Abstract

Fluorescence super-resolution foundation models perform well within their pretraining distribution, but whether they adapt efficiently to an imaging technique outside it is largely unexamined. We study this using fiber optic microscopy with ultraviolet (UV) surface excitation (FUSE). Applying such a model directly to FUSE data performs worse than the low-resolution input itself, a negative transfer. Tuning only the output side on a single image, about 0.0005% of all parameters, reverses this and significantly surpasses the baseline, whereas making more parameters trainable degrades performance monotonically through single-image overfitting, revealing a budget sweet spot at the low end of the parameter spectrum. Contrasting input-side and output-side tuning, with a training-length sweep, localizes the adaptation bottleneck to the output side. The output-layer update is low-rank and intensity-remapping, and closed-form controls rule out a reduction to global intensity alignment, indicating a structured correction. These findings hold across spleen and kidney, suggesting that for efficient cross-condition adaptation, placing a tiny correction at the right location can matter more than freeing model capacity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_030.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: Not Available

BibTex

@InProceedings{DenLin_ATiny_MICCAISAT2026,
        author = { Deng, Lingchi AND Liang, Kaicheng},
        title = { { A Tiny Correction in the Right Place: Adapting a Fluorescence Super-Resolution Foundation Model } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17265},
        month = {pending},
        page = {pending}
}


back to top