Aiforia Breast

Evaluation of the Aiforia breast AI in Memorial’s routine sign-out workflow.

Purpose

Assess a commercial breast AI in routine sign-out rather than in a curated research setting. The specific endpoint is not documented. [unverified] Likely candidates given the vendor’s scope are Ki-67 or other IHC quantification, but the chapter does not say. [unverified]

Data used

Not documented. [unverified]

Methods

Not documented. [unverified]

Current state / open questions

Active. Chapter is a stub: repo and site links plus “ECDP poster” under WIP, indicating a European Congress on Digital Pathology abstract is the current output. Site: aiforiabreast.patoloji.dev.

Open questions for the next refresh: which readout, what reference standard, and whether the comparison is against a single reader or a consensus panel. “Routine workflow” framing is a strength here — it makes the result closer to real-world performance than a curated test set would be — so it is worth documenting the routine-ness precisely.

Related reading

Nottingham Grading and Mitotic Count — added 2026-07-25 from sources/papers/rakha-2026-wsi-breast-grading.md. Relevant to this project in a specific way: mitotic counts are systematically underestimated on WSI relative to light microscopy. If the algorithm reads whole slide images while the reference standard was read on glass, that difference is a bias in the comparison design rather than a property of the algorithm. Whether it applies depends on the readout and on the reference standard’s modality — neither of which is documented here yet, which makes it worth settling before the ECDP output rather than after.

Ki-67 Proliferation Index — added 2026-07-26 from sources/papers/shaaban-2026-uk-ki67-recommendations.md. If this evaluation’s readout is Ki-67 — the most likely candidate given the vendor’s scope, though still [unverified] here — then the UK national recommendation is the reference standard the algorithm should be measured against, and its claim that AI scoring beats manual scoring on reproducibility is precisely what a routine-workflow evaluation would be testing locally. Confirming the readout is now doubly worth doing: it selects which of two literatures (grade/mitotic-count vs Ki-67) governs the comparison.

Derived from: patoloji-ai/aiforiabreast.qmd, projects/projects.yml.