Abstract
Precise instrument segmentation underpins advanced surgical navigation. Nevertheless, deploying a pre-trained natural image vision backbone for this task typically leads to degraded segmentation owing to severe cross-domain shifts. While adopting a self-supervised vision back bone pre-trained on endoscopic data can alleviate this issue, effectively adapting such a model to dense prediction tasks remains a formidable challenge. Visual prompt tuning is widely celebrated for cross-domain
transfer learning. However, its potential to facilitate intra-domain task generalization, specifically shifting a model from general scene understanding to precise semantic delineation, is largely unexplored. Motivated by this, we introduce a parameter-efficient framework that integrates an endoscopic-native Vision Transformer with a novel Gated Visual Prompt Tuning (GVPT) mechanism. GVPT dynamically regulates the injection of task-specific cues, allowing the model to gradually incorporate segmen
tation information while reducing perturbations to the pretrained surgical representation. Experiments on the EndoVis2018 dataset show that our method achieves 78.85% IoU, 60.17% mcIoU, and 82.56% mIoU with only 4.55M trainable parameters, highlighting the potential of prompt tuning for in-domain task adaptation in surgical image analysis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COLAS_023.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=WFNqRB8i1I
BibTex
@InProceedings{ZhuXug_Layerwise_MICCAISAT2026,
author = { Zhu, Xuguang AND Yu, Xiankang AND Wang, Changfei AND Zhang, Zhaochen AND Hayashi, Yuichiro AND Oda, Masahiro AND Mori, Kensaku},
title = { { Layer-wise Gated Adaptation of SAM for Endoscopic Surgical Instrument Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS pending},
month = {pending},
page = {pending}
}