VoxCare
Pretrained volumetric medical representations and 22 task-specific segmentation bundles.
GitHub: guangqianyang/VoxCare · Detailed model zoo · Reproducibility ledger
This repository is private. GitHub and Hugging Face permissions are granted separately.
Original-resolution Figure 1 · Figure PDF
Figure 1 is the manuscript overview of 42 task configurations across five families. The current downloadable release contains the backbone and 22 segmentation bundles. The 24,227-volume source pool differs from the 22,876 retained reference-training volumes, and radar axes use task-specific metrics rather than one shared scale. Original captions and extraction provenance.
Weights
| Bundle | Files | Use |
|---|---|---|
| M6-l100 encoder | 1 backbone checkpoint | 3D feature extraction / downstream initialization |
| Best-result downstream packages | 17 checkpoints + metadata/plans/labels/records | python -m voxcare infer |
| Archived task variants | 5 complete checkpoints + plans/labels | python -m voxcare legacy-infer |
All 23 checkpoints have been uploaded. Architectures, task-specific parameters, paths and SHA256 values are recorded in release_manifest.json. Download the entire chosen task folder; do not use a .pth file without its accompanying configuration.
Download and run
Install the code and verified environment from the GitHub README, then authenticate an account with access to this repository:
from huggingface_hub import login, snapshot_download
login()
snapshot_download(repo_id="Guangqian/VoxCare", local_dir="checkpoints",
allow_patterns=["pretrained/*", "downstream/Pancreas/**", "release_manifest.json"])
python -m voxcare infer --model-dir checkpoints/downstream/Pancreas \
--images /path/to/imagesTs --output outputs/pancreas
For one of the five legacy tasks, download downstream/legacy/<task>/** and pretrained/*, and use:
python -m voxcare legacy-infer --model-dir checkpoints/downstream/legacy/msd_heart \
--images /path/to/imagesTs --output outputs/msd_heart
The legacy entry point handles the archived NPU serialization on CPU without changing network tensor values, verifies the original checkpoint digest, and strictly loads the matching architecture on CUDA. No torch_npu installation is needed for these published inference commands. ToothFairy3 disables test-time mirroring.
Encoder features
import torch
from voxcare.backbone import load_backbone
model = load_backbone("checkpoints/pretrained/voxcare_m6_l100_backbone.pth", device="cuda")
with torch.inference_mode():
features = model.encode_spatial_features(torch.zeros(1, 1, 32, 64, 64, device="cuda"))
print(features.shape) # [1, 768, 4, 8, 8]
Results and validation
Original-resolution Figure 2 · Figure PDF
The 22-dataset table reports archived full-cohort local held-out results, not a newly completed full-volume replay from this renamed package. Figure 2 contains the manuscript's eight selected qualitative target-localisation examples, with shared slices and insets. Several historical recipes used test data for search/selection. Tasks use different downstream heads and training budgets.
Cross-modality CT Synthesis
Figure 3 compares brain and pelvis MRI-to-CT and CBCT-to-CT synthesis, using the same slices and enlarged regions across methods. High-resolution PNG · Figure PDF.
Brain MRI Registration
Figure 4 shows axial, coronal and sagittal views of the same OASIS3 moving–fixed pair. Dashed yellow contours mark the fixed lateral ventricle; cyan contours mark the moving or warped structure. High-resolution PNG · Figure PDF.
Figures 1–4 use the updated main_v2.pdf. Synthesis and registration are manuscript results; the current downloadable task bundles cover segmentation. Original captions and extraction details.
All 22 downstream architectures have passed strict loading and configured GPU patch forwards; the encoder feature example has passed as well. Reports are under code/reports. These checks do not establish an independent full-test reproduction or deterministic retraining. Exact historical pretraining launch/data provenance and complete renamed-package volume replay are still being audited.
No raw clinical images, labels or training caches are distributed. Source trees retain their upstream licenses; see upstream notices. This is a private reproducibility preview.



