Announcement

📢 June 3, 2026: FlexiCT ranked 2nd on the CVPR 2026 Foundation Models for General CT Image Diagnosis Challenge.

• Codabench leaderboard under name ricklisz

📢 June 1, 2026: The FlexiCT pretraining code is now released on the pretraining branch.

FlexiCT

FlexiCT is a CT foundation model family trained through agglomerative continual pretraining, progressing from 2D slice-level anatomy to 3D volumetric reasoning and report-aligned vision-language understanding.

The released model family includes FlexiCT-2D, FlexiCT-3D, and FlexiCT-3D-VLM. The models were trained on 266,227 CT volumes from 56 public datasets and evaluated across segmentation, registration, classification, clinical retrieval, and vision-language tasks.

Demos and notebooks

Notebook	What it shows
inference_demo.ipynb	Minimal inference for FlexiCT-2D, FlexiCT-3D, and FlexiCT-3D-VLM, including synthetic smoke inputs and optional user CT paths.
visualization.ipynb	Feature visualization, PCA maps, and similarity maps using the sample CT assets in assets/.

Launch either notebook from the repository root:

jupyter lab inference_demo.ipynb
jupyter lab visualization.ipynb

Checkpoints

Download the checkpoint you need and pass its local path explicitly or through the environment variables documented below.

Model	Model size	PyTorch checkpoint
FlexiCT-2D	144M parameters	Download
FlexiCT-3D	144M parameters	Download
FlexiCT-3D-VLM	741M parameters	Download

Local installation

Use Python 3.11 and run commands from the repository root.

conda create -n flexict python=3.11 -y
conda activate flexict
cd /path/to/FlexiCT
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Using checkpoints

Pass a checkpoint path when constructing a model:

from flexi_ct import Flexi_CT_3D

model = Flexi_CT_3D(checkpoint_path="/path/to/ct_3d_teacher.pth")

You can also set environment variables once per shell session:

export FLEXICT_2D_CHECKPOINT=/path/to/ct_2d_teacher.pth
export FLEXICT_3D_CHECKPOINT=/path/to/ct_3d_teacher.pth
export FLEXICT_VLM_CHECKPOINT=/path/to/ct_3d_vlm.pth

Flexi_CT_VLM uses the Qwen3 embedding text tower, so its tokenizer and config files are loaded from Hugging Face unless already cached. For offline or shared systems, point HF_HOME to the cache location:

export HF_HOME=/path/to/huggingface_cache

Quick validation

After installation, run a quick import check:

python -c "import torch; from flexi_ct import Flexi_CT_2D, Flexi_CT_3D, Flexi_CT_VLM; print(torch.cuda.is_available(), 'OK')"

Constructing a model loads the selected checkpoint.

Downstream workflows

Downstream examples are provided under downstream/ for segmentation, registration, retrieval, classification, and VLM evaluation.

For segmentation experiments, see downstream/segmentation/README.md for the nnU-Net setup and dataset directory conventions.

For retrieval reproduction commands, see downstream/retrieval/README.md.

Citing this work

Please cite the arXiv preprint:

@article{li2026universal,
  title={Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining},
  author={Li, Yuheng and Gao, Yuan and Dong, Haoyu and Lai, Yuxiang and Wang, Shansong and Safari, Mojtaba and Baciak, James E and Yang, Xiaofeng},
  journal={arXiv preprint arXiv:2605.21906},
  year={2026}
}

License and disclaimer

Code in this repository is released under the MIT License.

The released checkpoints are made available under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. The pretrained weights may also be subject to the licenses and usage terms of the original datasets used for training. Users intending to use FlexiCT in commercial settings should verify dataset and model licensing and obtain any required permissions.

FlexiCT is provided for research use. It is not a medical device and is not a substitute for professional medical judgment.