A method for developing a pdac prognosis prediction model

Abstract

The application relates to the technical field of intelligent medical systems, and provides a PDAC prognosis prediction model development and verification method, which is used for postoperative survival prediction and risk stratification of pancreatic ductal adenocarcinoma (PDAC) patients. The decision-level fusion model of deep learning and imageomics has the optimal prediction performance, and after further incorporation of body composition and key clinical pathological indexes, the model prediction accuracy is improved. The model interpretability is enhanced by means of gradient weighted class activation mapping (Grad-CAM) and SHAP analysis; through TCIA-CPTAC queue transcriptomics analysis, it is confirmed that the high-risk phenotype defined by the model is closely related to tumor stress adaptation, activation of inflammation-related signal pathways and imbalance of immune microenvironment. In conclusion, the application constructs a prognosis prediction model which has clinical practicability and biological mechanism support, and can provide a reference for individualized postoperative management of pancreatic ductal adenocarcinoma patients.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical system technology, and in particular to a method for developing a PDAC prognostic prediction model. Background Technology

[0002] Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive malignant tumor with a 5-year survival rate of less than 13%, placing a significant burden on patients and the healthcare system. Radical resection remains the only curative option, but postoperative prognosis exhibits significant heterogeneity, with some patients experiencing early recurrence or only achieving limited survival benefit. Therefore, postoperative survival risk is crucial for individualized risk stratification and optimized clinical management of PDAC.

[0003] Current prognostic assessments for PDAC remain inadequate. Although the American Joint Committee on Cancer (AJCC) 8th edition tumor-lymph node-metastasis (TNM) staging system is widely used in clinical practice, its ability to differentiate survival outcomes is limited. Traditional clinicopathological variables and serum biomarkers (including tumor size, lymph node status, and carbohydrate antigen (CA) 19-9 levels) have been shown to be associated with PDAC prognosis, but their clinical application value is limited by subjectivity and confounding factors such as obstructive jaundice. In recent years, genomics and molecular biomarkers have shown promise in prognostic assessment, but their clinical adoption is still constrained by the invasiveness of tissue sampling, the complexity of testing procedures, and high costs. Therefore, there is an urgent need for a non-invasive, efficient, objective, and biologically based prognostic assessment method to overcome the limitations of existing tools. Summary of the Invention

[0004] The purpose of this invention is to provide a method for developing a prognostic prediction model for pancreatic ductal adenocarcinoma (PDAC), which belongs to the category of health-related data systems and is used for postoperative survival prediction and risk stratification of patients with PDAC.

[0005] The above-mentioned technical objective of this invention is achieved through the following technical solution: a method for developing a PDAC prognostic prediction model, comprising the following steps,

[0006] Step 1: Dataset construction, including training queue, internal test queue, external test queue, and TCIA-CPTAC queue;

[0007] Step two, image processing, including image preprocessing, segmentation of intratumoral and peritumoral regions, and feature integration;

[0008] Step 3: Model development. Construct a radiomics survival model, a 2D deep learning survival model, and a 2.5D deep learning model. Construct two deep learning-radiomics fusion models: a feature-level fusion model and a decision-level fusion model. The scores generated by the best-performing model, as well as volume components and key clinicopathological factors, are incorporated into the RSF algorithm to construct a joint survival model.

[0009] Step four, model interpretability analysis, applies the Grad-CAM algorithm to a 2.5D deep learning model to visualize radiographic regions with prognostic relevance on CT images; the biological interpretability of the model is studied by performing correlation analysis between the risk scores predicted by the best-performing model and transcriptome data from the Clinical Proteomics Tumor Analysis Consortium database; and a Shap analysis is performed on the joint survival model to evaluate the contribution of each predictor at both global and individual levels.

[0010] Furthermore, in step two, the tumor in the portal venous phase CECT image is automatically segmented using the nnU-Net v2 framework, followed by independent manual review and optimization. Disagreements are resolved through consensus negotiation. Based on the final tumor flooding image, a 3mm peritumoral region is generated by automatically morphologically dilating the tumor boundary. The final region of interest covers both the intratumoral and peritumoral regions for subsequent radiomics and deep learning analysis.

[0011] Furthermore, in step three, the method for constructing the radiomics survival model is as follows: Using the PyRadiomics tool, radiomics features are extracted from each region of interest according to the guidelines of the Imaging Biomarker Standardization Initiative. All features are standardized by Z-score. To reduce feature redundancy, Spearman correlation analysis is performed. For features with a correlation coefficient ≥ 0.9, one feature is removed. Then, univariate Cox regression analysis is used to screen features that are significantly related to prognosis. The final radiomics feature set is determined by LASSO-Cox regression. Ten-fold cross-validation is used to select the optimal λ value. The screened features are incorporated into the Cox proportional hazards model to construct radiomics labels for survival prediction. Feature screening and model construction are both completed in the training queue.

[0012] Further, in step three, the construction method of the 2D deep learning survival model is as follows: Axial CT slices showing the largest cross-sectional area of ​​the tumor and including the peritumoral region are selected for each patient. A rectangular bounding box is used to crop the region of interest (ROI) that simultaneously covers both the tumor interior and the peritumoral region. All ROIs are first resampled to 256×256 pixels. During the training phase, data augmentation strategies such as random flipping and random cropping are used, and then the images are scaled to 224×224 pixels to adapt to the input size requirements of the pre-trained network. ResNet-50, ResNet-101, and ResNet-152 are evaluated as candidate backbone networks, and all networks are initialized using pre-trained weights based on the ImageNet dataset. ResNet is used as the convolutional backbone network to extract image features, which are then fed into the survival prediction head: the convolutional features are processed by a global average pooling layer and then input to a fully connected layer with L2 regularization. Finally, a single-neuron output layer generates an individualized risk score for each patient. To alleviate overfitting, a dropout method is introduced during training.

[0013] The model was trained using a stochastic gradient descent optimizer, with Nesterov momentum set to 0.9 and weight decay coefficient set to 1×10⁻⁶. -4 The system was configured with a cosine annealing learning rate scheduler with an initial learning rate of 0.001. Training was divided into two phases: the first five training epochs served as a warm-up phase, freezing the shallow layers of the backbone network; subsequently, all network parameters were unfrozen, and end-to-end optimization was performed, employing gradient pruning and early stopping strategies, with a batch size of 32. To adapt to right-censored survival data, optimization was performed based on the negative biased log-likelihood of the Cox proportional hazards model, defined as follows: ; For the i-th patient, the time of the event. This is an event indicator variable; it takes a value of 1 when the observed event occurs and a value of 0 when censored. The model predicts the logarithmic risk score for the i-th patient. Model training and parameter optimization are both completed in the training queue. The trained two-dimensional deep learning survival model generates an individualized risk score for each patient and extracts 2048-dimensional deep features from the global average pooling layer. Principal component analysis is used to reduce the dimensionality of these features to 256 dimensions. The resulting risk scores and the dimensionality-reduced deep features are used to construct the subsequent deep learning-radiomics fusion model.

