Pancreatic cancer is among the most lethal cancers of the digestive system, often called the "king of cancer" because of its aggressiveness, rapid metastasis, and dismal survival rates. The incidence has surged globally in recent decades, driven by ageing populations, smoking, obesity, chronic pancreatitis, diabetes, and hereditary factors. Surgery remains the primary treatment, but the absence of specific early clinical symptoms and reliable molecular markers means the disease is usually detected at advanced stages when surgical intervention is no longer effective.
Anatomical challenges: The pancreas is a deep-seated retroperitoneal organ surrounded by complex vascular structures. This highly vascularized environment facilitates rapid metastasis and makes imaging interpretation difficult. A study found that nearly 90% of pancreatic cancer misdiagnoses were due to the inability to identify vascular invasion and the difficulty of spotting tumour masses obscured by inflammation. Common symptoms, including abdominal pain, weight loss, jaundice, and changes in stool consistency, typically become prominent only at advanced stages.
Limitations of existing biomarkers: Serological markers like CA 19-9 (carbohydrate antigen) lack specificity for early-stage disease and indicate only advanced cancer, increasing mortality risk. Several imaging modalities, including multidetector CT (MDCT), MRI, endoscopic ultrasound (EUS), and PET, each have distinct merits and drawbacks. MDCT offers high sensitivity for vascular invasion but carries radiation risk and sometimes lacks attenuation contrast between cancerous tissue and pancreatic parenchyma. MRI avoids ionizing radiation and provides superior soft-tissue contrast but is expensive and limited in availability. EUS can detect lesions as small as 2-3 mm with the highest diagnostic accuracy but requires trained operators and is not widely available.
This review examines how AI models can improve pancreatic cancer diagnosis across different imaging modalities, along with emerging AI-driven approaches based on cytopathology and serological markers. The authors also address ethical concerns surrounding the clinical deployment of these tools.
The review traces the evolution of AI in medical imaging from basic "if-then" rule sets to the modern landscape of machine learning (ML) and deep learning (DL). ML uses computational algorithms to identify patterns in large datasets. Supervised ML trains on labelled data (e.g., CT scans already classified as cancerous or non-cancerous) and can extract subtle features that human clinicians might miss due to oversight or fatigue. Unsupervised ML identifies patterns in unlabelled data but generally produces slightly lower prediction accuracy. Reinforcement learning, which uses trial-and-error interaction with the environment, has not yet been applied to pancreatic cancer diagnosis but could eventually support remote clinical decision-making.
Key ML algorithms: The review catalogues several ML techniques applied to pancreatic cancer, including K-nearest neighbour (k-NN), artificial neural networks (ANNs), and support vector machines (SVMs). K-NN, introduced in 1967, classifies samples based on the distance between feature values in the training data. It was used by Kilic et al. to identify colonic polyps from CT colonography and by Reddy et al. for brain and pancreatic cancer classification using the gray-level co-occurrence matrix (GLCM). However, k-NN is limited by local structure sensitivity and overfitting. ANNs, conceptualized in the 1940s by McCulloch and Pitts, use layered architectures with back-propagation for error correction. Saftoiu et al. used ANNs to differentiate chronic pancreatitis from pancreatic adenocarcinoma on EUS images with 94% sensitivity. SVMs, developed by Vapnik et al. in 1995, define hyperplane boundaries for classification. Zhang et al. applied SVM to identify pancreatic cancer from EUS images, achieving a detection accuracy of 99.07% using 29 textural features.
Deep learning architectures: DL networks extract all features from medical images rather than pre-selected ones, making them preferred for digestive cancer detection. CNNs are the most extensively used DL technique, comprising input, convolutional, activation, pooling, fully connected, and output layers. U-Net architectures, which use fewer convolutional layers, have also been commonly applied for pancreatic segmentation. Oda et al. employed a 3D FCN model to segment the pancreas automatically from CT images, achieving an average Dice score of 89.7 plus or minus 3.8 (a Dice score above 88% is considered highly precise). Guo et al. used a Gaussian mixture model with a 3D U-Net segmentation technique, achieving a Dice score of 83.2 plus or minus 7.8%.
Other CNN variants discussed include LeNet (1989), AlexNet, VGGNet, Inception Net, and ResNet, introduced between 2012 and 2015, each differing in the number of convolutional and pooling layers. The one-stage classification method segments medical images into grids for direct classification, while the more time-consuming two-stage method (R-CNN, Fast R-CNN, Faster R-CNN) demarcates candidate regions of interest for more accurate predictions.
