Federated Learning for Multi-Center Imaging Studies (PPMI/ADNI)

Ideas and Inspiration
, , ,
1

In large-scale neuroimaging initiatives like PPMI and ADNI, data is collected across multiple sites, but direct pooling is often limited, or entirely hindered, by local IRB restrictions, regulatory constraints, and site-specific data sharing policies. Additionally, many sites lack the computing infrastructure or technical expertise to host or train centralized models on large-scale MRI/PET datasets.

Federated learning (FL) offers a powerful alternative: instead of moving sensitive imaging and clinical data, the model is sent to the data. Each site runs the analysis locally and returns model updates (not raw data) to a central server for aggregation and downstream analysis, ensuring that no raw data ever leaves the site.

This approach enables:

  • Cross-center collaboration without violating data governance or patient privacy laws
  • Scalable model training, even with limited local computational resources
  • Robustness to scanner and protocol variability across sites
  • Discovery of generalizable neuroimaging biomarkers for PD, AD, and other neurodegenerative diseases

Toolkits like Flower, Fed-BioMed, and MONAI FL provide robust frameworks to deploy FL in realistic, multi-site research settings not just in neuroscience and biomedicine, but across other domains where privacy and decentralization are key.

2

This is really interesting! Thanks for sharing.

Would this approach also be able to be applied to other types of data? Thinking of wearables or larger -omics datasets. @luc_evers do you use this for wearables already? Just curious.

3

Thanks for sharing these tools! I don’t have hands-on experience with federated learning, but it is definitely a relevant approach for wearable data as well as the volume of sensor data collected in different studies is growing rapidly. Especially for more privacy-sensitive sensor data (e.g. GPS).

4

Yes it seems like it could have a lot of applications with privacy-sensitive patient data across multiple data types. Neat!