---
license: cc-by-nc-nd-4.0
tags:
- medical-imaging
- ct
- vision
- vision-language
---

# FlexiCT

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

This family page links three child repos:

| Repo | Input | Output | Recommended use |
|---|---|---|---|
| `ricklisz/FlexiCT-2D` | `[B, 1, 512, 512]` CT slices | CLS and patch tokens | Slice-level feature extraction, classification, visualization, registration features |
| `ricklisz/FlexiCT-3D` | `[B, 1, 160, 160, 160]` CT volumes | CLS and patch tokens | Whole-volume feature extraction and downstream 3D workflows |
| `ricklisz/FlexiCT-3D-VLM` | CT volumes plus text | Image/text embeddings and similarity scores | Report-aligned retrieval and zero-shot text-image scoring |

## Preprocessing presets

`default` is recommended for whole-volume 3D and 3D-VLM inference. It orients/resamples path inputs to LPS at 2 mm spacing when spacing is available, clips HU to `[-1000, 1000]`, z-score normalizes, pads with the tensor minimum to at least `160^3`, then center crops to `160^3`. This best matches the released VLM evaluation path because it preserves physical scale better than globally resizing the anatomy.

`local_path` is a forgiving demo preset for arbitrary local CT files. It orients/resamples path inputs, clips, normalizes, pads to a cube, then trilinear-resizes to `160^3`. It is robust to heterogeneous scans but less faithful to the VLM inference scripts because it globally rescales anatomy.

`retrieval_roi` is retrieval-specific. It orients/resamples, clips, normalizes, crops an ROI cube from coordinates, a mask, or a bounding box, pads if the crop hits an image boundary, then resizes to `160^3`.

## License

The released checkpoints are made available under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. Code is MIT licensed in the source repository. Users must also comply with licenses and usage terms for the original datasets used to train the models.

## Citation

```bibtex
@misc{li2026universalctrepresentations,
  title = {Universal CT Representations from Anatomy to Disease Phenotype through Agglomerative Pretraining},
  author = {Yuheng Li and Yuan Gao and Haoyu Dong and Yuxiang Lai and Shansong Wang and Mojtaba Safari and James E. Baciak and Xiaofeng Yang},
  year = {2026},
  eprint = {2605.21906},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2605.21906},
  url = {https://arxiv.org/abs/2605.21906}
}
```

## Medical disclaimer

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