Problem, challenge, and solution to old school "AI Labeling"!!

Labeling was used extensively in building AI/ML algorithms. The idea is to apply the minds and knowledge of professionals to tell the AI/ML models to learn from the experts. Usually, in medical AI, the labeling identifies and marks the "Area of Interest" (AoI) for AI/ML as "positive". Army of pathologists have dedicated their time and efforts to this labor-intensive work, and we really appreciate it.

However, there is a problem: what if the professionals disagree with the EXACT AoIs?? We asked 4 cytologists, all with 10+ years of experience in clinical routine, to label cells to be likely "abnormal" (examples below). See the differences? How could an accurate AI/ML model be built with such inconsistency? Well, the current solution is to introduce MORE cytologists (as referees) to make judgments, but only to make the whole thing out of control.

The Solution: We are leading a team of #cytologists and #AI #ML computer experts to solve this problem using "#labelfree" #unsupervisedlearning. Would love to reveal our outcome in the near future. But I can share with you: the key to a successful label-free training is #HighMagnificationCytology (#HMC) images or otherwise garbage in garbage out #GIGO