Tag: AI

Harrison’s Teleradiology Play

Posted on by Brian Casey

AI developer Harrison.ai is expanding its business model by supporting a teleradiology venture called Frontier Radiology that’s scaling up and hiring radiologists. But the move drew unwanted attention from a critical article in the Australian news media this week.

Harrison is best known for its suite of AI algorithms for radiology applications, as well as its Harrison.Rad 1.5 foundation model that can draft reports for radiologist review.

  • Harrison is based in Australia, but has recently made the U.S. market a major focus, with co-founder and managing director Dimitry Tran relocating to California to spearhead the initiative.

But it was Harrison’s newfound U.S. focus that apparently drew the attention of ABC News, an Australian news outlet that on September 7 published a critical article that questioned the company’s moves, in particular its involvement with Frontier Radiology.

  • One bone of contention in the story was whether Harrison’s U.S. emphasis was appropriate given the Australian government’s 2025 investment of US$23M in the company. The story also claimed Harrison was planning layoffs of Australian staff and that its integration of AI with Frontier’s clinical services could be a conflict of interest.

But in an interview with The Imaging Wire, Dimitry Tran clarified Harrison’s strategic direction and its involvement with Frontier Radiology.  

  • Tran noted that Frontier is owned and operated by a radiologist – Joshua Ewell, DO – in accordance with U.S. rules requiring physician ownership of entities providing clinical services. 

Ewell’s LinkedIn profile says Frontier will be an “AI-native radiology group” built entirely around AI foundation models.

  • Harrison is providing Frontier with non-clinical services, including its suite of AI algorithms, which are already helping Frontier teleradiologists achieve industry-leading report turnaround times. But Frontier radiologists are free to use any AI solutions they wish – including those of Harrison’s competitors. 

Tran noted that close cooperation between AI developers and imaging services providers is hardly unusual anymore in radiology. 

  • Indeed, two of the specialty’s largest U.S. providers – Radiology Partners and RadNet – have formed their own AI divisions to provide algorithms to both their own radiologists and outside customers. 

The ABC News article conflated a number of recent developments into a narrative that doesn’t reflect reality, Tran believes. 

  • For example, the layoffs that occurred earlier this year at Harrison were connected to the company’s transition from an aggressive R&D phase into a commercialization push and weren’t related to Harrison’s U.S. entry.

And the Australian government’s funding was part of an investment that gave it a single-digit equity stake in the company – a stake it retains to this day and that will prove increasingly profitable with Harrison’s growing success.

  • Even as it supports Frontier, Harrison plans to continue its focus on AI algorithm development and commercialization, especially of the Harrison.Rad foundation model – while keeping the “vast majority” of its employees in Australia. 

The Takeaway

Putting aside the ABC article’s negative spin, Harrison’s move into teleradiology offers an intriguing twist on the growing integration between AI and imaging services providers. Given ongoing workforce shortages and rising imaging volume, it’s perhaps the best way to move the chains toward finding relief for beleaguered radiologists. 

Your Reporting Platform Is Now Your AI Strategy

Posted on by Partner Content

By Sheela Agarwal, MD, MBA, a practicing radiologist and chief medical information officer at Microsoft 

For most of my career, the radiology reporting platform was a workflow tool. Today, it is an AI strategy, and the way we evaluate it must change with it. 

I see that from two sides. As a practicing radiologist, I live in the reporting workflow. As CMIO at Microsoft, I help shape how technology serves it. 

  • The platform underlying radiology reporting now determines something bigger than features. It determines how effectively AI is integrated into the workflow and where it is making the biggest impact, for both radiologists and the care patients receive.  

The pressure is real, and it is human. Imaging volumes keep climbing while the workforce to read images struggles to keep pace, and the cognitive load on every radiologist grows with it. 

  • The question is no longer whether to adopt AI. It is which foundation can carry AI reliably, study after study. 

Most reporting solutions were not built to be that foundation. PowerScribe One is. The proof is in the scale: 

  • 280+ organizations, proven across IDNs, academic centers, community hospitals, and independent practices. 
  • 10,000+ radiologists have made PowerScribe One part of how they work. 
  • 10M+ reports every month, a volume that reflects real, sustained use. 

