Abstract
Cross-domain few-shot medical image segmentation aims to segment unseen target-domain structures from query images using only a few annotated support examples from the target domain. However, it remains challenging due to large variations in appearance across imaging modalities and anatomical structures. In this work, we propose HyProDINO, a support-conditioned framework for cross-domain medical image segmentation. The proposed model combines a trainable ResNet34-Transformer encoder with a frozen DINOv2 semantic encoder to jointly learn medical image-specific representations and transferable semantic features. Support-derived foreground and background prototypes are further refined through query-aware prototype calibration and support-conditioned query adaptation, followed by complementary prototype guided attention and dense support-query relation modeling for robust mask prediction. To further improve robustness against unreliable support information, we introduce a support-suppression fine-tuning strategy that discourages foreground hallucination under invalid support conditions. For evaluation, the proposed method is trained on the TN3K dataset and tested on seven unseen target datasets under 1-shot and 5-shot settings. Results demonstrate consistent cross-domain generalization across diverse medical imaging domains, achieving 5-shot Dice scores of 80.06%, 78.36%, and 75.75% on ISIC2018, BUS-BRA, and DDTI, respectively. Ablation studies further verify the frozen DINOv2 branch substantially improves generalization on visually distant domains, especially Chest X-ray and ISIC2018. Code will be available at https://github.com/Tasnim-Jahan/HyProDINO.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_054.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=x4TTLkLmsQ
BibTex
@InProceedings{JahTas_HyProDINO_MICCAISAT2026,
author = { Jahan, Tasnim AND King, Andrew P. AND Shatabda, Swakkhar},
title = { { HyProDINO: A Hybrid Prototype Guided DINO Framework for Cross-Domain Few-Shot Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}