Endoscopic ultrasound (EUS) uses high-frequency ultrasound probes positioned close to the pancreas for high-resolution imaging. EUS sensitivity for detecting pancreatic lesions ranges from 85-99%, which is considerably better than CT. For tumours with a diameter of 3 cm, EUS achieved 93% diagnostic accuracy, significantly outperforming CT (53%) and MRI (67%). EUS can detect lesions as small as 2-3 mm. EUS-guided fine needle aspiration (EUS-FNA) achieves diagnostic accuracies of up to 85% with nearly 100% specificity for evaluating primary tumour sites and lymph nodes, compared to just 50% accuracy for CT-assisted diagnosis. However, EUS-FNA requires experienced operators and is not available at many healthcare institutions.
Differentiating PDAC from chronic pancreatitis: A major clinical challenge is distinguishing pancreatic ductal adenocarcinoma (PDAC) from chronic pancreatitis (CP), since inflammation masks neoplastic features. Norton et al. (2001) used neural networks to analyse EUS images for differentiating PDAC and CP using four image parameters, achieving high sensitivity but only 50% specificity. Zhu et al. improved on this by using SVM to extract 105 features from EUS images of 262 pancreatic cancer patients and 126 CP patients, selecting 16 features to differentiate the conditions with 94% sensitivity. Zhang et al. applied SVM to differentiate PDAC from normal tissue using 29 EUS features with 97.98% sensitivity. Das et al. combined image analysis with ANNs for demarcating cancerous zones in EUS images with 93% accuracy.
Computer-aided diagnosis (CAD) systems: Ozkan et al. used multilayer perceptron neural networks (MNNs) to categorize EUS images of malignant and non-malignant tissues across age groups (under 40, 40-60, and over 60 years), achieving 97% accuracy with training data and 95% accuracy with testing data. A deep learning-based EUS-CAD system was developed to identify PDAC, CP, and normal pancreas, using 920 training images and 470 testing images. The detection efficiency was 92% in validation and 94% in testing.
Predicting IPMN malignancy: Deep learning algorithms were applied to EUS images of intraductal papillary mucinous neoplasms (IPMNs), precursors of PDAC, using 3,970 images acquired before pancreatectomy. The deep learning model achieved a malignant IPMN probability of 0.98 (p less than 0.001), with 95.7% sensitivity, 92.6% specificity, and 94.0% accuracy, significantly outperforming human diagnosis at 56.0%.
MRI provides non-invasive imaging of soft tissues by measuring the relaxation times (T1 and T2) of protons in water molecules under an external magnetic field. Contrast agents such as gadolinium chelates or iron oxide nanoparticles significantly enhance resolution and sensitivity. Despite these strengths, MRI struggles to identify subtle pre-malignant changes such as pancreatic intraepithelial neoplasia, which commonly precedes PDAC. Even individuals with stage I (localized) pancreatic cancer have only a 39% five-year survival rate, underscoring the need for better early detection.
Survival prediction: A supervised ML algorithm trained on a cohort of 102 MRI images (with 30 additional images for testing) was developed to predict overall survival rates in PDAC patients. The algorithm segmented images and extracted features, achieving 87% sensitivity and 80% specificity. The substantial overlap between the ML predictions and clinical histopathological conclusions indicates the promise of this approach for classifying pancreatic cancer subtypes.
GAN-augmented deep learning: A study investigating deep learning for distinguishing pancreatic diseases used T1 contrast-enhanced MRI data from 398 subjects (ages 16-85). Generative adversarial networks (GANs) were employed to generate synthetic images that augmented the training dataset. The Inception-V4 network (a CNN variant with multiple hidden layers) was trained on the GAN-augmented data and validated using MRI images from two separate hospital centres, comprising 50 images (ages 24-85) and 56 images (ages 26-80). Zhang et al. used a quantum genetic algorithm to optimize SVM parameters for improved paediatric pancreatic cancer detection on MRI, while Devi et al. combined ANNs with SVM for classification, extracting features using the GLCM method. Using the JAFER algorithm for feature selection and five classification techniques (ANN BP, ANN RBF, SVM Linear, SVM Poly, and SVM RBF), the ANN back-propagation approach achieved 98% classification accuracy.
