Faster QA for radiology training labels

Surface likely-mislabelled studies in the annotation set so clinical reviewers spend their time where it counts.

The problem

Reviewing training labels for quality is slow and uniform — every study gets the same attention. We want a model-assisted triage that ranks studies by likelihood of a labelling error (model/label disagreement, outliers) so expert reviewers focus on the riskiest 10% first.

What roles would make a great team?

Teams run 3–6 people, including the lead.