Pancreatic cancer has one of the worst prognoses of any cancer, with a 5-year survival rate of just 12% — the lowest of any common cancer. This poor outcome is largely because symptoms are vague and non-specific until the disease is advanced, and imaging-based diagnosis remains technically challenging.
The most common form, pancreatic ductal adenocarcinoma (PDAC), accounts for 80-85% of cases and presents with advanced local or distant metastatic disease in most patients at diagnosis. Only 15-20% of patients qualify for surgery — the only potentially curative treatment — highlighting the urgent need for better early detection tools.
This comprehensive review analyzed 198 research papers on AI applications in pancreatic cancer imaging. It covers five imaging modalities: CT (computed tomography), MRI (magnetic resonance imaging), EUS (endoscopic ultrasound), PET (positron emission tomography), and pathological slide images — each with distinct strengths and limitations for pancreatic cancer diagnosis.
Within each modality, AI tasks fall into four categories: segmentation (outlining tumor boundaries), classification (benign vs. malignant, cancer type), object detection (finding lesions), and prognosis prediction (estimating survival). Deep learning methods, particularly convolutional neural networks (CNNs), dominate all four task categories.
CT imaging is the most studied modality for PDAC, while EUS is highly valued for its ability to detect small lesions missed by CT or MRI. Pathological images — microscopy slides of biopsied tissue — are increasingly analyzed by AI to predict outcomes beyond what pathologists can assess visually.
In CT imaging, AI models have achieved comparable or better performance than experienced radiologists for detecting PDAC, segmenting tumor boundaries, and predicting lymph node metastasis. Deep learning models also show promise for identifying precursor lesions like intraductal papillary mucinous neoplasms (IPMNs) before they become malignant.
For EUS — the preferred modality for detecting small pancreatic lesions — AI has improved the ability to classify tumor types and distinguish PDAC from chronic pancreatitis, a common diagnostic challenge. Radiomics approaches on EUS images have shown particular promise for characterizing neuroendocrine tumors.
In pathological imaging, deep learning models analyzing whole-slide images can predict patient prognosis, molecular subtypes, and treatment response from tissue morphology — providing information that would require expensive genomic tests through imaging alone.
Current hot topics include multi-modal fusion — combining two or more imaging types to boost diagnostic accuracy — and large foundation models that can be adapted to multiple pancreatic imaging tasks from limited data. Federated learning, which trains AI across multiple hospitals without sharing patient data, is emerging as a solution to the small dataset problem that plagues pancreatic cancer AI.
Key challenges include the scarcity of large, well-annotated pancreatic imaging datasets (partly because PDAC is relatively rare), high inter-institutional variability in image acquisition protocols, and the need for better model explainability so that clinicians can trust and act on AI outputs.
AI has the potential to transform pancreatic cancer care by enabling faster, more accurate diagnosis at lower cost — potentially shifting diagnoses from stage IV to earlier, more treatable stages. Fast, low-cost AI screening models could be deployed at scale to identify at-risk individuals who would benefit from further workup.
For treatment planning and prognosis, AI integration with electronic health records and genomic data promises to deliver personalized risk assessments and treatment recommendations. Realizing this potential will require standardized data collection, regulatory frameworks for clinical AI deployment, and robust prospective validation studies.