This Week in AI Pathology — 2026, week 32

A consensus statement on what a laboratory should require before adopting an algorithm. This is the genre of paper the field has least of and needs most.

Notable for reporting error patterns, not just accuracy. How a model fails matters more than how often — see Ethics, Bias & Failure Modes.

Cytopathology remains thinly covered by AI work relative to its diagnostic burden.

A critique paper. Worth reading alongside the site-signature result — this field's most useful papers are often the ones testing whether a result means what it appears to.

Human factors rather than algorithms: how pathologists actually search a slide. Relevant to anyone designing a tool meant to fit into that behaviour.

A published methodological disagreement, with the reply. Reading both is more instructive than reading either.

PLOS ONE — multi-dataset, which is the part that matters.

Medical Image Analysis — compare against existing nucleus models.

New from the lab behind CLAM, UNI, CONCH and TITAN, all of which are catalogued. Worth watching on that track record alone; too new to catalogue until there is a paper and a stated licence.

A focused list for pathology spatial omics — deeper on that topic than our section intends to be.