Persistence of self-supervised backbone advantages for bruise segmentation

Determine whether the advantages of DINOv3 and LingBot-Vision self-supervised visual representations persist for diffuse, low-contrast targets such as bruises.

Background

The paper discusses DINOv3 and LingBot-Vision as self-supervised vision backbones that produce transferable visual representations for segmentation. Their reported benchmark settings primarily involve natural images with relatively well-defined object boundaries.

Bruises differ substantially from those benchmark targets because they often exhibit diffuse, low-contrast, and gradually blended boundaries. The paper explicitly identifies whether the backbones’ reported advantages transfer to this more ambiguous visual domain as an unresolved issue.

References

Both are benchmarked on natural images with well-defined object boundaries, and whether their advantages persist for diffuse, low-contrast targets such as bruises remains untested.

— BruNet: A Cross-Domain Transfer Framework for Bruise Segmentation  (2609.11463 - Wang et al., 10 Sep 2026) in Section 2.1, “Vision Transformer and Its Medical Uses”