AI radiology companies in 2026: the vendor landscape

An imaging director in 2026 gets pitched radiology AI more often than any other technology, and the market behind the pitches is real: per The Imaging Wire's March 2026 count, the FDA's AI-enabled medical device list held 1,451 cleared algorithms, and 1,104 of them are radiology devices, about 76 percent of everything the agency has cleared, per The Imaging Wire's running analysis of the list.

A disclosure before the map, because this page lives on a teleradiology practice's site: AstraRad performs no AI pre-read and no AI triage. Every study we receive is interpreted and signed by a board-certified, fellowship-trained subspecialist, and nothing here is a product pitch for us. It is the landscape we watch because facilities deploy these tools alongside outsourced reads, and because knowing what AI does well clarifies what it does not do at all. That vantage point is the article.

What do AI radiology companies actually sell?

Strip the branding and nearly every product on the FDA list falls into one of four categories, each with a different buyer and a different failure mode.

Category	What it does	Regulatory shape	Representative vendors
Triage and prioritization	Flags suspected critical findings and reorders the worklist	CADt clearances; assistive only	Aidoc, Viz.ai, Qure.ai
Detection and diagnosis support	Marks suspected findings for the reader; some quantify	CADe and CADx	Lunit, ScreenPoint, Gleamer, Annalise.ai
Quantification and analysis	Measures: coronary flow, nodule volumes, organ segmentation	Device-specific	HeartFlow, many platform tools
Reporting and workflow	Drafts impressions, tracks follow-ups, routes studies	Mostly non-device software	Rad AI, PACS and RIS vendors

The category determines the claim you should test. A triage tool should prove it moves time-to-treatment on your queue, a detection tool should prove sensitivity and a tolerable false-positive rate on your population, a quantification tool should prove reproducibility, and a reporting tool should prove minutes saved per radiologist without new error modes. A pitch that floats between categories is avoiding all four proofs at once.

The AI radiology companies by category

Aidoc is the most deployed pure-play vendor: more than 30 FDA-cleared tools spanning triage across body regions, and in January 2026 it received clearance for a foundation-model-powered CT triage device covering 14 conditions in one product, a first for the category. In April 2026 it closed a $150 million Series E led by Goldman Sachs with Nvidia's venture arm participating, taking total funding past $500 million. Its model is the enterprise platform: one integration, many algorithms, deployed across roughly two thousand sites by its own count.

Viz.ai built the time-critical-condition franchise, starting with large vessel occlusion stroke: the algorithm alerts the interventional team while the study is still in the queue. It holds 50 or more clearances and publishes vendor-cited clinical data showing meaningful treatment-time reductions in stroke workflows. Its historical significance is reimbursement: Viz LVO won the first Medicare New Technology Add-on Payment granted to AI software, approved in the FY2021 inpatient rule under the CMS new technology add-on payment program, which briefly made stroke AI the one radiology algorithm a hospital could bill for.

Rad AI leads generative reporting: automated impression drafting from the radiologist's dictated findings, follow-up recommendation tracking, and report-language tools, in use by thousands of US radiologists since its 2018 founding. This category touches the report itself, so the governing question is different: who is accountable for a drafted sentence the radiologist signs without editing? The signing radiologist, always, which is why adoption runs through radiologist preference rather than administrator mandate.

The single-domain specialists are often the strongest clinical products in their lane. HeartFlow computes fractional flow reserve from coronary CT and carries its own CPT codes, making it one of the few AI products with a standing reimbursement path. Gleamer's BoneView reads fracture studies alongside emergency clinicians. Lunit and Qure.ai built deep chest X-ray and screening portfolios with global-health deployments. ScreenPoint's Transpara supports mammography reading, one of the most studied AI use cases in the literature. Annalise.ai ships comprehensive chest X-ray detection covering over a hundred findings.

The platform layer is where consolidation happens: GE HealthCare, Siemens Healthineers, and Philips bundle their own and third-party algorithms into marketplaces, and PACS vendors increasingly resell AI as worklist features. For a buyer, the platform question is contractual: an algorithm bought through a marketplace inherits the marketplace's update, monitoring, and liability terms, which are usually thinner than a direct medical-device relationship.

Does AI replace the radiologist's read?

No cleared device in 2026 autonomously renders the final interpretation on the CT, MRI, and radiography volume that fills a worklist, and the FDA's clearance categories say so explicitly: triage devices are cleared on the condition that they notify and prioritize while the standard of care, a physician read of every study, continues unchanged. The interpreter of record signs, and the liability follows the signature.

That regulatory line matches the clinical evidence. AI performs best as a second set of eyes with a narrow assignment: surface the probable hemorrhage now, count what is countable, draft the boilerplate. The published failure modes are equally consistent: false-positive burden that trains clinicians to ignore alerts, performance drift when scanner fleets or populations shift from the pivotal study's, and degradation on the edge cases where an experienced subspecialist earns their training. The independent double-read remains the quality mechanism with the longest evidence trail, which is why blind double-reading is how AstraRad's protocol audits its reports: 1 in 20 signed reports re-read blind by a second subspecialist, humans auditing humans.

The honest synthesis: AI complements reads by managing time and attention, and it has not displaced the read. A facility that deploys triage AI still needs a radiologist within the hour for the study the algorithm flagged; the flag has no signature. How fast that signature arrives is a contracting question, covered by tiered commitments like the ones on our SLA page: STAT under 4 hours, with 2-hour and 1-hour STAT available on request at a higher multiplier, and routine 24 to 48 hours.

