VolumeProbe: guided sub-volume sampling for 3D medical self-supervised learning

Replacing uniform random cropping with coarse-to-fine anatomical probes eliminates observation waste and matches full-training transfer at 48% of SSL epochs.

Primary source: arXiv:2610.13915: https://arxiv.org/abs/2610.13915

What’s new: Self-supervised pretraining for 3D medical images typically relies on uniform random sub-volume cropping, implicitly treating all spatial regions as equally informative. In Where to Look Matters: Rethinking Sub-Volume Sampling in 3D Medical Self-supervised Learning (arXiv:2610.13915, 9 Oct 2026), the authors document that random crops contain 55.8% foreground tissue on average but overlap annotated organ structures only 19.8% of the time, resulting in extensive observation waste. By showing that downstream segmentation Dice correlates strongly with crop semantic density (Spearman $\rho = 0.913$), they reframe sub-volume cropping as an active where-to-look observation policy.

The authors propose VolumeProbe, a plug-and-play offline sampling framework that builds a coarse-to-fine candidate probe bank ranked by three principles: anatomy-awareness, local informativeness (intensity gradient and entropy), and spatial diversity. A lightweight variant, VolumeProbe-Lite, computes these scores entirely from geometric priors and image statistics without neural inference. In pretraining experiments across multiple 3D SSL backbones (including VoCo and SwinUNETR), VolumeProbe-Lite achieves downstream transfer parity with standard random cropping using only 47.9% of the pretraining budget, while consistently improving final segmentation Dice across CT and MRI benchmarks.

Why it matters: 3D medical foundation models require substantial compute to pretrain due to volumetric voxel dimensions. Establishing that the observation policy itself is a primary performance bottleneck allows institutions to significantly accelerate self-supervised training or improve representation richness without modifying loss functions or network architectures.

Caveats: VolumeProbe constructs its probe bank offline prior to training, meaning crop selection cannot dynamically adapt to changing model errors or latent uncertainties as representations mature. The lightweight heuristic variant relies on standard patient orientation and canonical scanning geometries, which may require retuning for non-standard clinical acquisitions or off-axis views.