[0014] Furthermore, in step three, the 2.5D deep learning survival model is constructed as follows: For each patient, the tumor-peritumor region of interest containing the largest tumor cross-section is extracted from the axial, sagittal, and coronal planes respectively; all regions of interest cover the tumor interior and peritumoral regions, are scaled to a fixed resolution, and then the regions of interest corresponding to the three planes are stacked along the channel dimension to construct a multi-channel 2.5D input tensor.

[0015] The ResNet series of networks is used as the backbone network, and ImageNet pre-trained weights are used for initialization to achieve transfer learning. To adapt to multi-channel 2.5-dimensional input, only the number of input channels of the first convolutional layer is modified, while the rest of the network structure remains consistent with the two-dimensional deep learning model. The convolutional features are processed by global average pooling and then input to a fully connected layer with L2 regularization. Finally, a single neuron output layer generates an individualized risk score. Dropout is introduced during training to alleviate overfitting. To adapt to right-censored survival data, the model is optimized using the Cox loss function.

[0016] Model training and parameter optimization were also completed in the training queue. After training, the model generated an individualized risk score for each patient and extracted 2048-dimensional deep features from the global average pooling layer. Principal component analysis was used to reduce the dimensionality of these features to 256 dimensions. The dimensionality-reduced deep features and the risk score output by the model were used to construct the subsequent fusion model.

[0017] Furthermore, in step three, the method for constructing the deep learning-radiomics fusion model is as follows:

[0018] The feature-level fusion model is constructed as follows: radiomics features are concatenated with 2D and 2.5D deep learning features to construct a joint feature matrix. Then, Z-score standardization is performed, and highly correlated features are removed by Spearman correlation analysis. When the absolute correlation coefficient between features is ≥0.9, only one feature is retained. Next, univariate Cox regression is used to screen features that are significantly related to overall survival. Then, LASSO-Cox regression is used to achieve feature dimensionality reduction and screening. Based on the final feature subset selected from the training queue, the RSF algorithm is used to build the feature-level fusion model.

[0019] The decision-level fusion model is constructed using a stacked fusion strategy. It takes the predicted risk scores output by the 2D deep learning model, the 2.5D deep learning model, and the radiomics model as inputs and uses the RSF algorithm to achieve integrated prediction.

[0020] The core hyperparameters of the RSF of both fusion models were optimized in the training queue through five-fold cross-validation. The final feature-level and decision-level fusion model was fitted based on the optimal hyperparameter combination and its performance was evaluated in the internal test queue, the external test queue, and the TCIA-CPTAC queue.

[0021] Furthermore, in step three, the joint survival model is constructed as follows: In the training queue, univariate Cox regression is first used to screen body components and clinicopathological variables that are significantly related to overall survival, and then multivariate Cox regression is used to determine independent predictors. The screened body component indicators, clinicopathological factors, and the risk score generated by the best-performing CT model are all incorporated into the RSF algorithm to construct the joint survival model; variable screening, model construction, and parameter optimization are all completed in the training queue, and the model performance is subsequently evaluated in internal and external testing queues.

[0022] Furthermore, in step four, Gradient Weighted Class Activation Mapping (Grad-CAM) is used to perform visual analysis on the 2.5D deep learning model to locate prognostic-related image regions on CT images.

[0023] Further, in step four, transcriptomics analysis was performed in the TCIA-CPTAC cohort. The optimal CT imaging model suitable for this cohort was selected to generate individualized risk scores, and patients were divided into high-risk and low-risk groups based on the threshold values ​​determined by the model. Using the R language's DESeq2 package, differentially expressed genes (DEGs) were screened based on the original transcriptome counting matrix, with a screening threshold of [insert threshold here]. The false detection rate (FDR) was <0.05. Functional enrichment analyses were performed using the R language's clusterProfiler package, including those based on the gene ontology GO and the Kyoto Encyclopedia of Genes and Genomes (KEGG). Based on the manually annotated gene set from the molecular feature database MSigDB, the sorted gene list underwent further gene set enrichment analysis (GSEA). The CIBERSORT algorithm was used to assess the differences in immune cell infiltration levels between high- and low-risk groups. SHAP analysis was performed on the final joint model to evaluate the contribution of each predictor at both the global and individual levels.

[0024] The beneficial effects of this invention are as follows: This invention provides a method for developing and validating a prognostic prediction model for pancreatic ductal adenocarcinoma (PDAC), used for postoperative survival prediction and risk stratification in patients with PDAC. The decision-level fusion model of deep learning and radiomics exhibits the best predictive performance; further incorporating body components and key clinicopathological indicators improves the model's predictive accuracy. Gradient-weighted class activation mapping (Grad-CAM) and SHAP analysis enhance the model's interpretability; transcriptomic analysis of the TCIA-CPTAC cohort confirms that the high-risk phenotype defined by this model is closely related to tumor stress adaptation, activation of inflammation-related signaling pathways, and imbalance of the immune microenvironment. In summary, this invention constructs a prognostic prediction model that combines clinical applicability with biological mechanism support, providing a reference for individualized postoperative management of patients with pancreatic ductal adenocarcinoma. Attached Figure Description

[0025] Figure 1 This is a flowchart of the prognostic prediction model construction and validation for pancreatic ductal adenocarcinoma (PDAC).

[0026] A. Dataset construction, including training queue, internal test queue, external test queue and TCIA-CPTAC queue.

[0027] B. Image processing, including image preprocessing, intratumoral and peritumoral region segmentation, and feature integration. C. Model construction, including radiomics models, 2D deep learning models, 2.5D deep learning models, feature-level fusion models, decision-level fusion models, and joint models. D. Model interpretability analysis, including model parsing and biological parsing.

[0028] Abbreviations: DL: Deep Learning; PDAC: Pancreatic Ductal Adenocarcinoma; TCIA: Cancer Imaging Archive; CPTAC: Clinical Proteomics Tumor Analysis Consortium.

[0029] Figure 2 Performance comparison of each prognostic model in four different queues.

[0030] A–D. C-index of each model in the four cohorts; error bars represent 95% confidence intervals. E–H. Time-dependent ROC curves of each model in the four cohorts. I–L. Calibration curves of all models in the four cohorts; the gray dashed line represents the ideal calibration state. M–P. Decision curve analysis of overall survival prediction in the four cohorts. Solid lines represent independent models, black lines represent the "all-intervention" strategy, and gray lines represent the "no-intervention" strategy.

[0031] Abbreviations: CI: Confidence Interval; ROC: Receiver Operating Characteristic (ROC).

[0032] Figure 3 It is a risk stratification and subgroup analysis of a 2.5D deep learning model.

[0033] A–D. Kaplan-Meier survival curves for the training cohort, internal test cohort, external test cohort, and TCIA-CPTAC cohort; E. Forest plot for subgroup analysis.

[0034] Figure 4 Distribution of risk scores for 2.5D deep learning models across four cohorts and model visualization.

[0035] AD. Distribution of risk scores, survival status, and risk stratification of 2.5D deep learning models in the four queues.

[0036] Figure EH is a violin plot comparing the 2.5D deep learning risk scores between the high-risk and low-risk groups in each cohort (***P<0.001).

[0037] I. Axial CT images of representative high-risk and low-risk patients from a 2.5D deep learning model and corresponding Grad-CAM visualizations, demonstrating radiographic patterns associated with risk status.

[0038] DL stands for Deep Learning; Grad-CAM stands for Gradient Weighted Class Activation Mapping.

