Brain scans now process in seconds instead of hours — but the speed comes with a hidden cost

A new AI method can analyze whole-brain metabolite maps from MRI scans in seconds instead of hours, matching the accuracy of the current gold-standard method. The catch: the AI's results shift by up to 30% depending on how you configure the baseline correction, a setting that humans have to choose but the AI can't explain.

Speed matters in clinical imaging — faster processing means faster diagnosis, more scans per day, lower costs. But this paper reveals a structural problem: the AI is fast because it's flexible, but that flexibility hides a choice that used to be visible. When a human ran the old method, they had to decide on baseline correction settings and live with the consequences. Now the AI adapts to whatever settings you give it, which sounds good until you realize that means the same brain scan can produce wildly different metabolite maps depending on which settings someone chose without thinking about it. The 30% variance isn't a measurement error — it's a configuration choice that now lives inside a black box.