Jun 2026 · Lightning talk · 2026 AI Research Summit, Mayo Clinic
Knowing When to Ask the Pathologist
Uncertainty-aware detection of rare leukemia cells in flow cytometry
Huynh T. T. Tran · Brian Jalaian, with collaborators at Mayo Clinic
My Ph.D. student Huynh Tran gave a lightning talk on this work at the 2026 AI Research Summit, hosted by Mayo Clinic in Rochester, Minnesota, in June. It is a collaboration with Mayo Clinic hematopathology. The paper is still being finalized, so this post covers the idea, not the results.
Measurable residual disease testing means finding a few leukemia cells among a great many normal ones. A model that is usually right is not enough: it also has to know which of its calls are shaky, so a pathologist can review exactly those.
The problem
After treatment, leukemia patients are tested for measurable residual disease: the small number of abnormal cells that can remain. Flow cytometry measures a very large number of cells, and the abnormal ones are rare, so the task is a search for needles in a haystack. Mistakes in either direction matter, and reviewing everything by hand is slow.
The idea
A classifier makes the call for each case, and a second model, an entropy-enhanced predictive meta-model, estimates how far that call can be trusted. Confident calls are decided automatically. Uncertain ones are routed to a pathologist.
The point is not to replace the expert. It is to spend the expert's time on the cases that actually need it.
Why it matters
In clinical settings, a model's errors are only acceptable if the system knows when it might be wrong. Uncertainty-aware triage is a practical way to bring AI into the lab while keeping a human in charge of the hard calls.