality of local pathological features, and the plurality of global context features, and calculates a radiomics score of each enhanced CT image; A training unit inputs the radiomics scores of the plurality of enhanced CT images and the clinical variables of the plurality of enhanced CT images into a logistic regression model for training to obtain a trained logistic regression model; A classification unit uses the trained logistic regression model to perform G-grade classification on the enhanced CT image to be classified.

9. A computer-readable storage medium, characterized in that, The storage medium has instructions stored thereon, which, when executed on a computer, cause the computer to perform the classification method for pancreatic neuroendocrine tumors according to any one of claims 1 to 7.

10. An electronic device, comprising: Comprise: One or more processors; One or more memories; the one or more memories store one or more programs, which, when executed by the one or more processors, cause the electronic device to perform the classification method for pancreatic neuroendocrine tumors according to any one of claims 1 to 7.

11. A computer program product comprising computer executable instructions, characterised in that, The instructions are executed by the processor to implement the classification method for pancreatic neuroendocrine tumors according to any one of claims 1 to 7.

Images