3D medical self-supervised pretraining drops random cropping: an offline probe bank lifts organ occupancy from 19.79% to 45.71%

In 3D medical self-supervised pretraining, the step that draws a fixed-size sub-volume from each scan now samples from a precomputed bank of candidate centers instead of cropping at random. Such pipelines previously assumed all spatial regions were equally worth observing, so random cropping spent the fixed crop budget on boundary-dominated or structure-sparse regions. On the mask-available subset, random crops averaged 55.83% foreground voxels but only 19.79% overlap with annotated organ masks, meaning many observations landed on non-background but anatomy-agnostic regions.

The probe bank scores candidate centers by anatomy-awareness, informativeness, and diversity, is built offline, and during pretraining one center is drawn uniformly from it before cropping. The lightweight variant uses only geometric priors and image statistics, namely normalized distance to the volume center, normalized minimum distance to the boundary, local mean gradient magnitude, and local intensity entropy, with no learnable parameters. On VoCo-10k, this scoring raised foreground occupancy from 55.83% to 75.63% and organ occupancy from 19.79% to 45.71%.

With the probe bank in place, transfer improves across all three self-supervised backbones on segmentation, classification, and registration. Against each backbone's own random-cropping baseline, gains are largest under 1% and 10% label segmentation, while 100% label and registration gains are smaller but consistently positive. VoCo on AMOS at 1% labels rose from 55.81% to 58.64%, S2DC on AMOS at 1% labels from 55.19% to 58.83%, and MedGMAE on WORD at 1% labels from 68.09% to 70.67%.

The method targets the sub-volume sampling step of 3D medical self-supervised pretraining, for pipelines that keep a fixed crop budget and observe repeatedly through sub-volumes, and it is validated on the VoCo, S2DC, and MedGMAE backbones across segmentation, classification, and registration. The probe bank is built once offline and cached, taking 41.5 MB on VoCo-10k and 21.8 MB on AbdomenAtlasMini, with total pretraining overhead of 4.4%, 2.1%, and 2.2% over random sampling. The authors propose future work on jointly optimizing sub-volume selection with SSL pretraining and on more flexible priors for non-canonical acquisition settings.

The probe bank is built offline and stays static, so it cannot adapt its probing to the model's evolving learning dynamics during training. The authors also note that both quality axes, the number of organ classes and the variance of frozen features, are unavailable during pretraining, since the former needs organ labels and the latter an external feature extractor, so the bank can only approximate that quality axis from volume-intrinsic signals. Multi-seed repeats show larger variance in low-label segmentation and classification, while the 100% label and registration settings stay relatively stable.