Pancreatic cancer is the 12th most common cancer worldwide, but it is among the deadliest. In 2020, there were 495,773 new cases globally and 466,003 deaths, a near 1:1 ratio of incidence to mortality that underscores just how lethal this disease is. The five-year survival rate sits below 10%, and approximately 80% of patients are diagnosed with locally advanced or distant metastatic disease, leaving very few candidates for curative surgery.
Why detection comes too late: The pancreas is a deep-seated retroperitoneal organ surrounded by complex vascular structures, making it difficult to image and biopsy. Patients rarely develop distinctive symptoms in the early stages. There are no widely adopted specific molecular markers for screening, and the pancreas's highly vascularized environment facilitates rapid metastasis. Even when imaging is performed using CT, MRI, 18FDG PET/CT, or endoscopic ultrasound (EUS), international guidelines for image-based stratification remain heterogeneous and lack consensus.
Histopathology challenges: Biopsies are frequently required to confirm the diagnosis, but histopathological analysis is complicated by significant morphological heterogeneity within tumors and the small volume of tissue that can typically be collected. Evaluating treatment response on histopathology slides also lacks precision, adding another layer of difficulty to clinical decision making.
This review, authored by researchers at Massachusetts General Hospital, Harvard Medical School, and Boston University, surveys the landscape of artificial intelligence applications across the entire pancreatic cancer care continuum, from early detection and diagnosis through treatment planning, biomarker discovery, and workflow optimization.
The review categorizes AI techniques applied to pancreatic cancer into three tiers: traditional machine learning, deep learning, and transfer learning. At the traditional ML level, support vector machines (SVMs) have been employed to classify pancreatic tumors from features extracted from CT or MRI data. SVMs learn decision boundaries in high-dimensional feature spaces, making them well suited for distinguishing cancerous tissue from normal pancreatic parenchyma when trained on labeled imaging datasets.
Random forests (RFs) are ensemble methods that combine many decision trees to handle high-dimensional clinical, molecular, and demographic feature sets. In pancreatic cancer research, RF models have been used to predict patient outcomes such as survival rates and treatment response. Their ability to incorporate heterogeneous data types, from genomic markers to clinical staging information, makes them particularly useful for prognostic modeling.
Unsupervised learning: Clustering algorithms like k-means and hierarchical clustering are used to identify patient subgroups based on clinical or molecular characteristics without predefined labels. Dimensionality reduction techniques such as principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) help researchers visualize high-dimensional datasets and discover hidden relationships. For instance, unsupervised hierarchical clustering of TCGA pancreatic adenocarcinoma data revealed mucin expression profiles linked to overall survival differences.
ML models have also been trained on genomic, proteomic, and transcriptomic data to predict treatment response and drug efficacy. By identifying molecular signatures associated with resistance or sensitivity, these models can help guide personalized therapy decisions and potentially spare patients from ineffective treatments.
Convolutional neural networks (CNNs) have been extensively applied to image analysis tasks in pancreatic cancer, including tumor detection, segmentation, and classification from CT and MRI scans. CNNs automatically learn hierarchical features from raw pixel data, eliminating the need for manual feature engineering. Recurrent neural networks (RNNs) have been applied to time-series data such as longitudinal patient monitoring and treatment response tracking, capturing temporal dependencies that static models cannot.
Natural language processing (NLP): NLP techniques enable the extraction of structured information from unstructured text sources, including electronic health records (EHRs), clinical reports, and biomedical literature. Named entity recognition (NER) identifies specific medical entities such as tumor types, biomarkers, genes, and drug names from free-text clinical documents. Text mining of the biomedical literature has revealed latent associations among pancreatic cancer risk factors, genes, and molecular pathways that might otherwise go unnoticed.
EHR mining: NLP applied to electronic health records enables large-scale retrospective studies by automatically extracting patient demographics, medical histories, treatment details, and clinical outcomes. This facilitates efficient data collection for building structured databases and conducting population-level analyses of treatment efficacy and prognostic factors. The automation of data extraction and analysis through NLP leads to more informed decision making and improved patient care.
Computer vision: Beyond simple classification, computer vision techniques in pancreatic cancer research encompass image segmentation (delineating tumor boundaries), lesion detection and localization, tumor characterization (extracting texture, shape, and intensity features), image registration and fusion across modalities, and quantitative morphological analysis. These capabilities provide clinicians with precise measurements of tumor dimensions, morphology, and growth patterns for diagnosis, treatment planning, and disease monitoring.
