Project state
closed
Project start
September 2024
Funding duration
12 months
Universities involved
UZH
Practice partners
Universitätsspital Zürich
Funding amount DIZH
CHF 75'000
TRUST-RAD develops AI tools for radiology that are not only powerful, but also reliable and transparent. A key achievement is RadVLM, a new AI model for chest X-ray analysis. It can explain findings, draft reports, answer clinical questions, and highlight relevant image regions, making AI support more understandable and interactive for radiologists. The project also released a large, publicly available dataset with over 1.1 million image, instruction pairs, enabling researchers worldwide to build more trustworthy medical AI systems. The team’s arXiv publication has already received 11 citations, was referenced in the Gemini technical report, and another one is presented at the ICML 2025 Workshop. Another major milestone is securing a 500,000 GPU-hour compute grant from the Swiss AI Initiative. This supports the development of a “3D Vision-Language Model for Radiology,” expanding AI assistance from 2D X-rays to 3D imaging such as CT scans. TRUST-RAD is funded through the third Rapid Action Call of DIZH, helping ensure that AI in radiology becomes safer, more transparent, and more clinically useful.
Scientific Publications & Presentations
- arXiv publication: The team published the RadVLM research paper on arXiv, which has already received citations and was referenced in the Gemini technical report.
- ICML 2025 Workshop presentation: The project’s work was accepted and presented at the ICML 2025 Workshop on Multimodal Foundation Models and Large Language Models for Life Sciences.
Public Releases (Open Science Outputs, available in https://physionet.org/)
- RadVLM Instruction Dataset: A large, publicly available dataset with over 1.1 million image–instruction pairs were released to support the development of reliable medical AI.
- RadVLM Model: The trained model and resources were made publicly accessible to the research community.
Team
Dr. Farhad Nooralazadeh, UZH Department of Quantitative Biomedicine
Dr. Nicolas Deperrois, UZH Department of Quantitative Biomedicine
Prof. Dr. Michael Krauthammer, UZH Department of Quantitative Biomedicine
Practice partner
Dr. med. Christian Blüthgen, Universitätsspital Zürich, Institute for Diagnostic and Interventional Radiology
Prof. Dr. med. Thomas Frauenfelder, Universitätsspital Zürich, Institute for Diagnostic and Interventional Radiology
Call type: 3. Rapid Action Call “Digital resilience: between deep fake and cyber creativity“