That scale is not the point on its own. It is evidence that the technology delivers where it counts. What matters to a radiologist is more specific: does it lighten a heavy worklist, or just add another click? 

As an AI companion to PowerScribe One, Dragon Copilot builds on that foundation, bringing prior reports, patient context, and information from credible sources to the same screen as they read, so radiologists work with a more complete picture without leaving their workflow. 

  • The deeper value is what it frees radiologists to do: practice at the top of their license, spending less time on the mechanics of report creation and more on the complex reads and the diagnostic judgment that shape a patient’s care. 

A reporting platform must be ready for what comes next, evolving alongside radiology and AI. 

  • For organizations still on PowerScribe 360, now is the time to evaluate the path forward. Migration to PowerScribe One preserves existing workflows and configurations, allowing organizations to modernize at their own pace without disrupting care. 

The Takeaway 

This is not simply a radiology reporting system upgrade. It is choosing a partner committed to practical innovation, built on decades of workflow expertise and deep collaboration with customers and partners. Connect with Microsoft’s team or click here to learn more about PowerScribe One and Dragon Copilot.

Where Do Humans Belong in the AI Loop?

Posted on by Jack Troy

AI has its skeptics and superfans when it comes to interpreting medical images, inspiring a wide range of approaches to checking its work, but a new RadioGraphics paper argues the right level of oversight lies somewhere in the middle. 

There’s a strong consensus on the importance of monitoring AI post-deployment, even as the right level of rigor remains an open question. 

  • The FDA and many of its global peers require institutions to track AI performance, implement human oversight, and have a corrective action plan. 
  • Past studies show oversight mechanisms are top of mind for many radiologists. 

The international researchers behind the article say a human-on-the-loop (HOTL) model is the sweet spot for radiology departments as they balance safety and reliability with efficiency. 

  • Under this framework, radiologists don’t have to review every AI output.
  • At the same time, the proposed model doesn’t take AI results at face value.
  • HOTL threads the needle by alerting humans to drift and accuracy problems.
  • If performance drops, the work of reading these scans reverts to radiologists. 

The model acknowledges that even the most cleverly designed radiology algorithms can be sensitive to changing inputs.

  • Scanner upgrades, workflow adjustments, and shifts in patient mix can leave AI that aced training tests out of step with clinical realities. 

HOTL has its perks, but the authors note a few challenges. 

  • Excessive alerts risk desensitizing monitoring teams, so it’s important not to set the notification threshold too low or treat minor deviations as urgent. 
  • Declines in subgroup performance can go unnoticed if overall metrics stay stable.

The Takeaway

For radiology departments, the ideal AI oversight plan protects patients without making the technology more trouble than it’s worth. It seems the HOTL system outlined in RadioGraphics checks those boxes by focusing on trends over individual outputs. Even so, the debate over how to best supervise AI is sure to remain lively. 

Top 10 AI Vendors by FDA Approvals

Posted on by Brian Casey

Who are the top 10 radiology AI vendors, based on the number of FDA regulatory authorizations? The agency provided some clarity this week with an update to its list of authorized AI-enabled medical devices through the end of Q1 2026. 

The FDA updates the list on more or less a quarterly basis, and it’s become a closely watched barometer for tracking not only the health of the AI industry but also which companies have received the greatest number of authorizations.

  • As we’ve noted in the past, the list includes both standalone AI algorithms as well as medical hardware that has AI functionality embedded in it, like a mobile X-ray machine with an onboard AI feature for detecting fractures.

The updated list tracks marketing authorizations through the end of March 2026, and shows that the FDA has…

  • Authorized 1,524 AI-enabled medical devices since it began keeping track in 1995, up 5.1% from Q4 2025. 
  • Authorized a total of 1,164 radiology devices, or 76% of all AI-enabled medical authorizations. 
  • In the first quarter of 2026, the FDA authorized 92 AI-enabled medical devices, or 28% more than in the fourth quarter of 2025.
  • For the quarter, 69 authorizations (75%) were for radiology devices, about the same ratio as in Q4 2025 (76%). 
  • GE HealthCare held its lead as the company with the most radiology AI authorizations at 130 (including recent acquisitions that had AI authorizations of their own).
  • Next is Siemens Healthineers at 95, then Philips at 58, Canon at 48, United Imaging at 40, Aidoc at 33, and DeepHealth at 29, with all numbers including acquisitions. 
  • Rounding out the top 10 are Samsung (21), Rapid.ai (20), and Hyperfine (13).

