FedPref: Federated Preference Learning for Structured Radiology Report Extraction

Flint Xiaofeng Fan, Cheston Tan, Yew-Soon Ong, Roger Wattenhofer

ELAMI @ MICCAI 2026 · Oral

Radiology reports express findings and locations in free text, while downstream analysis needs a fixed schema. FedPref uses frozen public language models to propose alternative JSON extractions and local annotations to rank them. Sites train compact adapters on their own preference pairs and collaborate through pair-count-weighted aggregation, without pooling reports or annotations.

Centralized pooling versus FedPref: reports, annotations and preference pairs stay at their sites; adapter states are exchanged.
Centralized pooling versus FedPref: reports, annotations and preference pairs stay at their sites; adapter states are exchanged.

← All publications

➤