3D medical self-supervised pretraining drops random cropping: an offline probe bank lifts AMOS low-label Dice from 55.81 to 58.64
Synopsis
3D medical self-supervised pretraining has long defaulted to random fixed-size sub-volumes from each scan, implicitly treating all spatial regions as equally worth observing. VolumeProbe replaces that step with an offline probe bank: candidate crop centers are scored for anatomical plausibility, informativeness and diversity, then sampled uniformly from the cache during pretraining. Across the VoCo, S2DC and MedGMAE backbones and segmentation, classification and registration tasks, swapping random cropping for this bank improves transfer consistently, most clearly in low-label segmentation, where VoCo on AMOS rises from 55.81 to 58.64 Dice at 1% labels.
Interpretation
In 3D medical self-supervised pretraining, where a sub-volume is cropped becomes a designable observation policy rather than a preprocessing detail. The prior default drew fixed-size sub-volumes at random from each scan, implicitly assuming all spatial regions are equally valuable for representation learning. In a controlled study on the VoCo backbone, candidate centers were ranked by the number of distinct organ classes per crop and formed into low, middle and high static banks; average Dice on BTCV and AMOS rose monotonically from low to high, with random sampling in between, and a feature-variance proxy from frozen MedicalNet features stayed positively correlated with downstream Dice at 10% labels.
The method builds a coarse-to-fine probe bank offline and draws crop centers only from that cache during pretraining. The bank scores candidates on anatomy-awareness, informativeness and diversity, refines only selected high-value coarse bins into fine centers, and uses an axis-wise redundancy penalty so centers do not crowd together. The lightweight variant uses only geometric priors and image statistics, while the feature variant uses frozen 3DINO-ViT features; both share the same candidate generation and bank-selection pipeline, so the comparison isolates the scoring source.
Replacing random cropping improves all three self-supervised backbones on segmentation, classification and registration. Pretraining protocols, crop size, number of sampled sub-volumes, backbone and downstream pipelines stay unchanged; only the source of crop centers is swapped. VoCo on AMOS rises from 55.81 to 58.64 Dice at 1% labels and from 73.34 to 74.95 at 10%, S2DC on AMOS from 55.19 to 58.83 at 1%, and MedGMAE on WORD from 68.09 to 70.67 at 1%; registration gains on IXI and OASIS are smaller but consistently positive.
Perspective
The method targets 3D medical self-supervised pretraining that uses fixed-size sub-volume cropping, validated on the VoCo, S2DC and MedGMAE backbones across segmentation, classification and registration, with pretraining on VoCo-10k and AbdomenAtlasMini 1.0. The probe bank is built once offline and cached, about 41.5 MB for VoCo-10k and 21.8 MB for AbdomenAtlasMini, with total pretraining time rising 4.4%, 2.1% and 2.2% over random sampling. The authors propose future work on jointly optimizing sub-volume selection with 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 lightweight-versus-feature comparison shows frozen semantic features do not reliably beat geometric and statistical cues alone, so the feature variant is positioned as a diagnostic. Multi-run variance is larger in low-label segmentation and classification, while the 100% label and registration settings remain relatively stable.
