CASE · Q1 — A slide-level melanoma model is offered for 14,000 mixed skin specimens. Which cases are in scope, what is missing from its output, and what blocks routine use?#
CASE · Q2 — Which of worklist triage, in-viewer overlay, and benign fast-track fit an adjunct intended use? How should slide scores become patient decisions?#
CASE · Q3 — For 10,000 slides and 5% prevalence, calculate TP/FP/FN, PPV, NPV, and flags/day at each threshold. Which point fits assistant versus rule-out?#
CASE · Q4 — A CE-IVD PD-L1 model changes antibody, stainer, scanner, and specimen type. Design site validation around 1% and 50% cut-offs and define ground truth.#
CASE · Q5 — Case mix changes and a new HER2-low therapy makes 0 versus 1+ clinically important. What drift checks and new validation are required?#
CASE · Q6 — If “IHC not needed” cases never receive IHC, why is zero disagreement not 100% NPV? Design an ethical evaluation of the prediction paradox.#
CASE · Q7 — Radiologist–AI agreement rises to 98.6%. Does that prove benefit or calibrated trust? Redesign policy and interface around automation bias.#
CASE · Q8 — Using the licence, integration, IHC, and reading-time figures, does pathology AI pay for itself? Who pays, who benefits, and what non-financial case remains?#