AI reading tumour slides picked out pancreatic patients facing double the recurrence risk

Mayo Clinic researchers analyzed pathology slides from 203 patients with pancreatic ductal adenocarcinoma and found that the physical layout of leftover tumor tissue, not just how much of it remained after treatment, flagged who was most likely to see their cancer come back. Patients whose residual cancer and surrounding tissue formed a fragmented, tangled pattern had earlier recurrence than those whose leftover disease sat in cleaner, more contained clusters. In one statistical model built from that spatial pattern, high-risk patients carried a 71% higher adjusted risk of recurrence; in a second model, their adjusted risk ran to more than twice that of lower-risk patients.

All 203 patients had received chemotherapy before surgery but had only a limited pathologic response, meaning a meaningful amount of cancer was still visible under the microscope once the tumor came out. That is the exact group where oncologists have struggled to sort who needs aggressive follow-up therapy from who does not, because the standard measure, how much cancer is left, has not reliably done that sorting on its own.

Two spatial models separated high-risk patients from low-risk ones

The findings, published in Clinical Cancer Research and described in a Mayo Clinic news release, came from a team led by Ryan Carr, a Mayo Clinic oncologist and the study’s senior author. Carr’s group combined an AI-enabled digital pathology platform with analytic methods borrowed from landscape ecology, the science normally used to study how patches of forest, grassland or wetland are arranged across a landscape, and applied it to standard hematoxylin and eosin slides already sitting in pathology archives. Rather than asking how much cancer remained, the software measured how the cancerous regions and the surrounding stroma formed patches, boundaries and mixed zones, then tied those geometric patterns to which patients had earlier disease-free survival.

Both resulting spatial signatures predicted disease-free survival even after Carr’s team accounted for tumor stage, lymph node status and the other clinical and pathologic risk factors oncologists already use. Carr said the standard pathology report mostly tells clinicians how much tumor is left after treatment, and that his team wanted to know whether the arrangement of that remaining cancer could reveal biology the volume measurement misses.

The digitized slides Carr’s team analyzed are the same H&E slides a hospital pathologist already reviews under a microscope after surgery. The AI platform simply re-renders them at a resolution fine enough to trace individual patches of tumor and stroma across an entire tissue section, the way an ecologist might map patches of trees, meadow and wetland across a forest plot. That patch-level view is what let the models separate patients with a fragmented, checkerboard-style pattern of residual cancer from patients whose leftover disease sat in fewer, more solid clusters, a distinction invisible to a pathology report that only counts how many square millimeters of tumor remain.

Standard pathology missed what the AI geography caught

The spatial models distinguished higher- and lower-risk patients in cases where the conventional measure, the raw amount of residual cancer, did not. That gap matters because a “limited pathologic response” is often treated in clinics as a reassuring label suggesting chemotherapy worked reasonably well, yet the Mayo data show plenty of patients inside that label are still heading toward recurrence. Carr’s team traced the difference not to how much cancer survived treatment but to how it was arranged across the tissue.

Because the technique works from slides pathology labs already generate as part of routine care, Carr said it could hand oncologists another data point on recurrence risk without ordering a new tissue biopsy or genetic panel. He described the underlying signal as information already present in the tissue, one AI-enabled analysis can measure in ways difficult to catch by eye, adding another layer of insight to how clinicians assess risk after surgery.

Fewer immune cells reached inside the highest-risk tumors

The spatial analysis also picked up an immune pattern: tumors with the high-risk geography contained fewer immune cells inside the cancer itself, with immune cells instead collecting around the tumor’s edge rather than penetrating it. Researchers say that pattern points to the tumor microenvironment, the mix of cells and tissue surrounding a tumor, as a factor in why some pancreatic cancers resist treatment and recur while others do not. An immune system that never breaches the tumor’s border cannot help clear the cancer cells chemotherapy left behind, which may explain why a fragmented, hard-to-defend tissue layout tracked with earlier recurrence in both spatial models.

The result feeds into a broader research push at Mayo Clinic, where Carr’s lab studies how pancreatic cancer cells interact with neighboring tissue using the same machine-learning and spatial-analysis tools, an effort the institution has folded into what it calls its Precure Research priority, using data and technology to flag risk earlier, before disease has a chance to advance.

The tool works from slides hospitals already have

Carr’s team says the approach still needs prospective testing, meaning it has to be applied ahead of time to new patients rather than validated after the fact against outcomes that already happened, before it could inform real treatment decisions such as which patients get more aggressive adjuvant therapy or closer surveillance after surgery. Carr said the long-term goal is identifying which patients remain at the greatest risk and using that information to guide more individualized surveillance schedules, adjuvant therapy choices and clinical trial design.

For now, the 203-patient study stands as a proof of concept rather than a bedside tool, a demonstration that the shape of what is left behind after chemotherapy, not merely its volume, carries a measurable recurrence signal routine pathology has been overlooking. The Gerstner Family Foundation, the Mayo Clinic Center for Clinical and Translational Science and the federal ARPA-H ADAPT program helped fund the work, published July 15 in Clinical Cancer Research.

More from Morning Overview

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.