[0039] Figure 5 Biological background of feature-level fusion model

[0040] A. GO functional enrichment analysis presented enrichment scores for three ontological aspects (biological processes, cellular components, and molecular functions), reflecting the core biological functions of genes associated with the DL_Rad_FB model. B. KEGG pathway enrichment analysis presented enriched signaling pathways and their enrichment scores and significance, elucidating the key signaling pathways involved in model-related genes. C. GSEA revealed the association patterns between different pathways and patient survival outcomes. D. Immune infiltration analysis showed the differences in the proportions of various immune cell infiltrations between high-risk and low-risk groups. GO: Gene Ontology; KEGG: Kyoto Encyclopedia of Genes and Genomes; GSEA: Gene Set Enrichment Analysis.

[0041] Figure 6 Shap analysis results of the joint model

[0042] AB: Global Interpretation: The bar chart shows the average absolute Shap value of each feature (A), reflecting its overall contribution to the model; the honeycomb plot shows the distribution of Shap values ​​of each feature (B), with colors indicating feature values ​​(red: high; blue: low).

[0043] CF: Local Interpretation: The waterfall plot shows the specific contribution of features to prediction in representative high-risk (C, E) and low-risk (D, F) individuals.

[0044] G: Population-level interpretation: This section attempts to illustrate the contribution distribution of each feature in the training cohort. The x-axis corresponds to the data for each patient, with red (high-risk) and blue (low-risk) areas representing the direction and intensity of feature contribution, respectively—the larger the red area, the higher the predicted risk of poor prognosis. This is based on the Shap method, also known as Shapley's additive interpretation. Detailed Implementation

[0045] Example: A method for developing a PDAC prognostic prediction model.

[0046] Materials and Methods

[0047] patient

[0048] This invention was approved by the ethics review committees of all participating institutions and was conducted in strict accordance with the Declaration of Helsinki (2013 revised edition). Because this is a retrospective study, written informed consent is not required. Patients and the public were not involved in the design, implementation, results reporting, or dissemination of the findings of this study.

[0049] This invention screened patients with primary pancreatic ductal adenocarcinoma (PDAC) who underwent radical surgical resection between June 2014 and October 2024 from four hospitals. Inclusion criteria: (1) pathologically confirmed pancreatic ductal adenocarcinoma; (2) follow-up duration and survival > 3 months; (3) completion of enhanced computed tomography (CECT) scan within 2 weeks before surgery. Exclusion criteria: (1) palliative or exploratory surgery; (2) neoadjuvant therapy before surgery; (3) other malignant tumors within the past 5 years; (4) incomplete clinicopathological data or prognostic information; (5) poor image quality.

[0050] The cohort included 545 patients from three hospitals affiliated with Huazhong University of Science and Technology: Wuhan Union Hospital (WHUH, n=353), Optics Valley Hospital (AVH, n=132), and Jinyinhu Hospital (JYH, n=60). This cohort was divided into a training cohort (70%) and an internal testing cohort (30%) at a 7:3 ratio. Guangdong Provincial People's Hospital (GPH, n=163) served as the external testing cohort. Furthermore, cases whose Cancer Imaging Archive (TCIA) imaging data matched with Clinical Proteomics Tumor Analysis Consortium (CPTAC) transcriptome data were screened; after excluding patients who only underwent plain CT scans, had poor image quality, or incomplete imaging data, 58 patients were ultimately included to form the TCIA-CPTAC cohort.

[0051] Data collection

[0052] This invention collects clinical and pathological variables, including age, sex, height, weight, body mass index (BMI), tumor location, tumor size, carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 12-5 (CA12-5), carcinoembryonic antigen (CEA), albumin, platelet-lymphocyte ratio (PLR), neutrophil-lymphocyte ratio (NLR), lymphocyte-monocyte ratio (LMR), and prognostic nutritional index (PNI) [serum albumin (g / L) + 5 × total lymphocyte count (10^6)]. 9 [ / L)], surgical margin status (R0 / R1), tumor differentiation degree, nerve invasion and vascular invasion (LVI).

[0053] Simultaneously, body composition indicators measured by CT were evaluated, including visceral fat index (VFI), subcutaneous fat index (SFI), skeletal muscle index (SMI), intramuscular fat index (IMFI), skeletal muscle density (SMD), visceral fat-subcutaneous fat index ratio (VSR), and visceral fat-skeletal muscle index ratio (VMR).

[0054] Overall survival (OS) was defined as the time from the first surgery to the patient's death or last follow-up. Patients were followed up every 3 months for the first 2 years after surgery, and then every 6 months thereafter, with the last follow-up date being September 30, 2025.

[0055] CT Image Acquisition and Region of Interest (ROI) Segmentation

[0056] All patients underwent enhanced computed tomography (CECT) scans within 2 weeks preoperatively. CT images from our hospital were retrieved from the Picture Archiving and Communication System (PACS), and publicly available CT data from the TCIA-CPTAC cohort were obtained from the Cancer Imaging Archive (TCIA) website (https: / / www.cancerimagingarchive.net / collection / cptac-pda / ). On portal venous phase enhanced CT images, the tumor was automatically segmented using the nnU-Net v2 framework. Subsequently, two abdominal radiologists independently reviewed and manually optimized the segmentation results using ITK-SNAP software (version 3.8.0), reaching a consensus in case of disagreements. Based on the final tumor mask, a 3mm wide peritumoral region was generated by automatically morphologically expanding the tumor boundary. The final region of interest (ROI) simultaneously covered both the tumor interior and the peritumoral region for subsequent radiomics and deep learning analysis.

[0057] Image group student survival model

[0058] Using PyRadiomics, radiomics features were extracted from regions of interest (ROIs) according to the **Imaging Biomarker Standardization Initiative (IBSI)** guidelines. Standardized preprocessing and extraction parameters were applied before feature extraction. A total of 1834 radiomics features were extracted in this study, and all features were subsequently Z-score standardized.

[0059] To reduce feature redundancy, Spearman correlation analysis was performed; for feature combinations with a correlation coefficient ≥ 0.9, one of the features was removed. Subsequently, univariate Cox regression analysis was used to screen features significantly associated with prognosis (P < 0.05). The final radiomics feature set was determined using LASSO-Cox regression, and the optimal λ value was selected using 10-fold cross-validation. The screened features were incorporated into a Cox proportional hazards model to construct radiomics labels for survival prediction. Both feature selection and model construction were completed in the training queue.

[0060] 2D deep learning survival model

[0061] For the two-dimensional deep learning (2D DL) survival prediction task, axial CT slices displaying the largest tumor cross-sectional area and including the peritumoral region were selected for each patient. Regions of interest (ROIs) covering both the tumor interior and peritumoral region were cropped using rectangular bounding boxes. All ROIs were first resampled to 256×256 pixels. During model training, data augmentation strategies such as random flipping and random cropping were employed. Subsequently, the images were scaled to 224×224 pixels to match the input requirements of the pre-trained network.

[0062] This study evaluated ResNet-50, ResNet-101, and ResNet-152 as candidate backbone networks, all initialized with ImageNet pre-trained weights. ResNet was used as the convolutional backbone network to extract image features, which were then connected to a survival prediction head: the convolutional features were processed by a global average pooling layer and then input to a fully connected layer with L2 regularization. Finally, a single-neuron output layer generated an individualized risk score for each patient. To reduce overfitting, dropout (dropout rate = 0.5) was used during training.