IPMN detection and organ segmentation: Corral et al. used CNNs to classify MRI scans for intraductal papillary mucinous neoplasia (IPMN) in 139 individuals, of whom 22% had a normal pancreas, 34% had low-grade dysplasia, 14% had high-grade dysplasia, and 29% had adenocarcinoma. The model achieved 92% sensitivity and 52% specificity for dysplasia detection, with an overall accuracy of 78% compared to 76% for American Gastroenterology Association guidelines. Chen et al. developed an automated deep learning model (ALAMO) for multi-organ segmentation from abdominal MRI, incorporating multiview, deep connection, and auxiliary supervision training procedures across ten organs including the pancreas.
Computed tomography is the most widely used imaging modality for pancreatic cancer, employing X-rays at multiple angles to generate reconstructed 3D images. However, CT poses challenges for accurate cancer diagnosis due to irregular contours, overlapping densities from vasculature, bony structures, and soft tissues, as well as fuzzy and noisy images with insufficient contrast. AI-driven methods that enable image segmentation, contour identification, and disease classification can significantly improve diagnostic performance.
Pre-diagnostic risk detection: In a notable study, approximately 19,500 non-contrast CT images from 469 scans were segmented using CNNs to compute mean pancreatic tissue density in Hounsfield units (HU) and pancreatic volume. Comparison of pre-diagnostic scans from individuals who later developed PDAC versus those who remained cancer-free revealed a significant reduction in mean whole-gland pancreatic HU (0.2 vs. 7.8 in PDAC developers), suggesting that HU attenuation in CT images could serve as an early risk indicator for PDAC.
Classification performance: A CNN with four hidden layers trained on CT images from 222 PDAC patients and 190 controls achieved 95% diagnostic accuracy, though it did not surpass human expert predictions. Zhang et al. employed feature pyramid networks with a recurrent CNN (R-CNN) on 2,890 CT images, achieving 94.5% classification accuracy but acknowledging limitations from closed-source data. A more advanced approach combined a 16-layer VGG16 CNN with R-CNN on 6,084 enhanced CT scans from 338 PDAC patients, reaching approximately 96% prediction accuracy with each image processed in just 0.2 seconds. Liu et al. achieved standout results with a CNN tested on 370 patients and 320 controls: 97.3% sensitivity and 100% specificity on one test set, and 99% sensitivity and 98.9% specificity on a second test set. Chen et al. developed a deep learning algorithm specifically for detecting tumours smaller than 2 cm, achieving 89.7% sensitivity and 92.8% specificity overall, with 74.7% sensitivity for sub-2 cm cancers.
Pancreatic cyst classification: CNN models were applied to CT images from 206 patients with different pancreatic cysts, including 64 with IPMNs, 66 with serous cystic neoplasms (SCN), 35 with mucinous cystic neoplasms (MCN), and 41 with solid pseudopapillary epithelial neoplasms (SPEN). The DenseNet architecture, which uses dense layers receiving inputs from all nodes via the shortest route, outperformed conventional CNNs across all cyst categories with a highest accuracy of 81.3% for IPMNs, followed by 75.8% for SCNs and 61% for SPENs. However, the study lacked information on tumour size and failed to explain the sources of classification errors.
Positron emission tomography uses short-lived radioisotope tracers (commonly 18F-FDG, exploiting the high glucose consumption of cancer cells) that emit positrons. PET provides functional information about organ activity and has demonstrated sensitivity above 85% for pancreatic cancer. However, PET has important limitations: dysregulated glucose metabolism and inflammation can produce false positives, and tumours smaller than 2 cm are often missed. PET is frequently combined with MRI or non-contrast CT to compensate for its poor spatial resolution.
Hybrid AI classification: Li et al. used 18F-FDG PET/CT images of 40 cancer patients and 40 controls with a multi-step AI pipeline. Regions of interest were segmented using simple linear iterative clustering (SLIC), features were extracted via dual-threshold principal component analysis (DT-PCA), and classification was performed using a hybrid feedback-SVM-random forest algorithm (HFP-SVM-RF). The hybrid model achieved 96.5% accuracy on the clinical dataset. When tested on 82 public PET/CT scans, it showed a Dice coefficient similarity of 78.9% and a Jaccard index of 65.4% against ground-truth contours, indicating room for improvement.
Radiomics for survival prediction: A combination of radiomics and machine learning was used for prognostic prediction from 18F-FDG-PET scans of 138 patients with pancreatic cancer. A random forest model classified 42 extracted features and identified the gray-level zone length matrix (GLZLM) and gray-level non-uniformity (GLNU) as the top factors influencing one-year survival, with total lesion glycolysis ranking second. This information was used to stratify individuals into poor-prognosis groups with a high risk of mortality.
