Domain-Constrained Distillation of DINOv3 into a Lightweight Foundation Model Toward Point-of-Care Ultrasound

Md Jaber Al Nahian, Shrimanti Ghosh, Jacob Jaremko, Abhilash Hareendranathan
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:3520-3541, 2026.

Abstract

Vision foundation models such as DINOv3 provide powerful representations but are too computationally demanding for point-of-care ultrasound (POCUS), whereas lightweight CNNs remain deployable yet brittle when faced with diverse anatomies and acquisition styles. We bridge this gap with a domain-constrained distillation framework that transfers DINOv3 ViT-B/16 knowledge into a compact ResNet-50, achieving roughly 3.4$\times$ compression while preserving the teacher’s billion-scale visual priors. Using a large, heterogeneous ultrasound corpus and physics-aware augmentations, the distilled model delivers substantial linear-probe improvements over standard CNN baselines and consistently outperforms the ViT teacher on challenging, heterogeneous datasets. It further offers marked gains in limited-label regimes, reflecting the realities of POCUS workflows where annotated data are scarce. Embedding visualizations show that the distilled encoder forms clearer, anatomy-aware clusters than the teacher, indicating successful alignment to ultrasound structure. Together, these results demonstrate that large-scale natural-image priors can be distilled into a lightweight, generalizable encoder suitable for resource-constrained clinical deployment.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-nahian26a, title = {Domain-Constrained Distillation of DINOv3 into a Lightweight Foundation Model Toward Point-of-Care Ultrasound}, author = {Nahian, Md Jaber Al and Ghosh, Shrimanti and Jaremko, Jacob and Hareendranathan, Abhilash}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {3520--3541}, year = {2026}, editor = {Huo, Yuankai and Gao, Mingchen and Kuo, Chang-Fu and Jin, Yueming and Deng, Ruining}, volume = {315}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v315/main/assets/nahian26a/nahian26a.pdf}, url = {https://proceedings.mlr.press/v315/nahian26a.html}, abstract = {Vision foundation models such as DINOv3 provide powerful representations but are too computationally demanding for point-of-care ultrasound (POCUS), whereas lightweight CNNs remain deployable yet brittle when faced with diverse anatomies and acquisition styles. We bridge this gap with a domain-constrained distillation framework that transfers DINOv3 ViT-B/16 knowledge into a compact ResNet-50, achieving roughly 3.4$\times$ compression while preserving the teacher’s billion-scale visual priors. Using a large, heterogeneous ultrasound corpus and physics-aware augmentations, the distilled model delivers substantial linear-probe improvements over standard CNN baselines and consistently outperforms the ViT teacher on challenging, heterogeneous datasets. It further offers marked gains in limited-label regimes, reflecting the realities of POCUS workflows where annotated data are scarce. Embedding visualizations show that the distilled encoder forms clearer, anatomy-aware clusters than the teacher, indicating successful alignment to ultrasound structure. Together, these results demonstrate that large-scale natural-image priors can be distilled into a lightweight, generalizable encoder suitable for resource-constrained clinical deployment.} }
Endnote
%0 Conference Paper %T Domain-Constrained Distillation of DINOv3 into a Lightweight Foundation Model Toward Point-of-Care Ultrasound %A Md Jaber Al Nahian %A Shrimanti Ghosh %A Jacob Jaremko %A Abhilash Hareendranathan %B Proceedings of The 9th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Yuankai Huo %E Mingchen Gao %E Chang-Fu Kuo %E Yueming Jin %E Ruining Deng %F pmlr-v315-nahian26a %I PMLR %P 3520--3541 %U https://proceedings.mlr.press/v315/nahian26a.html %V 315 %X Vision foundation models such as DINOv3 provide powerful representations but are too computationally demanding for point-of-care ultrasound (POCUS), whereas lightweight CNNs remain deployable yet brittle when faced with diverse anatomies and acquisition styles. We bridge this gap with a domain-constrained distillation framework that transfers DINOv3 ViT-B/16 knowledge into a compact ResNet-50, achieving roughly 3.4$\times$ compression while preserving the teacher’s billion-scale visual priors. Using a large, heterogeneous ultrasound corpus and physics-aware augmentations, the distilled model delivers substantial linear-probe improvements over standard CNN baselines and consistently outperforms the ViT teacher on challenging, heterogeneous datasets. It further offers marked gains in limited-label regimes, reflecting the realities of POCUS workflows where annotated data are scarce. Embedding visualizations show that the distilled encoder forms clearer, anatomy-aware clusters than the teacher, indicating successful alignment to ultrasound structure. Together, these results demonstrate that large-scale natural-image priors can be distilled into a lightweight, generalizable encoder suitable for resource-constrained clinical deployment.
APA
Nahian, M.J.A., Ghosh, S., Jaremko, J. & Hareendranathan, A.. (2026). Domain-Constrained Distillation of DINOv3 into a Lightweight Foundation Model Toward Point-of-Care Ultrasound. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:3520-3541 Available from https://proceedings.mlr.press/v315/nahian26a.html.

Related Material