[0063] The model was trained using a stochastic gradient descent (SGD) optimizer, with Nesterov momentum set to 0.9 and weight decay coefficient set to 1×10⁻. 4 The model is equipped with a cosine annealing learning rate scheduler with an initial learning rate of 0.001. Training is divided into two phases: a 5-round warm-up phase to freeze the shallow parameters of the backbone network; after the warm-up, all network parameters are unfrozen for end-to-end optimization. The study employs gradient pruning (norm threshold = 1.0) and early stopping strategies, with a batch size of 32. To adapt to right-censored survival data, the model is optimized based on the negative biased log-likelihood of the Cox proportional hazards model (hereinafter referred to as the Cox loss function), which is defined as follows: ;

[0064] For the i-th patient, the time of the event. This is an event indicator variable; it takes a value of 1 when the observed event occurs and a value of 0 when censored. The logarithmic risk score for the i-th patient predicted by the model.

[0065] Model training and parameter optimization were both performed in the training queue. After training, the trained two-dimensional deep learning survival model generated an individualized risk score for each patient and extracted 2048-dimensional deep features from the global average pooling layer. Principal component analysis (PCA) was used to reduce the dimensionality of these features to 256 dimensions. The resulting risk scores and the dimensionality-reduced deep features were then used to construct the subsequent deep learning-radiomics fusion model.

[0066] 2.5D Deep Learning Survival Model

[0067] To incorporate three-dimensional spatial information while preserving the computational efficiency of two-dimensional convolutional neural networks, this study constructed a 2.5D deep learning model for overall survival (OS) prediction based on multi-view feature fusion. For each patient, tumor-peritumor regions of interest (ROIs) containing the largest tumor cross-section were extracted from axial, sagittal, and coronal views. All ROIs covered both the tumor interior and peritumoral regions, were scaled to a fixed resolution, and underwent the same preprocessing and data augmentation procedures as the two-dimensional model. Subsequently, the ROIs corresponding to the three cross-sections were stacked along the channel dimension to construct a multi-channel 2.5D input tensor.

[0068] This study uses ResNet series networks as the backbone and initializes the network with ImageNet pre-trained weights to achieve transfer learning. To adapt to multi-channel 2.5-dimensional input, only the number of input channels in the first convolutional layer is modified; the rest of the network structure remains consistent with the 2D deep learning model. Convolutional features are globally averaged and then fed into a fully connected layer with L2 regularization, before being output by a single neuron to generate individualized risk scores. Dropout (dropout rate = 0.5) is used during training to alleviate overfitting, and the training scheme is consistent with the 2D model. To adapt to right-censored survival data, the model is optimized using the Cox loss function.

[0069] Model training and parameter optimization were also performed in the training queue. After training, the model generated an individualized risk score for each patient and extracted 2048-dimensional deep features from the global average pooling layer. Principal component analysis (PCA) was used to reduce the dimensionality of the features to 256 dimensions. The dimensionality-reduced deep features and the risk scores output by the model were used to construct the subsequent fusion model.

[0070] Deep learning-radiomics fusion survival model

[0071] To improve the predictive efficacy of overall survival (OS) for patients with pancreatic ductal adenocarcinoma (PDAC), this study constructed two types of deep learning-radiomics fusion models: feature-level fusion model and decision-level fusion model.

[0072] The feature-level fusion model (named Rad_DL_FB) concatenates radiomics features with 2D and 2.5D deep learning features to construct a joint feature matrix, followed by Z-score standardization. Highly correlated features are removed using Spearman correlation analysis; when the absolute correlation coefficient between features is ≥0.9, only one feature is retained. Univariate Cox regression is then used to screen features significantly correlated with overall survival (P < 0.05), followed by LASSO-Cox regression to complete feature dimensionality reduction and selection. Based on the final feature subset obtained from the training queue, the **Random Survival Forest (RSF)** algorithm is used to construct the feature-level fusion model.

[0073] The decision-level fusion model (Rad_DL_DB) is constructed using a stacked fusion strategy. It takes the predicted risk scores output by the two-dimensional deep learning model, the 2.5-dimensional deep learning model, and the radiomics model as inputs and achieves integrated prediction through the random survival forest algorithm.

[0074] The core hyperparameters of the random survival forest for both models were optimized through five-fold cross-validation in the training queue. The final feature-level and decision-level fusion model was fitted based on the optimal hyperparameter combination and its performance was evaluated in the internal test queue, the external test queue, and the TCIA-CPTAC queue.

[0075] Joint Survival Model

[0076] In the training queue, univariate Cox regression was first used to screen body components and clinicopathological variables significantly associated with overall survival (OS) (P < 0.05), followed by multivariate Cox regression to determine independent predictors. The screened body component indicators, clinicopathological factors, and risk scores generated by the best-performing CT model were then incorporated into the Random Survival Forest (RSF) algorithm to construct a joint survival model. Variable selection, model construction, and parameter optimization were all performed in the training queue, and model performance was subsequently validated and evaluated in internal and external testing queues.

[0077] Model interpretability analysis and biological mechanism research

[0078] Gradient-weighted class activation mapping (Grad-CAM) was used to visualize the 2.5-dimensional deep learning model and locate key imaging regions on CT images that are related to prognosis.

[0079] To explore the biological mechanisms of CT-based model risk phenotypes, this study conducted transcriptomic analysis in the TCIA-CPTAC cohort. The optimal CT imaging model suitable for this cohort was selected to generate individualized risk scores, and patients were divided into high- and low-risk groups based on the threshold values ​​determined by the model. Using the R language package DESeq2, differentially expressed genes (DEGs) were screened based on the original transcriptome counting matrix, with a screening threshold of |log2 fold change (log2FC)| > 0.5 and a false discovery rate (FDR) < 0.05. Functional enrichment analyses, including Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG), were performed using the R language package clusterProfiler. Gene set enrichment analysis (GSEA) was further conducted on the sorted gene list based on the manually annotated gene set from the Molecular Feature Database (MSigDB). The CIBERSORT algorithm was used to assess the differences in immune cell infiltration levels between the high- and low-risk groups. SHAP analysis was performed on the final combined model to evaluate the contribution of each predictor factor at both the global and individual levels.

[0080] Statistical analysis

[0081] All statistical analyses were performed using R software (version 4.3.3, website: www.r-project.org) and Python (version 3.8.1). Continuous variables were expressed as median and interquartile range, and the Mann-Whitney U test was used for comparisons between groups. Categorical variables were expressed as frequency and percentage, and the chi-square test or Fisher's exact test was used for comparisons between groups. For subsequent modeling, the selected continuous variables were binary-classified. The `surv_cutpoint` function from the `survminer` package in R was used to determine the optimal cutpoint values ​​for various inflammatory indicators in the training queue, including age, body mass index (BMI), tumor size, carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 12-5 (CA12-5), carcinoembryonic antigen (CEA), serum albumin, and platelet-lymphocyte ratio (PLR), neutrophil-lymphocyte ratio (NLR), lymphocyte-monocyte ratio (LMR), and prognostic nutritional index (PNI). This function selects the threshold that maximizes the log-rank statistic in survival analysis. Given the significant sex differences in body composition, sex-specific cutpoint values ​​were used to classify body composition indicators into high- and low-risk groups.

