Rethinking Support Aggregation: Similarity-Based Weighting for In-context Medical Segmentation

Kuan-I Chung, Daniel Moyer

Keywords: Segmentation, In-context Learning

TL;DR: We assign image-level-similarity-based weights to context samples to improve in-context segmentation model's performance.

Registration Requirement: Yes

Abstract: In-context medical segmentation models such as UniverSeg condition prediction on a support set of labeled examples, typically treating those examples equally. However, the relevance of support example to a given query can vary significantly. We propose a simple similarity-based weighting strategy that reweights support example's contributions without changing underlying model architecture. Experiments on CTSpine1K and WBC demonstrate improvement over uniform-weighed baseline.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 50