Deciphering Metastasis with
Multimodal Artificial Intelligence Foundation Models
Research Project
DECIPHER-M
About DECIPHER-M
Metastasis is the cause of death for most cancer patients. However, the understanding of its underlying mechanisms is very incomplete.
DECIPHER-M addresses this with a unique approach using a new form of artificial intelligence (AI) called multimodal foundation models. In this project, these models are used to analyze a wide range of data, such as radiological images, pathological reports and genetic information of a patient together. This will answer fundamental questions about metastasis, such as the mechanisms of its occurrence, the potential to predict who might develop it and where, and what type of treatment might be most effective for different patients.
In addition, DECIPHER-M will provide practical tools that can be applied to individual patients to customize screening and treatment in cases at high risk of metastasis. Specifically, these tools aim to predict the most effective treatment for individual patients with metastatic disease so that these patients can be treated more effectively.
Facts and figures
Coordinator: TU Dresden
Number of Partners: 7
Start Date: March 1, 2025
End Date: February 29, 2028 (with potential extension)
Total Funding: around € 5.5 million (additional €3.5 million upon extension)
This project has received funding from the German Federal Ministry of Education under grant agreement No. 01KD2420A.
Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the BMBF and the granting authority can’t be held responsible for them.
Project Plan
The objectives of DECIPHER-M are addressed with a strategic work plan, which is divided into 9 subprojects (SP).
First, data will be collected in SP1. In SP2, this data will be used and basic machine learning models will be developed that can use a variety of medical inputs (images, text, tabular data, etc.). The code and methodology to fine-tune these foundation models based on real-world clinical datasets will be developed in SP3. This will enable the models to learn specific medical knowledge and gain a comprehensive understanding of the relationships between image and non-image data. This code base will then be used to develop specialized, fine-tuned models for specific use cases. In particular, the focus is on two clinical scenarios that are of great importance: Identification of cancer of unknown cause in SP4 and risk prediction and prognosis in SP5. These advanced models will be validated in SP6, which deals with rigorous model evaluation. Once the models have been validated in this way, the clinical impact will be investigated in real clinical simulation studies in SP7. Based on these findings, the application scenarios of the models in SP4 and SP5 will be adapted to the needs of patients. Building on the previous sub-projects, SP8 is dedicated to translation into mechanistic research to potentially uncover new molecular changes with potential implications for treatment. Throughout the project, there will be close cooperation with patients and the collaboration of the consortium will be organized in SP9.
