Abstract
Accurate multi-class segmentation of interstitial lung disease (ILD) findings remains challenging because radiological patterns often exhibit overlapping appearances, substantial variability, and frequent coexistence within the same examination. Consequently, expert annotations are difficult and costly to obtain, resulting in a scarcity of densely labeled datasets for supervised segmentation. Vision foundation models (VFMs) provide transferable representations that may reduce the dependence of segmentation models on large amounts of annotated data. In this work, we investigate whether VFM representations can alleviate the scarcity of dense pixel-level annotations for fine-grained multi-class ILD segmentation, while comparing the transferability of large-scale generic and CT-adapted pretraining. Contrary to expectations, DINOv3 outperformed its CT-adapted counterpart for most ILD findings. The resulting method surpassed nnU-Net in six of eight classes of findings and improved the macro-average Dice score by 19.97 percentage points. External evaluation on an independent HRCT cohort further showed that the learned predictions retain class-specific information beyond the training domain. Overall, our findings establish VFMs as a promising foundation for annotation-efficient ILD segmentation and indicate that generic large-scale pretraining may transfer more effectively to this task than CT-specific adaptation. Code is available at: https://github.com/carlasbmon/VFM_ILD_segmentation.git.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TIA_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{MonCar_Generic_MICCAISAT2026,
author = { Monteiro, Carla AND Santos, Mariana AND Melo, Tânia AND Mendonça, Ana Maria AND Ramos, José AND Pedrosa, João},
title = { { Generic or Domain-Specific? Foundation Models for Annotation-Efficient ILD Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17269},
month = {pending},
page = {pending}
}