The thing is though, "move fast, break things" is a fundamentally unscientific ideological call to arms. Unsubstantiated white papers and preprints are the same buying of influence, only in scientific currency. This is wholly within character for guys who view Dale Carnegie's book as foundational.
!["Graph showing performance at a case (N = 115,973 participants) and breast (N = 231,268 breasts) level, across all five sites, using a ground truth of screen-detected cancers only (reflecting a common metric in the literature). This metric has a key limitation in that it cannot demonstrate incremental benefit of AI, as it does not include future cancers that developed over the subsequent 3 years. Readers naturally perform well using this metric, due to the gatekeeper effect - that is only cancers that were picked up by the historical double-reader process are considered as cancers, while incremental cancers that the AI system detects are penalized as false positives, rather than rewarded as a true positive. c) zoomed region of b)." [My inclusion]: AI performed much worse at false posiitives. Super weird and inappropriate to be arguing justifications in a figure legend. Also bullshit, as they are trying to argue pre-cancerous lesions the AI detected, or any patient who developed cancer in the next 3-years, knowing full well they don't have the data that says that same lesion AI detected turned into the eventual cancer. Also, pre-cancerous lesions are generally not treated, they are monitored and treated if they show growth over the years. So the AI is adding an inconsequential benefit?](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:ia6zvsuapxbocge7bb35iqzz/bafkreie63jvy4qkpvzmmlytejyhiehenyyv2hhbwp57m5it3nuwxj4nieq)
