When AI Guides the Scan, Governance Must Follow the Hands

When AI Guides the Scan, Governance Must Follow the Hands

The FDA has cleared ThinkSono Guidance, the first AI-powered ultrasound software authorized for evaluating suspected deep vein thrombosis, and the clearance is worth pausing on for what it is not. It does not interpret images and flag disease the way most cleared imaging AI does. It guides the clinician’s hand in real time, offering acquisition guidance so a non-specialist can capture a diagnostic-quality scan during the exam itself, and it reached market through the FDA’s Breakthrough Device Designation program before securing 510(k) clearance in early 2026 (Newswire, Diagnostic Imaging).

That distinction matters more than the headline. Diagnostic-interpretation AI fails into a human backstop: a radiologist reviews the flagged image before anything happens to the patient. Acquisition-guidance AI fails into the moment itself. If the software steers a clinician’s probe placement incorrectly, there is no second read before the scan is acted on. The failure mode is behavioral, not just statistical, and it lives inside the clinical encounter rather than downstream of it.

FDA’s own posture suggests it sees this distinction too. Despite a January 2025 executive order pledging to keep AI developers “free to innovate without cumbersome regulation,” the agency has continued issuing draft guidance specifically on AI device and software evaluation (Medical Device Network). That is not a contradiction. It is a signal that the agency is separating deregulation of classification and market entry from the underlying question of what evidence a device actually needs to generate, and for real-time guidance tools, that evidence increasingly has to include how humans behave under the AI’s direction, not just how accurate the AI’s underlying signal is.

For regulated buyers, hospital systems, and device sponsors evaluating this emerging category, the decision-relevant shift is in what counts as sufficient validation. Sensitivity and specificity metrics answer whether the algorithm is right. They do not answer whether a sonographer with six months of experience, guided by the software under real clinical time pressure, actually acquires the correct view, or whether guidance cues create new error patterns that didn’t exist in unguided practice. That is a human factors and workflow question, and it belongs in the same governance documentation as model performance, not treated as a separate usability afterthought.

This is exactly the terrain an AI management system under ISO 42001 is built to cover: intended use documented tightly enough to distinguish acquisition guidance from diagnosis, post-market monitoring that tracks operator behavior alongside model drift, and evidence plans that anticipate a notified body or FDA reviewer asking not just “is the model accurate” but “what happens when a clinician trusts it in the moment.”

Procurement and compliance teams sourcing this class of device should ask vendors for human factors testing data and real-world operator performance monitoring plans before clearance status becomes the only checkbox. Breakthrough Device Designation earns speed to market. It does not substitute for proof that the hands on the probe are being guided well.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The central argument—that acquisition-guidance AI has a fundamentally different failure mode than diagnostic-interpretation AI, requiring human-factors validation rather than just accuracy metrics—is
Source & Claim Verification	Qwen · local	cleared. The draft adequately cites sources for most factual claims, but a few lines lack direct citations, such as the discussion on the FDA’s posture and the distinction between diagnostic and acquisition gu
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects FDA, ISO 42001, and EU AI Act considerations for real-time AI guidance tools, particularly human factors and governance, but omits explicit MDR/IVDR alignment for medi
Technical Accuracy	Llama	cleared. The article accurately highlights the distinction between AI-powered diagnostic-interpretation and acquisition-guidance in medical imaging, and correctly identifies the need for additional governance
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and counters potential vendor hype by distinguishing between different types of AI clearance and emphasizing the critical role of human factors in real-time guidanc
Novelty & Non-Duplication	Grok	held. The ThinkSono clearance is pure wire copy, but the acquisition-vs-interpretation failure-mode distinction and its human-factors governance implication supply a differentiated analytical peg not duplic
Validation	DeepSeek	cleared. The central claim that acquisition-guidance AI requires different validation than diagnostic AI is strongly supported by the FDA’s clearance of ThinkSono and its distinct failure mode.

Sources cited: 13. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.