CARE
CARE is a foundation model for three-dimensional computed tomography (CT), developed through large-scale vision–language pretraining for representation learning and downstream analysis across five anatomical regions: head, chest, abdomen, spine and joints.
Code Repository
The model architecture, CT preprocessing, and representative downstream training and evaluation code will be available in the CARE GitHub repository. An accompanying model-loading Quickstart notebook demonstrates image and text encoding with CARE.
Model Checkpoints and Data Packages
The primary pretrained checkpoint is ckpts/care.pth. The release is organized to support the following downstream examples, subject to the availability and redistribution terms of each resource:
| Task | Public dataset | Checkpoint |
|---|---|---|
| Abnormality classification (linear probing) | CT-RATE | ckpts/downstream/abnormality_cls_ln/ctrate/care_abnormality_cls_ln_ctrate.pt |
| ROI classification | PANORAMA | ckpts/downstream/roi_cls/panorama/care_roi_cls_panorama.pt |
| Segmentation | KiPA22 | ckpts/downstream/segmentation/kipa22/care_segmentation_kipa22.pt |
| Image–report retrieval | CTRG | Uses ckpts/care.pth; no additional task checkpoint |
Where available and authorized for distribution, associated feature or evaluation packages are placed under datasets/downstream/ in the corresponding task and dataset directories.
For example, the CT-RATE linear-probing package uses train.npz, val.npz, and test.npz.
Dataset-specific access requirements and licenses continue to apply; the CARE model license does not replace them.
Download published files from Files and versions. If gated access is enabled, request access through the Hugging Face repository page. Model and data files are not bundled with the GitHub source code.
Basic Usage
Install the dependencies from the GitHub repository, place care.pth under ckpts/, and run from the repository root:
import torch
from model_loading.models import get_model, get_transform
# Load the pretrained CARE vision–language model.
device = "cuda"
model = get_model(device, "ckpts/care.pth")
# Read and preprocess a raw-HU 3D NIfTI CT volume.
transform = get_transform()
image = transform("path/to/ct_volume.nii.gz").unsqueeze(0).to(device)
# image: [1, 1, 112, 336, 336]
with torch.inference_mode():
pooled, tokens = model.visual(image)
print(pooled.shape) # [1, 1024]
print(tokens.shape) # [1, 4096, 1024]
The example uses the shared raw-HU NIfTI preprocessing pipeline. Pre-windowed CT data and downstream datasets may require different task-specific preprocessing; follow the relevant GitHub implementation or Quickstart notebook.
Intended Use
CARE is intended for medical imaging research, including CT representation learning, volume- and region-level recognition, segmentation, and vision–language applications. Users are responsible for obtaining datasets independently when those datasets are not distributed here.
Limitations
CARE is a research model and is not intended for direct clinical diagnosis, treatment decisions, or patient management without appropriate clinical validation and regulatory approval. Performance and applicability may vary across patient populations, scanners, acquisition protocols, and clinical settings.
License
CARE model materials are released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0). Third-party code, datasets, and derived resources remain subject to their own applicable licenses and access conditions.