The problem — AUC and accuracy tell you how well a model ranks cases; they say nothing about whether acting on the model at a given operating point actually helps. A pipeline that outperforms a naïve classifier on every standard metric can still deliver negative net benefit if the decision threshold is miscalibrated to the real cost asymmetry between false positives and false negatives.
What it is / how it works — Decision Curve Analysis (DCA), introduced by Vickers and Elkin in 2006, plots net benefit — true positives minus a threshold-weighted false-positive penalty — across the full range of probability thresholds a clinician or analyst might use. The curve makes two comparisons explicit: treating nobody (net benefit = 0) and treating everybody (a flat line whose slope is determined by prevalence). A model is useful only in the threshold range where its curve sits above both baselines. In pathology and imaging contexts, DCA surfaces the practical question: at the probability cutoff your pipeline triggers an action, does it produce more right decisions than either extreme default?
Where Fovea applies it — When a digital-pathology or Cell Painting readout feeds a downstream triage or prioritization decision, DCA sits at the end of the validation-without-ground-truth checklist. It is particularly useful for uncertainty-drift-and-failure-modes work: a model whose calibration drifts under distribution shift may maintain AUC while losing net benefit entirely at the clinically relevant threshold band. DCA makes that collapse visible before the pipeline reaches a laboratory or clinical workflow.