Two specialists look at the same breast image. One is an experienced doctor, the other is an artificial intelligence program. Sometimes their opinions match, sometimes they diverge. How often they agree determines how accurately they can together distinguish healthy tissue from a tumor.
In the new study, scientists took a mathematical tool, copulas, and used it to analyze how the errors of these two "rowers" are related in two different tasks. The first task is to rule out disease: if both say "healthy," the patient leaves without unnecessary worry. The second is to confirm disease: if both say "sick," the patient is sent for treatment. It turned out that for ruling out disease, it is better when the doctor and AI agree more often on images with actual tumors. For confirming disease, on the other hand, it is more useful when their opinions diverge on images of healthy women.
This was tested on three mammography datasets where AI helped doctors reduce workload or improve accuracy. The results matched the theory. The practical takeaway is simple: when tuning an AI assistant, you need to decide in advance which task it is for, screening out healthy patients or catching the sick ones, and adjust the degree of "harmony" with the doctor accordingly.
Copulas are mathematical functions that describe how random variables are related. The name comes from the Latin copula—"bond, connection."