Incentivizing DINOv3 Adaptation for Medical Vision Tasks via Feature Disentanglement

Published in Medical Imaging with Deep Learning (MIDL), 2026

The paper proposes DINOv3-FD, a feature disentanglement framework for adapting DINOv3 to medical vision tasks. It separates task-relevant and task-irrelevant features with a dual-stream adapter and orthogonality/distributional regularization to improve parameter-efficient fine-tuning performance.

Recommended citation: Zhicheng He, Yibing Fu, Yueming Jin. (2026). "Incentivizing DINOv3 Adaptation for Medical Vision Tasks via Feature Disentanglement." Medical Imaging with Deep Learning.
Download Paper