Abstract
Foundation models promise to unify multiple clinical tasks within a single framework, but recent ultrasound studies report that unified models can underperform task-specific baselines. We hypothesize that this degradation arises not from model capacity limitations, but from task aggregation strategies that ignore interactions between task heterogeneity and available training data scale. In this work, we systematically investigate how DINOv3 behaves under different multi-task learning paradigms for diverse ultrasound tasks, providing practical insights into task aggregation for unified ultrasound foundation models. We conduct a comprehensive empirical study using a pretrained DINOv3 backbone within a unified multi-organ, multi-task (M2) evaluation framework, termed M2DINO. We evaluate 26 ultrasound tasks spanning segmentation, classification, detection, and regression under three paradigms: task-specific (TS), clinically-grouped (CG), and all-task unified (AU) training. Our results show that the effectiveness of task aggregation is highly task-dependent. Unified training (AU/CG) consistently benefits dense prediction tasks, whereas classification and regression derive more limited gains. Embedding visualizations further provide qualitative insights into the representation behavior of DINOv3 under different multi-task learning paradigms. These findings provide new insights into foundation model adaptation for heterogeneous ultrasound tasks and offer practical guidance for future unified ultrasound foundation models. Our codes are available at: \href{https://github.com/13204942/Multitask_Foundation_model_for_ultrasound}{GitHub}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_045.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ASMUS_045_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=q2R4cdVRQu
BibTex
@InProceedings{WanFan_From_MICCAISAT2026,
author = { Wang, Fangyijie AND Akumu, Tanya AND Dang, Vien Ngoc AND Jiménez-Sánchez, Amelia AND Bai, Jieyun AND Silvestre, Guénolé C. M. AND Lekadir, Karim AND Curran, Kathleen M.},
title = { { From Pretraining to Clinical Tasks: Transfer of Foundation Models in Ultrasound Multi-Task Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}