ORCA-3DCT — precomputed 3D CT visual tokens for CT-RATE

📄 arXiv · 💻 GitHub · 🤗 Models & Data

Encoder outputs for the whole of CT-RATE, released so that work on 3D CT token compression, pooling, retrieval and probing does not have to start by spending several hundred GPU-hours re-encoding the corpus.

Two layers:

• ct_rate/uncompressed/ — the raw encoder grids, plus the pooled encoders' native output. Bring your own compressor.
• ct_rate/compressed/ — ORCA and grid-average token bundles at four budgets, the exact inputs behind ORCA.

---

What is here

Encoder	ORCA B=216	ORCA B=64	ORCA B=27	Grid avg B=216	Grid avg B=64	Grid avg B=27	Uncompressed	Organ segmentation
COLIPRI	216×792 8.79 GB	64×792 2.60 GB	27×792 1.10 GB	216×768 8.53 GB	64×768 2.53 GB	27×768 1.07 GB	24×24×24×768 545 GB	11×24×24×24 0.37 GB
CT-CLIP	216×536 13.0 GB	64×536 3.44 GB	27×536 1.45 GB	216×512 12.4 GB	64×512 3.29 GB	27×512 1.39 GB	24×24×24×512 710 GB	11×24×24×24 0.93 GB
ViSD-Boost + CT-CLIP	—	—	—	—	—	—	5×256 0.13 GB	—
ViSD-Boost	—	—	—	—	—	—	4×256 0.10 GB	—
FVLM	—	—	—	—	—	—	4×256 0.10 GB	—

Everything else:

path	contents	size
ct_rate/checkpoints/reportgen_{orcafull,avgpack}_b{27,216}__s2/step_*/	our four report-generation LoRA + projector checkpoints — stage 2, at the epoch with the highest clinical F1	11 GB

Coverage. CT-CLIP and the pooled encoders cover all of CT-RATE (47,149 train / 3,039 valid). COLIPRI covers 24,128 / 1,564, because its authors take one reconstruction per scan to be sufficient.

Only the arrays we computed are hosted here. The reports, the 18 abnormality labels and the report-generation question set are CT-RATE's own files — take them from CT-RATE (dataset/vqa/, dataset/radiology_text_reports/, dataset/multi_abnormality_labels/).

Loading

import numpy as np
from huggingface_hub import snapshot_download

d = snapshot_download("LiangRenjie/ORCA-3DCT", repo_type="dataset",
                      allow_patterns="ct_rate/compressed/colipri/colipri_ORCA_b216_d792_lam0p5/*")

tokens = np.load(f"{d}/ct_rate/compressed/colipri/colipri_ORCA_b216_d792_lam0p5/valid.npy", mmap_mode="r")  # (1564, 216, 792) fp16
ids    = open(f"{d}/ct_rate/compressed/colipri/colipri_ORCA_b216_d792_lam0p5/valid_ids.txt").read().split()
tokens[ids.index("valid_1000_a_2")]        # -> (216, 792)

Organ masks are .npz, one member per volume, so a single volume decompresses on its own:

z = np.load(f"{d}/ct_rate/organ_masks/colipri/valid.npz")
z["valid_1000_a_2"]        # -> (11, 24, 24, 24) fp16, soft occupancy in [0, 1]
z["_channel_names"]        # -> lung, airway, heart, aorta, ...

vid = "valid_1000_a_2"
study = vid.rsplit("_", 2)[0]                                      # -> valid_1000
g = np.load(f"{d}/ct_rate/uncompressed/colipri/valid/{study}/{vid}.npy")   # (768, 24, 24, 24) fp16
g = g.transpose(1, 2, 3, 0)                                          # -> (24, 24, 24, 768)

Running ORCA yourself

ct_rate/organ_masks/ holds, for every volume, an (11, *grid) float16 array giving the fraction of each token occupied by each of eleven thoracic structures — lung, airway, heart, aorta, central vessels, mediastinum, hiatus, pleura, chest-wall bone, thoracic spine, upper abdomen — as soft occupancy in [0, 1] rather than a hard label, already resampled onto the encoder's token grid. Producing it means running TotalSegmentator over 47k volumes and resampling the 347 GB of output; 1.1 GB here replaces that. Match the set to your encoder — encoders resample and crop the volume differently, so a mask built for one grid does not align with another. With the uncompressed grids plus these masks, the compressor in the repository reproduces every ct_rate/compressed/ bundle.

License and terms

CC-BY-NC-SA-4.0, inherited from CT-RATE. Academic and research use only; no commercial use; no attempt at re-identification. These arrays are encoder features derived from CT-RATE — if you use them, you are bound by CT-RATE's terms and should cite CT-RATE and the encoder whose outputs you used, alongside ORCA.

@article{liang2026orca,
  title={ORCA: ORgan-Centroid Aggregation for Training-Free 3D CT Visual Token Compression},
  author={Liang, Renjie and Xu, Zijian and Pan, Jinqian and Sun, Chengkun and Fan, Zhengkang and Li, Shawn and Qin, You and Liu, Mei and Xu, Jie},
  journal={arXiv preprint arXiv:2608.00345},
  year={2026}
}
@article{hamamci2024ctrate,
  title   = {Generalist Foundation Models from a Multimodal Dataset for 3D Computed Tomography},
  author  = {Hamamci, Ibrahim Ethem and others},
  journal = {Nature Biomedical Engineering},
  year    = {2026},
  doi     = {10.1038/s41551-025-01599-y}
}
@misc{wald2025colipri,
  title  = {Comprehensive Language--Image Pre-training for 3D Medical Image Understanding},
  author = {Wald, Tassilo and Hamamci, Ibrahim Ethem and Gao, Yuan and others},
  year   = {2025},
  eprint = {2510.15042},
  archivePrefix = {arXiv}
}