Autonomy at the Med Bay: AI-Guided Scans and Self-Service Diagnostics
NVIDIA, GE HealthCare and Franka Robotics are collaborating to advance innovation in autonomous diagnostic imaging with physical AI, including autonomous X-ray technologies and ultrasound applications.
Credit: NVIDIA / GE HealthCare / Franka Robotics
When latency and limited staff force judgement onboard, checklists evolve into algorithms
For exploration missions, medical autonomy isn’t a luxury—it’s survival. On Mars, latency kills telemedicine; on the Moon, bandwidth and schedules will still conspire against real-time consults. The path forward is clear from ISS practice: move from remote-guided procedures to on-board AI guidance that helps crews collect and interpret data safely when experts aren’t available.
Start with what’s proven. NASA’s Ultrasound-2 protocols show non-physician crew can capture decision-grade images using disciplined, codified guidance. The documentation is strikingly specific—probe moves quantified in centimetres and degrees; colour-coded keyboard prompts; standard phrasing to avoid confusion. In parallel, the station’s ocular-health work uses Optical Coherence Tomography (OCT) to monitor Spaceflight-Associated Neuro-ocular Syndrome (SANS), a well-characterised risk on long missions. These are the two pillars of autonomy: reliable acquisition and focused interpretation.
AI now has credible toeholds on the interpretation side. A 2024 study introduced SANS-CNN, a model trained on real astronaut OCT images to flag SANS-like features quickly—exactly the kind of onboard aid that could triage vision risks when comms are delayed or crew are busy. More broadly, reviews of SANS imaging and NASA’s latest “Vision Changes” brief underscore that OCT, ultrasound and fundus photography are the core modalities—good targets for AI quality-grading and pre-reads before ground review.
On the acquisition side, Tele-SUS showed astronauts can self-scan muscles accurately with ground guidance. Translate that to autonomy and the near-term product is AI that coaches probe handling—“slide 1 cm left,” “rotate 10° toward the head,” “image quality acceptable”—borrowing directly from Ultrasound-2 playbooks. The aim isn’t diagnosis in a box; it is image quality assurance and structured capture so later interpretation (human or AI) is trustworthy.
Considering NVIDIA, GE Healthcare and Franka Robotics’s recent push toward autonomous diagnostic imaging with physical AI, when latency and limited staff force judgment onboard (think Moon/Mars missions), checklists must evolve into algorithms. The path is to embed edge AI that coaches acquisition and triages interpretation so non-physician crews capture decision-grade data without real-time specialists—building on NASA’s Ultrasound-2 and OCT workflows for SANS, early models like SANS-CNN, and self-scanning lessons from Tele-SUS.
Commercially, the near-term deliverable is a modular kit: ultrasound with AI quality guidance, OCT workflow with AI triage, and SOPs that tell crews when to use which tool and what to log. Package everything as upgradeable software with independent test witnessing and versioned release notes so safety boards can audit changes. For terrestrial buyers (submarine fleets, polar stations, remote mines), the same kit reduces medevacs and elevates first-line care.
Risks centre on over-trust and drift. Guardrails include explainable prompts (“quality low due to motion”), strict scope limits (AI flags, humans decide), and periodic ground truth checks where onboard results are reviewed by specialists post-hoc. Privacy and data-sovereignty concerns are mitigated by on-device processing with selective sync.
The end-state isn’t a doctor in a box; it’s a crew-operated diagnostic stack that turns latency and personnel limits into manageable constraints. We already have the protocols and devices; the job now is to embed them with AI that keeps quality high and decisions safe—even when no one answers on the other end of the line.
