Samsung teams with DeSci startup Galeon to train ultrasound AI without pooling patient data

Samsung brings decentralized AI training to ultrasound imaging

Samsung is integrating its ultrasound devices with Galeon's electronic health record (EHR) platform to fuel AI model training while keeping patient data private. The EHR is already live across 18 interconnected hospitals, including Rouen University Hospital, Caen University Hospital, Toulon Hospital and Sud Francilien Hospital in France.

The approach separates data from algorithms. Patient data stays within each institution. The AI algorithm is coordinated onchain for auditability, while all training data is anonymized before use.

How the integration works

Hospitals contribute anonymized ultrasound-derived data to a shared training process without centralizing PHI. Think of it as federated learning across sites: models learn from distributed datasets, and participating centers retain control of their records.

Blockchain is used for traceability of algorithm actions and versioning, not for storing medical data. This helps with provenance, reproducibility and compliance reviews.

What this means for hospitals

• Faster model improvement with multi-center diversity, without shipping raw data.
• Local control over data governance, retention and consent.
• Transparent audit trails for algorithm updates and performance shifts.
• Potential for better ultrasound triage, image quality aids and streamlined reporting.

Early use cases from Galeon

Galeon reports progress on practical tools: an automatic billing assistant for medical services, an AI that summarizes consultations and an in-progress speech-to-text system. These focus on administrative and documentation workloads that slow teams down.

Decentralized science gains traction in healthcare

DeSci initiatives are accelerating in medicine as researchers and communities look for faster, lower-cost pathways. Community groups like VitaDAO have argued that traditional drug development timelines push people to experiment with new funding and data-sharing models.

Some projects are ambitious. HydraDAO has claimed a study where rats with fully transected spinal cords walked again within five days. Claims like these draw attention-and scrutiny-from clinicians and investors alike.

Investors are paying attention

DeSci platform Bio Protocol recently raised $6.9 million from backers including Maelstrom Fund and Animoca Brands, following an earlier investment from Binance Labs at the end of 2024. The common thread: data access and governance. Several groups are vying for large-scale biomedical datasets, including genetic data, as assets change hands in restructuring processes. Ethics, consent and legal clarity remain non-negotiable.

What healthcare leaders should do next

• Interoperability: Confirm your ultrasound fleet, PACS and EHR can push de-identified data into decentralized training workflows without breaking existing reporting.
• De-identification: Validate your process against recognized standards such as HIPAA de-identification methods. See HHS guidance here.
• Governance: Establish a model registry with clear approval gates, version control and rollback plans. Require site-level opt-in and IRB oversight where applicable.
• Security: Keep PHI offchain. Review vendor key management, access controls and incident response.
• Validation: Run multi-site prospective evaluations for each model update. Track drift by device type, protocol and patient cohort.
• Operations: Quantify gains (report turnaround time, sonographer clicks, denial rates) and tie them to staffing and reimbursement.
• Contracts: Specify data rights, audit rights, liability for model errors and processes for model recall.

Clinical and technical questions to ask vendors

• Which data elements are used for training and how are they de-identified?
• What runs locally vs. onchain or in the cloud, and who can view audit logs?
• How often are models updated and how are updates validated on our devices?
• What bias checks exist across demographics, scanners and protocols?
• What is the fall-back workflow if the AI is unavailable or underperforms?

Key risks and how to mitigate them

• Privacy leakage: Use strict de-identification, k-anonymity checks and prohibit re-linkage across datasets.
• Model drift: Monitor sensitivity/specificity per site and device; gate deployments with thresholds.
• Over-reliance: Keep human-in-the-loop review and clear accountability in clinical notes.
• Regulatory gaps: Align with local regulations and standards; for federated approaches in medicine, see peer-reviewed context here.

Bottom line

Samsung's integration with Galeon points to a practical path for AI in imaging: shared learning without moving patient data. If you lead a hospital or imaging service, treat this as a chance to pilot decentralized training with tight governance and clear ROI metrics.

Want your team fluent in AI options by job role? Explore focused programs at Complete AI Training.