PROJECT / MICCAI 2026LOTFI ABDELKRIM MECHARBAT & COLLABORATORS

MetaVFM

Meta-Learned Efficient Adaptation of Vision Foundation Models for Medical Imaging

Use meta-learning to adapt foundation models to new medical imaging tasks with adapters and LoRA.

Read the paper ↗View method overview ↓Poster (PowerPoint) ↓

MetaVFM research illustration — Meta-space analysis from the MICCAI 2026 poster

METHOD & EXPERIMENTS

A shared-weight Super-Adapter supports an adapter zoo built across 90+ datasets. The learned dataset–adapter meta-space retrieves configurations for unseen tasks; the framework is evaluated on 17 datasets.

2–3×faster convergence
under a fixed epoch budget

Reported in the manuscript; see the paper for evaluation settings.

AN INTERACTIVE EXPLANATION

Method overview

METAVFM / PAPER DATA

Adapter placement and rank

Adapter placement and rank interact. Explore the validation accuracy reported in Figure 1; the best configuration changes with the dataset.

Choose a dataset

MetaVFM learns a dataset–adapter meta-space to retrieve placement, rank, and initialization together.

Full results & uncertainty ↗

Adapter configuration sensitivityVALIDATION ACCURACY (%)

Lower within this datasetHigher

Selected columns from MetaVFM, Fig. 1. Cells show mean ± standard deviation over seeds. Gold outline marks the best displayed mean. Q/V: query/value; P: output projection; F1/F2: MLP projections; All: all six module types. This sensitivity study does not show MetaVFM’s final predictions.

MICCAI 2026