Contrast-Gated Parameter-Efficient Adaptation of MedSAM for Skin-Tone-Robust Lesion Segmentation

Overview

CG-MedSAM investigates whether contrast-conditioned adaptation can support skin-lesion segmentation across skin tones and variations in bounding-box prompts. The author manuscript is available as a Zenodo preprint.

Research question

How does contrast-conditioned parameter-efficient adaptation affect segmentation quality across skin-tone subgroups and imperfect bounding-box prompts?

Methods

Residual bottleneck adapters are modulated by a contrast signal computed within the prompt region. The manuscript examines segmentation quality, prompt perturbations, subgroup performance, and contrast-conditioning ablations.

Data & evaluation setup

ISIC 2018 and sDDI. Dataset access and usage remain subject to their original terms; the preprint deposit does not include datasets or model weights.

Limitations

This is an author preprint, not a peer-reviewed or accepted journal article. Interpret findings, subgroup comparisons, and generalization using the manuscript and repository documentation; clinical readiness is not established.

Collaborators & acknowledgements

The work builds on MedSAM and the original ISIC 2018 and sDDI datasets. Their creators retain credit and their original licensing terms.