Where to Look Matters: Rethinking Sub-Volume Sampling in 3D Medical Self-supervised Learning
What it is
This paper argues that random sub-volume sampling is a bottleneck in 3D medical self-supervised learning because it ignores anatomical value. It proposes a plug-and-play framework that replaces random cropping with a guided probe bank based on anatomy-awareness, informativeness, and diversity to improve downstream transfer.
The problem it solves
Random sub-volume sampling in 3D medical self-supervised learning treats all spatial regions as equally valuable, ignoring anatomical priors and spatial redundancy, which limits representation learning quality.
What you could use it for
- Improving pretraining efficiency for 3D medical image models
- Enhancing downstream task performance in medical imaging
- Designing better observation policies for self-supervised learning
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.