Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers, with a five-year survival rate below 12%. The poor prognosis is largely due to the fact that most patients are diagnosed at an advanced, inoperable stage when symptoms finally appear.
Medical imaging plays a central role in diagnosing and staging pancreatic cancer. Computed tomography (CT) is the primary modality used, but interpreting pancreatic lesions accurately requires significant radiological expertise, and even experienced clinicians can miss subtle early-stage tumors.
Computer-aided diagnosis (CAD) systems are designed to assist radiologists by automatically analyzing medical images and flagging areas of concern. Advances in artificial intelligence have made CAD systems for pancreatic cancer increasingly viable, motivating this comprehensive review of recent progress.
A typical CAD workflow for pancreatic cancer begins with acquiring a CT or MRI scan, followed by preprocessing steps such as noise reduction and intensity normalization to ensure consistent image quality across different scanners and institutions.
The next step is segmentation, which involves delineating the pancreas and any lesions within the image. This is a challenging task because the pancreas is a small, irregularly shaped organ that varies greatly in appearance between patients and has low contrast with surrounding tissues.
After segmentation, relevant features are extracted either by hand-crafting mathematical descriptors or by using deep learning to automatically learn discriminative patterns. These features are then fed into a classifier to produce a diagnostic prediction such as malignant versus benign.
Sensitivity measures the proportion of true cancer cases that the AI system correctly identifies. In cancer diagnostics, high sensitivity is critical to avoid missing malignancies, which can have fatal consequences for patients.
Specificity measures the proportion of healthy cases correctly identified as non-cancerous. A system with low specificity generates many false positives, leading to unnecessary procedures, patient anxiety, and increased healthcare costs.
The Area Under the Receiver Operating Characteristic Curve (AUC) summarizes overall diagnostic performance by plotting sensitivity against 1-specificity at various classification thresholds. An AUC of 1.0 represents perfect discrimination, while 0.5 indicates no discriminative ability.
For segmentation tasks, the Dice Similarity Coefficient (DSC) is the primary metric, measuring the overlap between the predicted segmentation mask and the ground truth annotation. DSC ranges from 0 (no overlap) to 1 (perfect overlap).
Radiomics refers to the high-throughput extraction of large numbers of quantitative features from medical images. These features capture tumor shape, texture, intensity distribution, and heterogeneity in ways that are not visible to the naked eye.
Common radiomics features include first-order statistics (mean, variance, skewness of pixel intensities), texture features derived from gray-level co-occurrence matrices (GLCM), and shape descriptors that quantify tumor geometry. Hundreds of features can be extracted from a single tumor region.
Because radiomics generates far more features than patients, dimensionality reduction techniques such as LASSO (Least Absolute Shrinkage and Selection Operator) and principal component analysis are applied to select the most informative features and prevent overfitting in downstream classifiers.
Popular classifiers used in radiomics studies include Support Vector Machines (SVM), Random Forests, and XGBoost, a gradient boosting framework that consistently achieves strong performance on structured data and has become a leading method in radiomics pipelines.
Convolutional Neural Networks (CNNs) have transformed medical image analysis by automatically learning hierarchical feature representations directly from pixel data. For pancreatic segmentation, encoder-decoder architectures such as U-Net are particularly effective because they combine high-level semantic understanding with fine-grained spatial detail.
Transformer-based models apply self-attention mechanisms to capture long-range spatial relationships in medical images. Hybrid architectures that combine CNN local feature extraction with transformer global context have shown promising results for pancreas segmentation, where the organ's irregular shape requires understanding both local texture and global anatomical context.
Generative Adversarial Networks (GANs) are used in pancreatic cancer AI primarily for data augmentation, generating synthetic training images to address the common problem of small labeled datasets. GANs can produce realistic CT slices showing pancreatic lesions, helping models generalize better to real patients.
A persistent challenge for deep learning in pancreatic imaging is the scarcity of large, annotated public datasets. Unlike natural image benchmarks, annotating medical images requires specialized radiological expertise, making it time-consuming and expensive to build the training sets needed for robust deep learning models.
Studies reviewed in this paper demonstrate consistent improvement in pancreatic cancer detection over the past five years. Recent deep learning segmentation models achieve Dice Similarity Coefficients above 0.80 on well-curated single-center datasets, a substantial improvement from early approaches that struggled to exceed 0.70.
For lesion detection and classification, several radiomics studies report AUC values above 0.90 for distinguishing malignant from benign pancreatic lesions using CT features. However, performance often degrades when models are evaluated on external datasets from different hospitals, highlighting generalizability as a major unresolved challenge.
Fusion models that combine radiomics features with deep learning representations consistently outperform either approach alone. The complementary nature of hand-crafted texture descriptors and automatically learned spatial features allows fusion models to capture a broader range of diagnostically relevant information from each scan.
One of the most pressing needs in the field is the development of large, multicenter, publicly available datasets for pancreatic cancer imaging research. Most existing studies use small private datasets from single institutions, making it difficult to verify whether results reflect true generalizability or dataset-specific patterns.
Federated learning is a promising approach for building larger effective training sets without sharing raw patient data. In federated learning, AI models are trained locally at each hospital and only model parameter updates are shared, preserving patient privacy while enabling collaborative model development across institutions.
Clinical adoption of AI diagnostic tools requires not only high performance but also interpretability. Explainability methods such as gradient-weighted class activation maps (Grad-CAM) visualize which image regions drove a model's prediction, helping radiologists understand and trust AI recommendations rather than treating them as black-box outputs.
Future CAD systems will likely incorporate multimodal data integration, combining imaging findings with genomic profiles, clinical laboratory values, and patient history to produce more holistic and accurate assessments of pancreatic cancer risk and staging.