Pancreatic cancer is notoriously difficult to detect on medical imaging. The pancreas is a small organ buried deep in the abdomen, surrounded by other structures, and pancreatic tumors often lack clear borders or distinctive features that distinguish them from surrounding tissue. Even experienced radiologists miss early-stage tumors.
Artificial intelligence—particularly deep learning—offers a way to analyze CT, MRI, and endoscopic ultrasound (EUS) images with a level of consistency and pattern recognition that exceeds what human reviewers can achieve alone. AI doesn't get tired, doesn't introduce interobserver variability, and can be trained to detect subtle changes invisible to the human eye.
This systematic review analyzed 95 published studies (selected from 1,069 screened) to map the current state of AI applications in pancreatic imaging, covering tumor detection, segmentation, classification, prognosis, and treatment planning.
Following PRISMA guidelines (the international standard for systematic reviews), the authors searched three major medical databases and identified 1,069 potentially relevant studies. After removing duplicates and screening by relevance and quality, 95 studies were included in the final analysis.
The included studies spanned all major imaging modalities used in pancreatic disease: CT (the most common), MRI, endoscopic ultrasound (EUS), and PET. AI approaches studied included traditional machine learning, convolutional neural networks (CNNs), U-Net and its variants (specialized for medical image segmentation), and other deep learning architectures.
The review categorized AI applications into four main areas: segmentation (drawing boundaries around the pancreas or tumor), detection (identifying tumors that might be missed), diagnosis/classification (distinguishing cancer from benign conditions), and prognosis (predicting outcomes).
Pancreas segmentation—drawing precise boundaries around the pancreas in 3D CT images—is a challenging task because the organ varies widely in size, shape, and position across individuals. Deep learning models, particularly U-Net variants and cascade networks, have achieved Dice Similarity Coefficients (DSC) of 0.6–0.96, with the best approaches approaching expert human performance.
Hierarchical deep learning models—those that first identify the general region, then refine the boundary—showed the most consistent performance. The best models achieved DSC scores above 0.90, meaning their delineations overlapped with expert manual segmentations more than 90% of the time, while being far faster than manual methods.
Detection of small pancreatic lesions (under 2cm) remains the hardest challenge. Several DL models showed promising results in identifying tumors that were invisible or very subtle on routine CT scans, sometimes catching cancers up to 475 days before clinical diagnosis when retrospectively applied to historical scans.
CT imaging dominates current AI pancreatic research, partly because CT datasets are large and standardized. MRI provides better soft-tissue contrast but has more variability in acquisition protocols. EUS offers the highest spatial resolution for small lesions and is critical for tissue sampling, but AI training data for EUS is limited by smaller dataset sizes.
For CT-based detection of PDAC, leading models achieved sensitivities of 75–98% with specificities generally above 85%, substantially outperforming radiologists working without AI assistance in several studies. When AI was used as a 'second reader'—checking scans after the radiologist—detection rates improved further.
AI also showed promise in differentiating PDAC from benign pancreatic conditions such as autoimmune pancreatitis, which can appear almost identical on imaging. Distinguishing these conditions without AI currently requires endoscopic biopsy, but several models achieved accuracies of 85–94% from imaging alone.
Several AI applications reviewed are approaching clinical readiness. Computer-aided detection tools for pancreatic tumors on CT are being tested in prospective clinical studies, with some already deployed in specialist centers as decision-support aids for radiologists. Real-time AI assistance during endoscopic ultrasound procedures is also in active development.
A key clinical application is identifying 'incidental' pancreatic findings on CT scans done for other reasons. Many pancreatic cancers are discovered when a patient has a CT for an unrelated problem—an AI system watching for subtle pancreatic abnormalities in routine scans could systematically catch these cases earlier than current practice.
Risk stratification—using AI to predict which pancreatic cysts or lesions will progress to cancer—is another high-value application. Current management guidelines for pancreatic cysts involve surveillance imaging every 6-12 months, but AI models could identify which patients truly need close follow-up and which can be safely monitored less intensively.
The review identified several barriers to wider clinical adoption of AI in pancreatic imaging. Dataset size is a persistent limitation—pancreatic cancer is relatively rare, making it hard to amass the tens of thousands of annotated cases that large deep learning models ideally need. Many studies used fewer than 500 patients, raising concerns about generalizability.
Variability in CT acquisition protocols across hospitals means that a model trained at one institution may perform differently when deployed at another. Standardization of imaging protocols and development of AI models that are robust to this variability are essential steps before widespread clinical deployment.
Regulatory approval, liability frameworks, and clinician acceptance are non-technical challenges that will shape how quickly AI tools move from research to routine radiology. The evidence base for clinical benefit—improved outcomes, not just improved detection accuracy—is still being built.
This systematic review confirms that AI has already demonstrated significant capability in pancreatic imaging, with deep learning models achieving performance that rivals or exceeds human radiologists in several specific tasks. Segmentation, detection, and differential diagnosis have all benefited substantially.
The field is maturing rapidly: from early proof-of-concept studies using small single-center datasets to larger multicenter validations and prospective clinical studies. Endoscopic ultrasound AI in particular is entering an exciting phase, with real-time assistance tools showing strong performance in expert hands.
The ultimate promise of AI in pancreatic imaging is shifting pancreatic cancer from a disease diagnosed too late to one caught while surgery remains curative—a goal that would save thousands of lives annually. Achieving this will require continued investment in large, diverse, annotated imaging datasets and rigorous prospective clinical trials.