[0082] Survival curves were plotted using the Kaplan-Meier method, and log-rank tests were used for inter-group comparisons. Independent prognostic factors related to overall survival (OS) were screened using univariate and multivariate Cox proportional hazards regression analysis. Model discrimination was assessed using the concordance index (C-index) and time-dependent receiver operating characteristic (ROC) curves; the consistency between the model's predicted survival probability and the actual outcome was assessed using calibration curves; and decision curve analysis was used to quantify the net clinical benefit of each model at different threshold probabilities. Differences in C-index among different models were compared using the compareC function. All statistical tests were two-tailed, and p < 0.05 was considered statistically significant.

[0083] result

[0084] Patient baseline characteristics

[0085] This study included 766 patients with pancreatic ductal adenocarcinoma (PDAC) from 4 medical centers and 1 public dataset, of whom 452 were male (59.01%) and 411 were over 60 years of age (53.66%). Patients were divided into a training cohort (n=381), an internal testing cohort (n=164), an external testing cohort (n=163), and a TCIA-CPTAC cohort (n=58). The median follow-up time for each group was 52.2, 44.5, 33.1, and 59.2 months, respectively, and the median overall survival (OS) was 24.1, 21.2, 25.6, and 25.7 months, respectively. The baseline clinicopathological characteristics of the patients are summarized in Table 1.

[0086] Feature extraction and model building

[0087] The overall research process is shown in Figure 1. When constructing the radiomics model, 35 key radiomics features were retained after LASSO feature selection (e.g., ...). Figure 4 Among all the evaluated deep learning networks, ResNet-152 demonstrated the best predictive performance in both 2D and 2.5D models, and was therefore selected as the final backbone network. The 2D and 2.5D deep learning models built on ResNet-152 generated individualized survival risk scores, which were used to construct the subsequent decision-level fusion model. When constructing the feature-level fusion model, 27 radiomics features, 6 two-dimensional deep learning features, and 10 2.5D deep learning features were ultimately incorporated.

[0088] Univariate and multivariate Cox regression analyses confirmed that high skeletal muscle density, high visceral fat-to-subcutaneous fat ratio (VSR), CA19-9 > 539.7 U / mL, CA12-5 > 13.9 U / mL, poor tumor differentiation, TNM stage ≥ IIB, vascular invasion, and adjuvant therapy were independent prognostic factors for overall survival in patients with pancreatic ductal adenocarcinoma (Table 2). Based on these findings, the Random Survival Forest (RSF) algorithm was used to integrate these independent prognostic factors with the optimal CT survival prediction model to construct a joint prediction model.

[0089] Performance comparison of various prognostic models

[0090] The consistency index (C-index) results for all models are shown below. Figure 2 AD and Table 3. In the four cohorts, the C-index of the radiomics model was 0.631–0.659, the 2D deep learning model was 0.674–0.709, and the 2.5D deep learning model was 0.695–0.735, indicating that the 2.5D deep learning model outperformed both the radiomics model and the 2D deep learning model in predictive performance. Among the two fusion strategies, the feature-level fusion model (DL_Rad_FB) had a C-index of 0.688–0.736, and the decision-level fusion model (DL_Rad_DB) had a C-index of 0.762–0.788, achieving the highest C-index in all cohorts. This invention subsequently integrated body components and clinicopathological factors independently related to overall survival based on the DL_Rad_DB model to construct a joint model, further improving its predictive performance in training, internal testing, and external testing cohorts. Detailed comparison results of all models can be found in [link to relevant documentation]. Figure 6 .

[0091] Figure 2 EP further summarized the predictive performance of each model for 2-year overall survival (OS). Time-dependent receiver operating characteristic (ROC) curves ( Figure 2 EH) shows that, across all four queues, the 2.5D deep learning model outperforms the image omics model and the 2D deep learning model, with the DL_Rad_DB model demonstrating stronger discriminative ability; the joint model further optimizes its prediction performance in the training, internal testing, and external testing queues.

[0092] Calibration curve ( Figure 2 The results (IL) show that the DL_Rad_DB model demonstrates good consistency between predicted and actual survival outcomes across all four cohorts; the joint model shows even better calibration performance in the training, internal testing, and external testing cohorts. Decision curve analysis ( Figure 2MP) showed that, across all four cohorts, the DL_Rad_DB model provided higher net clinical benefit compared to the "full intervention" and "no intervention" strategies, as well as the radiomics, 2D deep learning, 2.5D deep learning, and DL_Rad_FB models; the combined model further improved net benefit in the training, internal testing, and external testing cohorts.

[0093] Survival risk stratification of each prognostic model

[0094] Among single radiomics and deep learning models, the 2.5D deep learning model showed the best risk stratification performance. Using the median risk score of the training cohort as the cutoff value, Kaplan-Meier analysis showed that in all four cohorts, the overall survival (OS) of patients in the high-risk group was significantly shorter than that in the low-risk group (all log-rank P < 0.001). Figure 3 The median overall survival (OS) in the training cohort for high-risk and low-risk groups was 15.4 months and 37.8 months, respectively; in the internal testing cohort, it was 16.8 months and 35.3 months, respectively; in the external testing cohort, it was 17.8 months and 38.7 months, respectively; and in the TCIA-CPTAC cohort, it was 16.2 months and 34.7 months, respectively. Overall, the risk of death was significantly increased in the high-risk group (hazard ratio HR = 3.63; 95% confidence interval CI: 3.03–4.35; P < 0.001). Subgroup analysis further confirmed the robustness of the model in the pre-specified subgroups. Figure 3 E).

[0095] Among the two fusion models, the decision-level fusion model (DL_Rad_DB) demonstrated the strongest risk stratification capability and was the best-performing CT imaging model. Using the same threshold division method, the DL_Rad_DB model effectively distinguished differences in overall survival across all four cohorts (all log-rank P < 0.001). The median overall survival in the high-risk and low-risk groups of the training cohort was 14.6 months and 37.8 months, respectively; in the internal testing cohort, it was 14.5 months and 36.6 months, respectively; in the external testing cohort, it was 15.6 months and 39.2 months, respectively; and in the TCIA-CPTAC cohort, it was 14.2 months and 42.1 months, respectively. Overall, the high-risk group segmented by this model had the highest risk of death (HR = 5.21; 95% CI: 4.33–6.27; P < 0.001), indicating its superior discriminative ability. Subgroup analysis further validated the robustness of the model.

[0096] The joint model also achieved significant overall survival stratification in the training, internal testing, and external testing cohorts (all log-rank P < 0.001). The median overall survival in the high-risk and low-risk groups in the training cohort was 13.3 months and 40.2 months, respectively; in the internal testing cohort, it was 13.1 months and 36.6 months, respectively; and in the external testing cohort, it was 15.5 months and 39.2 months, respectively. The risk stratification effect of the model was further improved after integrating body composition and clinicopathological factors.

