What we will prove, and how.

A second-reader tool makes no findings of its own. What it changes — if it changes anything — is how a radiologist reads. So "how accurate is the software?" is the wrong question to lead with.

The unit of evaluation is not the algorithm. It is the radiologist reading with it.

The design built for that question is the multi-reader, multi-case study, MRMC for short. Several radiologists read the same retrospective cases twice: once without the software, and — after a pause long enough that no one remembers individual cases — once with it. Because every reader reads every case both ways, the study can attribute a change in reading performance to the software itself, rather than to one sharp-eyed reader or one easy case mix.

That is the standard of evidence we consider appropriate here, and this site holds itself to it: until the study is run and independently analysed, no clinical performance claim appears on any page.

The study in preparation uses one hundred retrospective brain MRI cases across three clinical strata: multiple sclerosis, primary brain tumours, brain metastases. Every participating radiologist reads every case twice — first unaided, later aided by Lävi Delta — with an enforced washout period between the phases.

Fig. 1 — Planned study designin preparation; final design subject to methodology review

The reference standard and the statistical analysis are planned to be independently defined and reviewed. The endpoints being prepared cover localisation performance, false-positive burden, reading time and reader confidence, alongside standalone technical performance.

We publish no sample-size justification, no power figures, no target effect sizes and no dates. Those belong in the study protocol, not on a website.

Clinical validation programme in preparation

Reader studies fail on logistics about as often as on science, so we built the reading environment first. A dedicated browser-based reader workstation exists and is deployed — not a slide deck, a running system.

What is implemented and tested in it: role-based access with two-factor authentication; the blinding gate between arms, so a reader in the unaided phase cannot reach the software's outputs even on purpose; washout enforcement between phases; structured capture of every reader response; a full audit trail; and a results export for the analysts.

The pipeline and this reader environment have been run together end to end in a full production-form deployment. That included the unglamorous half of the test: cases the pipeline rejects were confirmed to stay rejected, not slip quietly through.