The review notes that every imaging modality requires customized robust algorithms to extract the subtle but distinctive features of pancreatic cancer. No single AI model performs universally well across all modalities, and the choice of algorithm must be tailored to the specific characteristics and limitations of each imaging technique.
Serological biomarkers: CA 19-9 is the most extensively explored protein biomarker for pancreatic cancer, but multiple studies have shown it cannot serve as a standalone predictor. AI-based approaches aim to integrate multiple biomarkers for improved accuracy. An SVM algorithm with recursive feature elimination (RFE) screened gene expression datasets from 78 samples and identified seven gene targets, including FOS (leucine zipper protein), MMP7 (matrix metalloproteinase-7), and A2M (alpha-2-macroglobulin), as more accurate diagnostic markers detectable in both serum and urine. ANN-based analysis of CA 19-9, CA125, and carcinoembryonic antigen (CEA) levels from 913 samples showed improved detection accuracy compared to single-marker prediction.
Exosome-based miRNA profiling: Exosomes containing cancer-specific miRNA (including miR-21, miR-196a, miR-1290, miR-1246, and others) are gaining importance for pancreatic cancer diagnosis. In a seminal study, exosomes were captured using antibodies against EpCAM (epithelial cell adhesion molecule), their RNA cargo was isolated, and miRNA signatures were identified using qPCR. A custom machine learning algorithm was then validated on samples from normal individuals and pancreatic cancer patients with good prediction accuracy. A neural network algorithm screened 140 datasets of pancreatic cancer patients for urinary gene biomarkers (REG1A/1B, LYVE1, TFF1, and CA 19-9) and predicted REG1A/1B as the most important urinary biomarker with an importance ratio exceeding 80%.
Digital cytopathology: Deep learning algorithms such as VGG, DenseNet, and ResNet, along with ML algorithms based on SVM and random forest, are being applied to extract tumour features from tissue biopsies. SA-GAN (stain acclimatization generative adversarial network) addresses the challenge of staining intensity variation across laboratories by normalizing colour across images. A deep learning-based spiral algorithm transformed 3D MRI images of pancreatic tissue into 2D images while preserving texture and edge parameters. CNN models with bilinear pooling modules classified images as having TP53 gene mutations or not, with prediction results agreeing well with actual mutation status. This offers a non-invasive alternative to painful biopsies for gene mutation classification.
Risk prediction: Muhammad et al. used ANNs to predict and stratify pancreatic cancer risk from personal health data into low, medium, or high categories, achieving 80.7% sensitivity and specificity. The detection of subtle textural and morphological changes in abdominal scans through customized AI algorithms could enable pre-symptomatic risk identification.
Data limitations: The widespread clinical deployment of AI-driven pancreatic cancer diagnosis has not yet been realized, largely because most studies rely on small, single-centre datasets that are insufficient to convincingly train and validate algorithms. Many of the studies reviewed used closed-source or proprietary data, limiting reproducibility. The variation in study populations, tumour sizes, and comparison conditions (cancer vs. normal, cancer vs. chronic pancreatitis) across different reports makes direct performance comparisons difficult. Most models have not undergone rigorous multi-centre prospective validation, which is essential before any clinical adoption.
Black-box problem: Most AI models operate in a "black box" mode, meaning clinicians cannot understand or explain the basis of the identification or stratification decisions. This lack of interpretability creates reluctance among medical professionals to trust and adopt these tools. The absence of explainable AI frameworks specifically designed for pancreatic cancer imaging remains a significant gap that must be addressed before clinical integration.
Ethical concerns: The review identifies several ethical issues. AI tools require large datasets for training and validation, raising concerns about data privacy and confidentiality. There is currently no structured regulatory framework for AI-based tool development, covering data collection, storage, processing, and sharing. Frequent comparisons between AI and expert clinician predictions have fueled concerns about potential de-skilling of future clinicians due to over-dependence on AI. Additional issues include the erosion of the patient-doctor relationship and questions of accountability when AI tools produce wrong diagnoses with potentially disastrous consequences.
Future directions: The exponential growth in computing resources and open-source tools has triggered an increasing number of studies focused on more robust algorithms for accurate, rapid, and early diagnosis. Key areas for advancement include the development of large, multi-centre annotated datasets, the integration of multi-modal data (combining imaging with biomarkers and genetic data), the creation of explainable AI models that clinicians can trust, and new regulatory policies governing AI deployment in healthcare. The emergence of digital cytopathology, combining AI with tissue biopsy analysis, may establish a new standard for cancer detection and stratification. Reinforcement learning, still unexplored in pancreatic cancer, could eventually enable automated clinical decision-making in remote settings.