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

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.