[0097] Visual Interpretation of 2.5D Deep Learning Models

[0098] Figure 4 AD illustrates the association between the 2.5D deep learning model risk score and patient survival status in training, internal testing, external testing, and the TCIA-CPTAC cohort. In the figure, the red and cyan curves represent the trend of risk score changes, the colored scatter dots represent patient survival outcomes (red for death, cyan for survival), and the dashed line represents the risk stratification threshold. The results show that there are significant differences in risk scores among patients with different survival outcomes in all cohorts; higher risk scores are associated with worse prognosis, and the two are significantly positively correlated.

[0099] Figure 4 EH showed significant differences in risk scores between high- and low-risk groups in each cohort, with high-risk patients generally having higher risk scores. To analyze the decision-making mechanism of the 2.5D model, this invention generates a gradient-weighted class activation mapping (Grad-CAM) heatmap (…). Figure 4 I). Heatmap results show that in high-risk patients, the model pays more attention to tumor heterogeneous regions and the overall lesion, with large areas of red / yellow high-activation regions visible; while in low-risk patients, the activation intensity of the tumor region is significantly reduced, mainly manifested as blue low-attention regions.

[0100] Biological basis for the decision-level fusion model (DL_Rad_DB) to identify risk phenotypes

[0101] Given that the decision-level fusion model (DL_Rad_DB) is the best-performing CT imaging model, this invention further analyzes the transcriptomic data of the TCIA-CPTAC cohort to explore the biological mechanisms behind the risk phenotypes defined by this model. Patients were divided into high-risk and low-risk groups using a threshold (194.87) determined by the model. RNA-seq analysis screened for significantly differentially expressed genes (DEGs); false discovery rate (FDR) < 0.05 and In the high-risk group, LUZP4, PPIAP73, MKRN6P, LCA10, and SLC16A1P1 genes were upregulated, while MYADML2, IGHD6-25, SPATA31C1, and RNU6-528P genes were downregulated, indicating significant transcriptomic differences between the two groups.

[0102] Gene Ontology (GO) enrichment analysis showed that the differentially expressed genes were mainly enriched in biological processes such as phagocytosis, B cell activation, complement activation, and telomere maintenance, as well as molecular functions such as immunoglobulin binding, antigen binding, and enzyme inhibitor activity. Figure 5 A) suggests that the high-risk phenotype may be related to immune dysregulation and remodeling of host defense-related processes. Further pathway analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) revealed that differentially expressed genes were enriched in key pathways such as ferroptosis, autophagy, and NOD-like receptor signaling. Figure 5 B), these pathways are closely related to tumor cell stress adaptation, activation of inflammation-related signals, and microenvironment interactions.

[0103] Gene set enrichment analysis (GSEA) showed that the high-risk group exhibited significant negative enrichment of gene sets related to solute carrier (SLC) transporter function, ribosome-related processes, and tumor progression. Figure 5 C), indicating a disruption in metabolic transport and protein synthesis processes. Further immune infiltration analysis showed significantly higher abundance of naive B cells and resting memory CD4⁺ T cells in the low-risk group, while the high-risk group was predominantly enriched with activated mast cells. Figure 5 (D) indicates that there are significant differences in the tumor immune microenvironment between the two groups.

[0104] In summary, the results of this invention suggest that the high-risk phenotypes defined by the DL_Rad_DB model may be associated with tumor cell stress adaptation, activation of inflammation-related signaling pathways, and imbalance of the immune microenvironment. These factors collectively promote tumor progression and ultimately lead to poor patient prognosis.

[0105] SHAP Interpretability Analysis of Joint Model

[0106] To further analyze the contribution of each integrated predictor in the final joint model, this study uses SHAP analysis to conduct interpretability analysis at the model level. At the global level, SHAP bar charts and scatter plots ( Figure 6 (AB) shows that the 2.5D deep learning risk score is the primary contributor to survival prediction, followed by the two-dimensional deep learning risk score and multiple clinicopathological indicators. At the individual level, the SHAP analysis plot ( Figure 6 CF revealed significant differences in the contribution patterns of various factors between typical high- and low-risk patients.

[0107] In one high-risk patient, the model-predicted hazard increased from a baseline of 183.146 to 328.44, primarily driven by the 2.5D and 2D model risk scores, while radiomics indicators and some clinicopathological factors had a mitigating effect. Conversely, in one low-risk patient, the model-predicted hazard decreased to 98.13, because the negative contributions of the 2.5D deep learning risk score, radiomics score, and TNM stage outweighed the positive contributions of other variables. SHAP force-directed plot of the overall population ( Figure 6 G) Further confirmation that there is significant heterogeneity in the risk contribution patterns of various factors among patients suggests that the individualized risk prediction of the joint model is jointly determined by deep learning features, radiomics features, body component indicators and clinicopathological features.

[0108] analyze

[0109] Pancreatic ductal adenocarcinoma (PDAC) exhibits high tumor heterogeneity and extremely poor clinical prognosis; therefore, accurate prognostic stratification is crucial for individualized treatment and improved survival. This study employed a feature-level and decision-level fusion strategy, integrating two-dimensional and 2.5D deep learning, radiomics, and clinicopathological indicators to construct a multimodal prognostic assessment system. This integrated model demonstrated excellent predictive performance in multiple independent cohorts, providing a novel approach to postoperative prognostic management for PDAC patients.

[0110] Previous studies have confirmed that the peritumoral region plays a crucial role in tumor invasion, metastasis, and immune regulation, and its prognostic information can complement intratumoral features. Therefore, both the radiomics and deep learning models in this study extract combined intratumoral and peritumoral region features to more comprehensively capture tumor spatial heterogeneity. At the radiomics level, the concordance index (C-index) of the intratumoral-peritumoral radiomics model constructed in this study reached 0.659, slightly higher than previous models built solely based on the intratumoral region (C-index=0.63), suggesting that peritumoral information can further enhance prognostic predictive value. However, radiomics features rely on manually designed descriptors, and their feature representation capabilities have inherent limitations.

[0111] Based on this, this study introduces a deep learning framework to further improve prediction performance. Among the three candidate backbone networks, ResNet-50, ResNet-101, and ResNet-152, ResNet-152 achieved the best performance in this study task, consistent with previous findings. The two-dimensional deep learning model built based on this backbone network achieved a C-index of 0.709, outperforming radiomics models. In contrast, the C-index of existing two-dimensional convolutional neural networks trained on single-layer axial CT images is only 0.651. These comparisons demonstrate that even with a two-dimensional network structure, optimizing the backbone network design and employing an end-to-end learning strategy can still effectively improve model prediction performance.

[0112] Given that single-layer CT images cannot fully capture the spatial heterogeneity of tumors, this study further constructs a 2.5D multi-planar fusion model, integrating axial, sagittal, and coronal image information to enhance spatial representation capabilities while minimizing computational complexity. This model achieves a C-index of 0.735, outperforming 2D deep learning models, radiomics models, and several previously reported 3D models. For example, existing multi-phase 3D multi-task models have a C-index of only 0.667. Previous reviews have indicated that while 3D models provide richer spatial information, they are prone to overfitting when sample size is limited and data anisotropy exists. In contrast, the 2.5D modeling strategy better balances computational efficiency and predictive performance. Under the same network architecture, the 2.5D model outperforms both 2D and 3D models in the lung nodule classification task. In conclusion, the 2.5D framework can effectively balance computational efficiency and the completeness of spatial information in volumetric medical image analysis.

