AI Spatial Analysis Predicts Pancreatic Cancer Recurrence Risk

AI reads tumor geography to forecast relapse

Mayo Clinic researchers have demonstrated that artificial intelligence can detect spatial patterns in routine pathology slides that help predict which pancreatic cancer patients face the highest risk of recurrence after surgery and chemotherapy.

The approach analyzes how residual cancer tissue is organized—its fragmentation, boundaries, and intermixing with surrounding stromal tissue—rather than simply measuring how much tumor remains. In a study of 203 patients with pancreatic ductal adenocarcinoma who showed limited response to pre-surgical chemotherapy, spatial configuration proved more predictive than tumor volume alone.

"Current pathology assessments largely tell us how much tumor is left after treatment. We wanted to know whether the geography of that remaining cancer could reveal additional biology about recurrence risk," said Ryan Carr, M.D., Ph.D., a Mayo Clinic oncologist and senior author of the study published in Clinical Cancer Research.

Why it matters

Pancreatic cancer remains one of the deadliest malignancies, and identifying which patients face the highest recurrence risk could enable more targeted surveillance and adjuvant therapy decisions. Because this method analyzes standard hematoxylin and eosin (H&E) slides already generated during routine care, it requires no additional tissue sampling—only computational analysis of existing pathology materials. That positions it as a potentially scalable tool for risk stratification without adding procedural burden.

Fragmentation signals danger

The research team used an AI platform to identify cancer and stromal regions on standard pathology slides, then applied methods adapted from landscape ecology to measure tissue shape, fragmentation, and intermixing patterns. Patients whose tumors showed more fragmented, intermixed patterns of cancer and stroma experienced earlier recurrence.

Two spatial signatures predicted disease-free survival even after accounting for stage, lymph node status, and other established clinical risk factors. In one model, high-risk patients had a 71% higher adjusted risk of recurrence. In another model, high-risk patients faced more than twice the adjusted risk. These spatial models successfully distinguished patients at higher and lower risk when standard measures, including residual tumor volume, did not.

Immune exclusion patterns emerge

The analysis also revealed that high-risk spatial patterns contained fewer immune cells within the cancer tissue itself. Instead, immune cells tended to accumulate around the tumor perimeter rather than infiltrating it—a finding that highlights the role of the tumor microenvironment in treatment resistance and disease progression.

Carr's broader research applies ecological principles to cancer biology, using machine learning and spatial analysis to map the pancreatic cancer ecosystem and study how cancer cells interact with neighboring cells and tissues.

Path to clinical implementation

The researchers emphasized that while results are promising, prospective studies are needed before this approach can inform clinical decision-making. The work aligns with Mayo Clinic's Precure Research priority to use data and technology to predict risk earlier and create opportunities to intercept serious disease before it advances.

"Our long-term goal is to better identify which patients remain at greatest risk and ultimately use that knowledge to guide more individualized surveillance, adjuvant therapy and clinical trial design," Carr said.

The study was supported by the Gerstner Family Foundation Career Development Award, the Grand Forks Career Development Award, the Mayo Clinic Center for Clinical and Translational Science, and the ARPA-H ADAPT program. Details were first reported by Mayo Clinic and published in Clinical Cancer Research.