Self-supervision drives representational convergence in medical foundation models more than clinical supervision

In 30 Sekunden

Was passiert ist

Self-supervised pretraining drives convergence in medical image encoders more than clinical supervision, across 18 image and 7 text models tested.

Warum es zählt

Engineers building interoperable medical imaging systems should know that shared representations come from training objectives, not model scale or clinical labels.

Achtung

Convergence is modest, within-modality only, does not match radiologist judgment, and varies across patient subgroups and hospitals.