[0113] In terms of multimodal fusion, this study systematically compares two strategies: feature-level fusion and decision-level fusion. Feature-level fusion achieves joint modeling by concatenating feature vectors, but it is more susceptible to feature redundancy and noise interference in high-dimensional feature spaces. Decision-level fusion, on the other hand, integrates the risk scores independently generated by each sub-model, reducing the model instability caused by direct coupling between different modalities. Our results show that the decision-level fusion model has a C-index of 0.788, which is superior to the feature-level fusion model (0.736), consistent with previous research on risk score-based integration strategies. This suggests that, under conditions of limited sample size, step-by-step modeling helps improve model stability.

[0114] The model performance was further improved by incorporating CA19-9, CA12-5, TNM stage, tumor differentiation grade, adjuvant therapy, and body composition indicators. Notably, body composition indicators such as skeletal muscle density and the visceral-to-subcutaneous fat ratio demonstrated independent prognostic value in the combined model, suggesting that patient metabolic status can provide supplementary information for survival assessment of pancreatic ductal adenocarcinoma. While previous studies have confirmed the correlation between body composition and prognosis in pancreatic ductal adenocarcinoma patients, studies systematically integrating these indicators into prognostic prediction models are still limited. This study incorporates body composition indicators into a comprehensive model that integrates radiomics, deep learning, and clinicopathological factors, expanding the application scope of multimodal prognostic modeling.

[0115] Compared to multi-omics models that rely on molecular sequencing, this study builds upon routine clinical imaging and laboratory data, ensuring predictive efficacy while exhibiting stronger clinical generalizability. Furthermore, this study enhances model interpretability through Gradient Weighted Class Activation Mapping (Grad-CAM) and SHAP analysis: Grad-CAM visualizes the imaging regions of interest to the deep learning model, while SHAP analyzes the decision-making logic of the machine learning model; SHAP quantifies the contribution of each variable to the individualized risk score, improving the transparency and clinical interpretability of the model output.

[0116] Meanwhile, multi-omics data from the TCIA-CPTAC cohort validated the biological implications of this multimodal model. The high-risk pancreatic ductal adenocarcinoma subgroup exhibited unique molecular characteristics, such as upregulation of the mRNA export adaptor protein LUZP4, suggesting dysregulation of RNA processing. Pathway analysis revealed simultaneous enrichment of autophagy and ferroptosis pathways in high-risk tumors. Autophagy can regulate ferroptosis through NCOA4-mediated ferritin phagocytosis, indicating a potential association between the high-risk phenotype and ferroptosis sensitivity in the context of adaptive autophagy. Significant differences were also observed in the tumor microenvironment: low-risk tumors showed higher abundances of naive B cells and resting memory CD4⁺ T cells, while high-risk tumors exhibited an immunosuppressive, innate immune-dominated phenotype. This microenvironment may be regulated by intratumoral microbes, thereby mediating chemotherapy resistance, which is consistent with the downregulation of solute carriers (SLCs) and metabolism-related genes in high-risk patients. Therefore, this model can not only perform prognostic stratification non-invasively, but also uncover potential therapeutic targets for pancreatic ductal adenocarcinoma (such as ferroptosis sensitivity and immune microenvironment characteristics), providing a basis for developing targeted therapy strategies for high-risk patients.

[0117] Table 1. Baseline clinical data of patients

[0118] Note: Unless otherwise stated, data represent the number of patients, and percentages are in parentheses; * represents the median and interquartile range.

[0119] BMI: Body Mass Index; CA19-9: Carbohydrate Antigen 19-9; CA12-5: Carbohydrate Antigen 12-5; CEA: Carcinoembryonic Antigen; LMR: Lymphocyte / Monocyte Ratio; NLR: Neutrophil / Lymphocyte Ratio; OS: Overall Survival; PLR: Platelet / Lymphocyte Ratio; PNI: Prognostic Nutritional Index; IMFI: Intramuscular Fat Index; SFI: Subcutaneous Fat Index; SMD: Skeletal Muscle Density; SMI: Skeletal Muscle Index; TNM: Tumor-Lymph Node-Metastasis Staging; VFI: Visceral Fat Index; VSR: Visceral Fat Area / Subcutaneous Fat Area Ratio; VMR: Visceral Fat Area / Skeletal Muscle Area Ratio.

[0120] Table 2. Univariate and multivariate Cox regression analyses of clinical characteristics

[0121] Note: BMI, Body Mass Index; CA19-9, Carbohydrate Antigen 19-9; CA12-5, Carbohydrate Antigen 12-5; CEA, Carcinoembryonic Antigen; LMR, Lymphocyte-Monocyte Ratio; NLR, Neutrophil-Lymphocyte Ratio; PLR, Platelet-Lymphocyte Ratio; PNI, Prognostic Nutritional Index; IMFI, Intramuscular Fat Index; SFI, Subcutaneous Fat Index; SMD, Skeletal Muscle Density; SMI, Skeletal Muscle Index; TNM, Tumor-Lymph Node-Metastasis Staging; VFI, Visceral Fat Index; VSR, Visceral Fat Area / Subcutaneous Fat Area Ratio; VMR, Visceral Fat Area / Skeleton Muscle Area Ratio.

[0122] Table 3. C-index of each prediction model in the training and testing queues.

[0123] Note: Data in parentheses represents the 95% confidence interval. CI stands for confidence interval.

Claims

1. A method for developing a PDAC prognostic prediction model, characterized in that: Includes the following steps, Step 1: Dataset construction, including training queue, internal test queue, external test queue, and TCIA-CPTAC queue; Step two, image processing, including image preprocessing, segmentation of intratumoral and peritumoral regions, and feature integration; Step 3: Model development. Construct a radiomics survival model, a 2D deep learning survival model, and a 2.5D deep learning model. Construct two deep learning-radiomics fusion models: a feature-level fusion model and a decision-level fusion model. The scores generated by the best-performing model, as well as volume components and key clinicopathological factors, are incorporated into the RSF algorithm to construct a joint survival model. Step four, model interpretability analysis, applies the Grad-CAM algorithm to a 2.5D deep learning model to visualize radiographic regions with prognostic relevance on CT images; the biological interpretability of the model is studied by performing correlation analysis between the risk scores predicted by the best-performing model and transcriptome data from the Clinical Proteomics Tumor Analysis Consortium database; and a Shap analysis is performed on the joint survival model to evaluate the contribution of each predictor at both global and individual levels.

2. The method for developing a PDAC prognostic prediction model according to claim 1, characterized in that: In step two, the tumor in the portal venous phase CECT image is automatically segmented using the nnU-Net v2 framework, followed by independent manual review and optimization. Disagreements are resolved through consensus negotiation. Based on the final tumor flooding image, the tumor boundary is automatically morphologically expanded to generate a 3mm peritumoral region. The final region of interest covers both the intratumoral and peritumoral regions for subsequent radiomics and deep learning analysis.

