At 7:12 a.m., a radiologist opens the worklist and gets hit by a squall of scans - chest CTs, stroke alerts, follow-up MRIs, all piling up before the coffee has even found the bloodstream. Off to one side sits the promise of AI, polished like a new brass compass. It can flag bleeds, triage urgent cases, maybe even shave minutes off a long day. And yet, in hospital after hospital, that compass is still sitting in the drawer.

The paper by Towbin, Kim, Kottler, and colleagues is not another "our model got 0.97 AUC, please clap" affair. It is something rarer and, frankly, more useful: a field report on why medical imaging AI keeps stalling between demo day and daily practice [1].

In March 2025, the Academy for Radiology & Biomedical Imaging Research gathered people from academia, industry, government, and clinical leadership to talk through where AI implementation goes sideways. They used the AI lifecycle and a simplified failure modes and effects analysis, or sFMEA, which is basically a disciplined way of asking, "All right, mates, where is this contraption likely to spring a leak?" [1]

What they found will surprise absolutely nobody who has ever tried to install new software in a large institution and discovered that three committees, two procurement forms, and one mystery server are involved. The trouble clustered around five areas: governance, use cases, implementation, cost, and regulation [1].

The math is fine. The plumbing is mutiny.

This is the heart of it. Medical imaging AI often does reasonably well in studies, but hospitals do not run on benchmark scores alone. They run on workflow, staffing, budgets, trust, interoperability, and legal responsibility - all the boring beams that keep the ship afloat.

That theme shows up all over the recent literature. A 2023 scoping review found recurring barriers such as lack of trust, weak familiarity with AI tools, and fears about professional autonomy [2]. A 2024 multi-society statement from the ACR, CAR, ESR, RANZCR, and RSNA argued that implementation needs active monitoring, local validation, and real operational planning rather than blind enthusiasm [3]. Another 2024 RSNA-MICCAI expert report made the same call in plainer terms: align expectations, resources, data flow, and accountability before you toss AI onto the deck and hope it rows [4].

That last part matters. AI in radiology is not magic. It is more like hiring an extremely fast junior mate who never sleeps, sometimes spots things you miss, and occasionally says something odd with the confidence of a pirate describing tax law.

Five storms on the horizon

Towbin and colleagues lay out five recurring trouble spots [1].

Governance: Who owns the decision to buy, test, monitor, pause, or retire an AI tool? If the answer is "sort of everybody," the real answer is nobody.

Use cases: A model can be technically impressive and still solve a problem nobody ranks as urgent. Hospitals need tools that fit actual pain points, not just conference slides.

Implementation: Even good models fail when they do not mesh with PACS, reporting systems, or clinician habits. This is where many elegant ideas drift onto the shoals.

Cost: AI is not just a license fee. It is IT time, validation work, training, oversight, and sometimes infrastructure upgrades. The GPU may be the glamorous deck cannon, but the invoice usually arrives from below deck.

Regulation: The FDA has been building out guidance for AI-enabled medical devices, including transparency principles in June 2024 and draft lifecycle guidance on January 7, 2025 [5][6]. That is progress, but for hospitals trying to compare tools, monitor updates, and assign responsibility, the waters are still choppy.

Why this matters outside the radiology harbor

If this paper lands with a thud instead of fireworks, good. That is the sound of a field growing up.

Recent evidence suggests the market is filling fast, but the proof for real-world value is still patchy. A 2025 review of 173 commercial radiology AI products found more peer-reviewed support than a few years ago, yet higher-level evidence on patient or socioeconomic outcomes remains relatively thin [7]. Another 2025 scoping review found a boom in published studies but noted that practical implementation, experience, and cost evidence are still uneven [8].

Meanwhile, the adoption story is moving whether everyone feels ready or not. Pew reported in March 2025 that many hospitals had used FDA-cleared imaging AI with limited piloting and monitoring, and that radiology still dominates the FDA's AI-enabled device landscape [6]. That is the part where the old sea captain narrows his eyes. More tools in the harbor means more need for charts, rules, and competent watchkeeping.

The big contribution of this roundtable paper is that it stops pretending the bottleneck is just model performance. The bottleneck is coordination. Industry wants scalable products. Academics want rigor. Clinicians want tools that help today, not someday. Regulators want safety. Administrators want costs that do not hit like a rogue wave. Everybody is, in a sense, correct. Everybody is also standing on a different part of the ship yelling into the wind.

The course correction

The paper's implied advice is plain enough: treat implementation as a full system problem, not a software install [1]. Pick use cases that matter. Build governance before rollout. Validate locally. Monitor continuously. Keep humans accountable. In sailor's terms, do not celebrate a fine new compass if the hull is taking on water.

That may sound less thrilling than tales of superhuman AI, but it is how useful technology actually reaches patients. Not with a trumpet blast. With checklists, shared incentives, fewer blind spots, and a crew that agrees on where the harbor entrance actually is.

References

1. Towbin AJ, Kim W, Kottler N, et al. Bridging Industry and Academia: Proceedings from the 2025 Academy Roundtable on AI Implementation in Medical Imaging. Radiology: Artificial Intelligence. Published online May 6, 2026. DOI: 10.1148/ryai.250671. PubMed: PMID 42089795

2. Eltawil FA, Atalla M, Boulos E, Amirabadi A, Tyrrell PN. Analyzing Barriers and Enablers for the Acceptance of Artificial Intelligence Innovations into Radiology Practice: A Scoping Review. Tomography. 2023;9(4):1443-1455. DOI: 10.3390/tomography9040115. PMCID: PMC10459931

3. Brady AP, Allen B, Chong J, et al. Developing, purchasing, implementing and monitoring AI tools in radiology: practical considerations. A multi-society statement from the ACR, CAR, ESR, RANZCR & RSNA. Insights into Imaging. 2024;15(1):16. DOI: 10.1186/s13244-023-01541-3

4. Linguraru MG, Bakas S, Aboian M, et al. Clinical, Cultural, Computational, and Regulatory Considerations to Deploy AI in Radiology: Perspectives of RSNA and MICCAI Experts. Radiology: Artificial Intelligence. 2024;6(4):e240225. DOI: 10.1148/ryai.240225

5. U.S. Food and Drug Administration. Artificial Intelligence in Software as a Medical Device. Updated March 25, 2025. FDA page

6. The Pew Charitable Trusts. Research Reveals Gaps in Oversight of Artificial Intelligence for Radiology. March 19, 2025. Article

7. Antonissen N, Tryfonos O, Houben IB, et al. Artificial intelligence in radiology: 173 commercially available products and their scientific evidence. European Radiology. 2025. DOI: 10.1007/s00330-025-11830-8

8. Lawrence R, Dodsworth E, Massou E, et al. Artificial intelligence for diagnostics in radiology practice: a rapid systematic scoping review. eClinicalMedicine. 2025;83:103228. DOI: 10.1016/j.eclinm.2025.103228. PMCID: PMC12140059

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.