Acquisition Shift Explorer

Change slice thickness, reconstruction kernel and noise, then watch a model trained on one protocol degrade on another.

domain shift
external validation
protocol
generalization

Internal test AUC

0.660

External AUC

0.697

Generalization gap

-0.037

modest

ROC: same model, two acquisition protocols

Internal (thin slice)External protocol

Texture feature distribution by site

Development siteExternal site

Reading the gap

With concept shift at zero the entire loss is covariate shift: the tumours are unchanged, only their measurements moved. Harmonization recovers much of that. Any gap that persists once harmonization is on, and all of the loss once concept shift is non-zero, cannot be fixed by rescaling features.

Preparing the simulation…