Pancreatic cancer datasets are notoriously small compared to those available for more common cancers. Transfer learning addresses this constraint by leveraging knowledge from pre-trained models, typically CNNs trained on large general medical image datasets or datasets from other cancer types, and fine-tuning them on pancreatic cancer-specific data. The pre-trained layers retain universal visual features (edges, textures, structural patterns) that transfer well to the pancreatic cancer domain, allowing the model to converge faster and achieve higher accuracy with limited training samples.
Knowledge transfer across layers: Rather than retraining entire networks, researchers can freeze the lower-level layers of a pre-trained model (which capture generic features) and add custom layers tailored to the pancreatic cancer task. This layered approach retains the computational investment of the original training while adapting to the specific morphological and imaging characteristics of pancreatic tumors.
Domain adaptation: A related strategy involves training a model on data from analogous cancer types or other biomedical domains, then refining it with the available pancreatic cancer data. This allows the model to capture cancer-related features that possess transferability across tumor types. For example, a model trained on liver or colorectal CT data could learn general abdominal anatomy and tumor morphology patterns before being fine-tuned on pancreatic scans.
The authors emphasize that transfer learning is particularly important for pancreatic cancer because the disease is relatively rare compared to breast or lung cancer, and large, well-annotated pancreatic cancer datasets remain scarce. By building on pre-existing model weights rather than starting from scratch, researchers can develop competitive models without requiring tens of thousands of labeled pancreatic cancer images.
Early detection: AI has the potential to identify pancreatic cancer before clinical symptoms appear by integrating diverse patient data, including age, family history, lifestyle factors, medical history, laboratory results, and diagnostic reports. Deep learning models trained on large datasets can recognize characteristic imaging features of early-stage tumors, while segmentation frameworks can delineate the pancreas itself on imaging to flag morphological changes. AI can also mine electronic health records to detect subtle patterns and abnormalities that clinicians might miss.
Diagnosis and classification: A CNN-based deep learning model was developed to automatically segment pancreatic tumors from CT images and classify them as resectable or unresectable with high accuracy. An RNN-based model was proposed to analyze EUS images and differentiate benign from malignant pancreatic cysts with high sensitivity and specificity. On the genomic side, an SVM model classified pancreatic cancer into four molecular subtypes (basal-like, classical, quasi-mesenchymal, and exocrine-like) based on gene expression data. A random forest model predicted survival from DNA methylation data. A multimodal autoencoder (MAE) fused CT, PET, and gene expression data to classify tumors as resectable or unresectable.
Treatment planning and monitoring: AI integrates patient-specific data, including clinical history, genomic profiles, and imaging results, to help oncologists formulate customized treatment plans spanning surgery, chemotherapy, radiation, and immunotherapy. AI-enabled predictive modeling accounts for genetic markers, tumor characteristics, and patient-related factors to forecast treatment response. Real-time monitoring through biomarker levels and imaging data allows early identification of treatment failure, enabling timely modifications to therapy regimens.
Biomarker discovery: A multimodal neural network (MNN) was proposed to combine whole-slide images (WSI), gene expression data, clinical data (age, gender, tumor location), and biomarker data (miRNA) to forecast survival. AI methods applied to genomics, proteomics, metabolomics, and microbiomics can identify novel biomarker signatures with higher sensitivity and specificity than single biomarkers alone, overcoming the heterogeneity and limited sample challenges that plague traditional biomarker research.
An emerging application highlighted in this review is AI-based analysis of contrast-enhanced ultrasound (CEUS) images for pancreatic cancer characterization. A 2022 study developed a deep learning radiomics model trained on CEUS images from 558 patients to aid radiologists in diagnosing pancreatic ductal adenocarcinoma. The model achieved an AUC above 0.95 across training, internal validation, and two external validation cohorts, a notably strong performance that held up across independent datasets.
Predicting chemotherapy response: In a 2023 study of 38 patients, investigators applied deep learning to contrast-enhanced ultrasound videos to predict the efficacy of neoadjuvant chemotherapy for pancreatic cancer, achieving AUCs above 0.89 in two separate CNN models. While the sample size was small, the results suggest that CEUS-based AI could help identify patients who would benefit from preoperative chemotherapy before subjecting them to potentially ineffective treatment cycles.