3. The method for developing a PDAC prognostic prediction model according to claim 2, characterized in that: In step three, the radiomics survival model is constructed as follows: Using PyRadiomics, radiomics features are extracted from each region of interest according to the guidelines of the Imaging Biomarker Standardization Initiative. All features are Z-score standardized. To reduce feature redundancy, Spearman correlation analysis is performed. For features with a correlation coefficient ≥ 0.9, one feature is removed. Then, univariate Cox regression analysis is used to screen features that are significantly related to prognosis. The final radiomics feature set is determined by LASSO-Cox regression. Ten-fold cross-validation is used to select the optimal λ value. The screened features are incorporated into the Cox proportional hazards model to construct radiomics labels for survival prediction. Feature screening and model construction are both completed in the training queue.

4. The method for developing a PDAC prognostic prediction model according to claim 3, characterized in that: In step three, the construction method of the 2D deep learning survival model is as follows: axial CT slices showing the largest cross-sectional area of ​​the tumor and including the peritumoral region are selected for each patient, and regions of interest covering both the tumor interior and the peritumoral region are cropped using rectangular bounding boxes; all regions of interest are first resampled to 256×256 pixels. During the training phase, data augmentation strategies such as random flipping and random cropping are employed. Subsequently, the images are scaled to 224×224 pixels to fit the input size requirements of the pre-trained network. ResNet-50, ResNet-101, and ResNet-152 were evaluated as candidate backbone networks, and all networks were initialized using pre-trained weights based on the ImageNet dataset. ResNet was used as the convolutional backbone network to extract image features, which were then fed into the survival prediction head: the convolutional features were processed by a global average pooling layer and then input into a fully connected layer with L2 regularization. Finally, a single-neuron output layer was used to generate an individualized risk score for each patient. To alleviate overfitting, a dropout method was introduced during training. The model was trained using a stochastic gradient descent optimizer, with Nesterov momentum set to 0.9 and weight decay coefficient set to 1×10⁻⁶. -4 It is paired with a cosine annealing learning rate scheduler with an initial learning rate of 0.001; Training is divided into two phases: the first five training epochs are a warm-up phase, freezing the shallow layers of the backbone network; then all network parameters are unfrozen, and end-to-end optimization is carried out, employing gradient pruning and early stopping strategies, with a batch size of 32; to adapt to right-censored survival data, optimization is performed based on the negative biased log-likelihood of the Cox proportional hazards model, defined as follows: ; For the i-th patient, the time of the event. This is an event indicator variable; it takes a value of 1 when the observed event occurs and a value of 0 when censored. The logarithmic risk score predicted by the model for the i-th patient; Model training and parameter optimization were both completed in the training queue. The trained two-dimensional deep learning survival model generated an individualized risk score for each patient and extracted 2048-dimensional deep features from the global average pooling layer. Principal component analysis was used to reduce the dimensionality of these features to 256 dimensions. The resulting risk scores and the dimensionality-reduced deep features were used to construct the subsequent deep learning-radiomics fusion model.

5. The method for developing a PDAC prognostic prediction model according to claim 4, characterized in that: In step three, the 2.5D deep learning survival model is constructed as follows: For each patient, the tumor-peritumor region of interest containing the largest tumor cross-section is extracted from the axial, sagittal, and coronal planes respectively; all regions of interest cover the tumor interior and peritumoral regions, are scaled to a fixed resolution, and then the regions of interest corresponding to the three planes are stacked along the channel dimension to construct a multi-channel 2.5D input tensor. The ResNet series of networks is used as the backbone network. ImageNet pre-trained weights are used for initialization to achieve transfer learning. To adapt to multi-channel 2.5-dimensional input, only the number of input channels of the first convolutional layer is modified, and the rest of the network structure is consistent with the two-dimensional deep learning model. After global average pooling, the convolutional features are input to a fully connected layer with L2 regularization, and finally, individualized risk scores are generated through a single neuron output layer. During training, a dropout method is introduced to alleviate overfitting. To adapt to right-censored survival data, the model is optimized using the Cox loss function. Model training and parameter optimization were also completed in the training queue. After training, the model generated an individualized risk score for each patient and extracted 2048-dimensional deep features from the global average pooling layer. Principal component analysis was used to reduce the dimensionality of these features to 256 dimensions. The dimensionality-reduced deep features and the risk score output by the model were used to construct the subsequent fusion model.

6. The method for developing a PDAC prognostic prediction model according to claim 5, characterized in that: In step three, the method for constructing the deep learning-radiomics fusion model is as follows: The feature-level fusion model is constructed as follows: radiomics features are concatenated with 2D and 2.5D deep learning features to construct a joint feature matrix. Then, Z-score standardization is performed, and highly correlated features are removed by Spearman correlation analysis. When the absolute correlation coefficient between features is ≥0.9, only one feature is retained. Next, univariate Cox regression is used to screen features that are significantly related to overall survival. Then, LASSO-Cox regression is used to achieve feature dimensionality reduction and screening. Based on the final feature subset selected from the training queue, the RSF algorithm is used to build the feature-level fusion model. The decision-level fusion model is constructed using a stacked fusion strategy. It takes the predicted risk scores output by the 2D deep learning model, the 2.5D deep learning model, and the radiomics model as inputs and uses the RSF algorithm to achieve integrated prediction. The core hyperparameters of the RSF of both fusion models were optimized in the training queue through five-fold cross-validation. The final feature-level and decision-level fusion model was fitted based on the optimal hyperparameter combination and its performance was evaluated in the internal test queue, the external test queue, and the TCIA-CPTAC queue.

7. The method for developing a PDAC prognostic prediction model according to claim 6, characterized in that: In step three, the joint survival model is constructed as follows: In the training queue, univariate Cox regression is first used to screen body components and clinicopathological variables that are significantly associated with overall survival, and then multivariate Cox regression is used to determine independent predictors. The screened body component indicators, clinicopathological factors, and the risk score generated by the best-performing CT model are all incorporated into the RSF algorithm to construct the joint survival model. Variable screening, model construction, and parameter optimization are all completed in the training queue, and the model performance is subsequently evaluated in internal and external testing queues.

8. The method for developing a PDAC prognostic prediction model according to claim 1, characterized in that: In step four, Gradient Weighted Class Activation Mapping (Grad-CAM) is used to visualize and analyze the 2.5D deep learning model, locating prognostic-related image regions on CT images.

9. The method for developing a PDAC prognostic prediction model according to claim 8, characterized in that: In step four, transcriptomics analysis was performed in the TCIA-CPTAC cohort. The optimal CT imaging model suitable for this cohort was selected to generate individualized risk scores, and patients were divided into high-risk and low-risk groups based on the threshold values ​​determined by the model. The DESeq2 package in R was used to screen differentially expressed genes (DEGs) based on the original transcriptome counting matrix, with a screening threshold of [insert threshold value here]. The false detection rate (FDR) was <0.

05. Functional enrichment analyses were performed using the R language's clusterProfiler package, including those based on the gene ontology GO and the Kyoto Encyclopedia of Genes and Genomes (KEGG). Based on the manually annotated gene set from the molecular feature database MSigDB, the sorted gene list underwent further gene set enrichment analysis (GSEA). The CIBERSORT algorithm was used to assess the differences in immune cell infiltration levels between high- and low-risk groups. SHAP analysis was performed on the final joint model to evaluate the contribution of each predictor at both the global and individual levels.

Images