Where AI genuinely helps a read stream

Four deployments earn their cost most reliably, and each has a measurable claim you can hold it to.

Queue reordering on deep queues. When routine studies wait hours, a CADt flag that promotes a probable pneumothorax to the top demonstrably shortens time-to-read for critical findings. The benefit scales with queue depth, which means it shrinks as your turnaround improves; measure your worst-hour queue before buying, and revisit the STAT discipline framework first, since mislabeled priorities cause more delay at most facilities than missing algorithms do.

Time-critical team activation. Stroke and pulmonary embolism response teams paged in parallel with the read, the Viz.ai model, moves treatment clocks in ways trial data supports. This is workflow engineering as much as AI, and it works where the downstream team actually exists.

Quantification nobody enjoys doing by hand. Nodule volumetrics, vessel measurements, bone age: automation here removes tedium and inter-reader variance without touching the interpretive core. Adoption friction is lowest and the claims are easiest to verify.

Report drafting under radiologist control. Impression generation saves dictation minutes at scale when the radiologist reviews every word. The productivity numbers are real; so is the automation-complacency risk, and mature deployments audit edit rates to catch rubber-stamping.

What does not work is deploying any of these as a substitute for reader capacity. An algorithm cannot absorb your unread overnight list; the subspecialty match on complex studies is still a human variable; and a triage alert on a study nobody is staffed to read within the hour is a documented finding waiting for a plaintiff's attorney.

Who pays, and what that tells you

Reimbursement reality disciplines this market more than any analyst report. Viz.ai's NTAP was time-limited, as all NTAPs are; HeartFlow's CPT codes are the durable exception; the overwhelming majority of the 1,104 cleared radiology algorithms bill nobody. They are bought as operational spend against throughput, length of stay, transfer avoidance, or liability posture, and the purchasing decision looks like any other capital-adjacent one: a pilot with defined endpoints, a measured before-and-after, and a kill criterion.

That is also the lens for reading vendor durability. Companies with enterprise platform contracts (Aidoc), reimbursed products (HeartFlow), or radiologist-pulled workflow tools (Rad AI) have revenue that survives budget cycles. Single-algorithm companies without a reimbursement path consolidate into the platforms; several already have. A buyer signing a three-year AI contract should ask the same change-of-control questions a teleradiology contract gets, and the company evaluation framework translates almost line for line.

How to pilot an AI tool without fooling yourself

The pilot design matters more than the vendor selection, because a badly designed pilot validates whatever was going to be bought anyway. Four rules keep the answer honest.

Run shadow mode first. For the first weeks the algorithm sees live studies and its outputs are logged without reaching any clinician. Shadow mode gives you the true alert volume, the false-positive rate on your actual population, and the prevalence math, all before anyone's workflow changes or anyone's alert fatigue begins. A vendor that resists shadow mode is resisting measurement.

Define endpoints and a kill criterion before day one. A triage pilot's endpoint is time-to-read for the target finding, measured from image arrival, compared against your own baseline quarter. A reporting pilot's endpoint is dictation minutes and edit rate. Write the number that constitutes failure, with a date attached, before the kickoff call; pilots without kill criteria become permanent by inertia, invoiced annually.

Govern it like a medical device, because it is one. Someone clinical owns the deployment: configuration changes are logged, alert thresholds are reviewed on a schedule, and there is a named path for a radiologist to report a miss or a bad flag. The FDA cleared the device; your facility owns its operation, and surveyors and plaintiffs' attorneys both know the difference.

Re-measure quarterly after go-live. Scanner replacements, protocol changes, and population shifts all move performance quietly. The vendor's monitoring answer from the question set below gets tested here: a real monitoring program sends you drift reports without being asked, the same way a real teleradiology QA program sends the monthly discrepancy log unprompted.

The parallel with buying reads is exact, and deliberate. A pilot with a baseline, defined measures, and a written gate is how this site recommends trialing a teleradiology vendor too; rigor is cheaper than switching, in both markets.

The five-question filter for any AI radiology pitch

1. Which exact indication is cleared, and does our intended use match it? Off-label deployment of a cleared algorithm transfers risk to you. Verify the clearance on the FDA list by device name.
2. What was the pivotal study population, and how far is ours from it? Scanner mix, prevalence, demographics. Performance is measured there and only promised here.
3. What is the false-positive burden at our volume? Sensitivity headlines; specificity bills you in attention. Ask for the numbers at the operating point they will actually configure.
4. Whose worklist does it touch, and who integrates? If your reads are outsourced, confirm whether the AI reorders your queue, your teleradiology vendor's queue, or neither. AstraRad's answer, for clarity: our worklist routing is subspecialty and SLA driven, no algorithm pre-reads or re-prioritizes studies inside our practice, and a facility running AI on its own side sends us studies the same way it would without it, per how we work.
5. How is post-deployment performance monitored, and by whom? Drift is a when. A vendor without a monitoring answer is selling a snapshot.

AI radiology companies in 2026 are a real industry solving real queue and attention problems, and the reads themselves still belong to people. If the read side of your operation is the gap, that is the problem we quote: a written rate card from a subspecialist practice, no minimums and no platform fees, within one business day of a request.