Home / Projects / DALE-CT
Project Software Apache-2.0

DALE-CT

GitHub

Details

Architecture Python
License Apache-2.0

DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CT

Training code and model release for DALE-CT (Depth-Aware Latent-Euclidean
Computed Tomography) — a family of 2D slice-based Vision Transformers trained
entirely from scratch on chest CT with the heuristics-free
LeJEPA objective and depth-aware slab
sampling: self-supervised views are drawn from across a physical $z$-axis
slab rather than a single slice, so the frozen representations form an
anatomical world model — they linearly decode volumetric slice position
($R^2 approx 0.97$), recover slice ordering without labels, and localize
organs and findings, despite no 3D or positional supervision.

Paper: DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CT
· Benchmark: chest-ct-foundation-model-benchmark —
the single-protocol evaluation harness, patient-disjoint splits, and full
result tables for every model in the paper

Released models

All weights are on the Hugging Face Hub (CC-BY-NC-SA-4.0); the DALE-CT
variants load in one line via timm (Finetuned DINOv2 is a transformers
model — see its card):

import timm
model = timm.create_model("hf-hub:Kentucky-Open-Science/DALE-CT-0-L", pretrained=True)
Model CT-RATE Macro AUROC RAD-ChestCT AUROC (frozen / retrained probe) Role
DALE-CT-0-L ⭐ 0.8156 0.6281 / 0.7572 Recommended general-purpose backbone — best 2D external-transfer point estimates; supervision-free at ~287k-scan scale
DALE-CT-2S 0.8247 0.6252 / 0.7389 Best in-domain (CT-RATE)
DALE-CT-1S-v2 0.8098 0.6284 / 0.7334 Anatomical (TotalSegmentator) dense supervision only
DALE-CT-0 0.8057 0.5946 / 0.7477 Pure self-supervised, CT-RATE
Finetuned DINOv2 0.7953 0.6252 / 0.7550 Continual-pretraining baseline

The earlier DALE-CT-1S
release (patch-14, [CLS]-only anatomical supervision) remains available as
the backbone used by Ker-VLJEPA-3B;
DALE-CT-1S-v2 is the configuration benchmarked in the paper.

All numbers are our own head-to-head measurements: every model in the paper
(including public 3D baselines COLIPRI-CRM, Merlin, CT-FM, CT-CLIP) is probed
under one linear-probing MIL protocol on shared splits with bootstrap
confidence intervals. The protocol, splits, and result tables are maintained
in the chest-ct-foundation-model-benchmark
repository. Each model card documents the exact Hounsfield-Unit
preprocessing its backbone expects — DALE-CT-0-L uses different
normalization statistics than the CT-RATE-trained variants.

Repository map

Path Contents
train_lejepa.py Pre-training entry point (all DALE-CT variants)
lejepa_core/ LeJEPA architecture: SSL meta-arch, trainer, SIGReg
models/ ViT backbone, projector, MIL pooling heads
supervised_heads/ Dense auxiliary supervision (TotalSegmentator / ReXGroundingCT soft labels)
dataloaders/ CT-RATE WebDataset, multi-scale multi-crop, multi-source zarr, embedding loaders
configs/ OmegaConf YAML for every pretraining / probing / evaluation run
scripts/run_error_bars.py Head-to-head benchmark: probe selection, seed variance, bootstrap CIs
scripts/generate_benchmark_embeddings.py Embedding extraction for the public 3D baselines
scripts/run_model_comparison_probes.py Dense 2D probing (auxiliary-task tables)
scripts/exp_c_zposition.py, scripts/exp_c_worldmodel_probes.py Anatomical world-model probes ($z$-regression, slice ordering, organ identity)
train_gridsearch.py, eval_rad.py Linear-probe grid search; RAD-ChestCT transfer
train_e2e_lora.py LoRA fine-tuning

The evaluation harness above is kept as released with the paper; the
maintained, self-contained version — including the protocol documentation —
is the chest-ct-foundation-model-benchmark
repository.

Dataset preparation (CT-RATE → WebDataset shards, TotalSegmentator masks)
lives in a separate preprocessing pipeline that is not yet public.

Paths in configs refer to our cluster layout; point them at your own data
roots. Raw datasets and checkpoints are never stored in this repository.

Citation

@article{damron2026dalect,
  title   = {DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CT},
  author  = {Damron, Evan W. and Gokmen, Mahmut S. and Klusty, Mitchell A. and
             Leach, Caroline N. and Collier, Emily B. and Bumgardner, V. K. Cody},
  journal = {arXiv preprint arXiv:2606.07775},
  year    = {2026}
}

License

Code is released under Apache-2.0; model weights on the Hugging
Face Hub are CC-BY-NC-SA-4.0.

Imported from gh:Kentucky-Open-Science/DALE-CT. Source last updated 2026-08-26. Synced 2026-08-26.