Unified Supervision for Vision-Language modeling in 3D computed tomography

Official Code Release for ICCV 2025 3DVLM Workshop Paper

Title: Unified Supervision for Vision-Language modeling in 3D computed tomography
Conference: ICCV 2025, Vision-Language Modeling in 3D Medical Imaging (VLM3D) Workshop

<p align="left"> <img src="assets/logo.png" alt="Model Logo" height="120"/> </p>

---

Overview

Uniferum is a volumetric vision-language model designed for radiology. Uniferum integrates classification labels and segmentation masks into a single unified training framework.

• Harmonizes classification and segmentation across multiple CT datasets.
• Improves State-of-the-Art Results on the CT-RATE benchmark by +7% compared to CLIP-based models.
• Robust out-of-distribution performance
• zero-shot capabilities on RAD-CHEST and INSPECT datasets.

---

Citation

If you find this code useful for your research, please consider citing our work:

@inproceedings{iccv2025uniferum,
  title={Unified Supervision for Vision-Language modeling in 3D computed tomography},
  author={Hao-Chih Lee, Zelong Liu, Hamza Ahmed, Spencer Kim, Sean Huver,
Vishwesh Nath, Zahi A. Fayad, Timothy Deyer, Xueyan Mei},
  booktitle={ICCV VLM3D Workshop},
  year={2025}
}

<p align="left"> <img src="assets/BMEII_logo.png" alt="BMEII" height="120"/> </p>