Tumor aggressiveness prediction: A 2022 study of 104 patients developed a nomogram combining clinical factors (tumor size, arterial enhancement level, deep learning predictive probability) with deep learning CEUS analysis to predict the aggressiveness of pancreatic neuroendocrine neoplasms preoperatively. The combined nomogram significantly outperformed the clinical-only model (AUC 0.97 vs. 0.87, p = 0.009). A 2021 study demonstrated automatic segmentation of pancreatic tumors on contrast-enhanced endoscopic ultrasound with a decent concordance rate.
Prospective validation: A 2023 prospective trial demonstrated a deep learning system that diagnosed pancreatic masses significantly better than endoscopists and also improved first-pass diagnostic yield when used to guide fine needle aspiration. This prospective design is particularly noteworthy because most AI studies in pancreatic cancer remain retrospective.
Limited treatment options: Only 15-20% of pancreatic cancer patients are candidates for surgical resection, and nearly 75% of those who undergo surgery develop recurrence within 2 years, suggesting widespread micro-metastatic disease. First-line chemotherapy regimens (FOLFIRINOX and nab-paclitaxel) are the standard, but there are no standardized algorithms for second-line treatment decisions. Two recognized tumor progression pathways exist: Pancreatic Intraepithelial Neoplasia (PanINs), which are microscopic and undetectable on imaging, and Intraductal Papillary Mucinous Neoplasms (IPMNs), which are imageable. This biological complexity adds a fundamental layer of difficulty for AI models trying to predict disease course.
Data quality and availability: The lack of large, centralized, well-annotated datasets is a major barrier. Currently, the NIH-NCI-sponsored EDRN project is the only major centralized data effort for pancreatic cancer. Studies using smaller datasets have often not accounted for suboptimal image quality, post-treatment status artifacts, or the presence of biliary stents, all of which can introduce errors and biases into model training.
Black-box opacity: AI algorithms are frequently described as "black boxes" because the code and complexity behind them make it difficult to reproduce results independently. General descriptions of model architectures do not provide sufficient information to replicate findings. Without interpretability, clinicians cannot critically interrogate model outputs, which has increased resistance to clinical adoption. Efforts to address this include explainable AI techniques such as feature importance scores, visualizations, and natural language explanations of model reasoning, as well as alignment of AI outputs with existing clinical guidelines.
Algorithmic bias and ethics: Datasets used to train pancreatic cancer AI models tend to underrepresent women and minorities, producing biased models that may not generalize to the diverse patient populations seen in clinical practice. Deploying such skewed algorithms could widen, rather than narrow, health outcome disparities. Questions of data ownership and patient privacy are especially acute when large imaging datasets are assembled for model training. The authors call for improving diversity in training data, validating models across various populations, and establishing ethical guidelines and regulatory frameworks that prioritize fairness and transparency.
Multi-omics data integration: Pancreatic cancer is molecularly heterogeneous, and single-modality analyses capture only part of the picture. The authors argue that future AI algorithms must integrate genomic, transcriptomic, proteomic, and metabolomic data simultaneously to uncover robust molecular signatures, biomarkers, and therapeutic targets. This requires not just algorithmic innovation but also standardized data collection protocols and interoperable data platforms.
Improved imaging analysis: Future work should focus on developing and validating AI models specifically designed for pancreatic cancer imaging, enabling more precise disease detection, staging, and longitudinal monitoring. A recent example of what is possible is a radiomics-based machine learning model that detected pancreatic ductal adenocarcinoma at the prediagnostic stage with substantial lead time, suggesting that AI-powered imaging could shift detection much earlier in the disease course.
Predictive modeling and risk stratification: AI systems that combine clinical, genetic, and imaging data to identify patients at high risk of disease progression, recurrence, or therapy resistance could transform treatment planning. Real-time decision support systems that incorporate patient-specific data, clinical guidelines, and current research evidence would give oncologists actionable recommendations at the point of care. The authors stress that these systems must be user-friendly, interpretable, and scalable for integration into existing clinical workflows.
Data sharing and clinical validation: The effectiveness of AI in pancreatic cancer depends on access to high-quality, diversified, and well-annotated datasets. The authors call for standardized data-gathering techniques across institutions, ethical data-sharing frameworks, and secure platforms for data integration. Most critically, large-scale prospective validation studies are needed to establish the clinical efficacy, safety, and cost-effectiveness of AI tools before they can be responsibly deployed. Regulatory and ethical factors, including privacy protection, informed consent, and algorithm transparency, must be addressed in parallel.