The Takeaway

The FDA’s new numbers on AI marketing authorizations show that the agency is keeping pace with rapid developments in the healthcare AI industry. Indeed, the FDA is even accelerating its pace of product approvals compared to its last update, with radiology still securing the lion’s share of authorizations.

AI for Breast Cancer Risk

Posted on by Jake Fishman

Artificial intelligence may be capable of identifying subtle mammographic signs of breast cancer years before conventional diagnosis, according to a new study published in Radiology. Researchers from Sweden found that three commercially available AI algorithms for mammography screening generated elevated cancer scores as early as 10 years before diagnosis, with detection signals strengthening as diagnosis approached.

Predicting breast cancer risk offers the prospect not only of detecting cancer earlier, but also of tailoring mammography screening to women most likely to benefit from it.

  • Clinical risk calculators like Tyrer-Cuzick and breast density analysis are available, but AI-based algorithms are showing promise by predicting risk from screening mammograms.

In the new study, researchers analyzed 89k mammograms from 31.4k women collected over a 10-year period, drawn from Sweden’s national screening program, where women aged 40-74 undergo biennial mammography interpreted by two radiologists.  

  • During the study period, 12.1k women (39%) were ultimately diagnosed with breast cancer. Three commercially available AI algorithms were used to generate risk scores (Vara AI from Vara, Lunit Insight MMG from Lunit, and MammoScreen from Therapixel). (It’s worth noting all three were originally designed for cancer detection rather than risk prediction.) 

AI scores increased progressively over time in women who later developed cancer, while remaining relatively stable among cancer-free participants…

  • At 90% specificity, AI systems flagged 19%-20% of future breast cancer cases six years before diagnosis.
  • Detection increased to 23%-25% at four years before diagnosis.
  • Performance rose further to 35%-39% at two years before diagnosis.
  • Even 10 years before diagnosis, the systems identified 13%-17% of future cancers.
  • Across all pre-diagnostic examinations, AI achieved AUC values of 0.63-0.67, outperforming mammographic density alone (AUC = 0.57).

The findings suggest that AI tools developed for cancer detection may also have value as early-alert systems for identifying women who could benefit from closer surveillance or supplemental imaging.

  • While prospective validation is still needed, sequential AI scoring may ultimately help identify women who would benefit from supplemental imaging, closer surveillance, or earlier intervention.

The Takeaway

The study adds to growing evidence that mammography AI can extend beyond cancer detection to long-term risk stratification. By identifying subtle imaging patterns years before diagnosis, AI-derived detection scores could provide an additional layer of longitudinal risk monitoring and help guide more personalized screening strategies.

AI for Chest X-Ray Varies

Posted on by Brian Casey

Not all AI is created equal when it comes to analyzing chest X-rays. A new study in Radiology found wide variation in performance for seven commercially available chest X-ray algorithms to detect lung cancer. 

X-ray is by far the most widely used imaging modality. Radiography is often the first imaging exam a patient receives, and it frequently serves as a gateway to other more advanced imaging modalities. 

  • But radiography also has well-known shortcomings (which is why advanced imaging is needed for follow-up). Could AI help unlock X-ray’s value and make it more useful?

That’s what a host of AI algorithm developers are banking on, but the wide variety of solutions can create confusion for clinicians.

  • So U.K. researchers decided to hold an AI bake-off, comparing commercially available algorithms from seven developers for detecting lung cancer on chest X-rays. 

The competing companies included Annalise/Harrison.ai, Gleamer, Infervision, Milvue, Oxipit, Qure.ai, and Rayscape. Researchers anonymized performance results from the different products.

In all, chest radiographs from a dataset of 5.2k patients with a real-world lung cancer prevalence rate were included, with researchers finding…

  • Significant variance in algorithm performance by each of the major accuracy measures: sensitivity (21%-78%), specificity (59%-98%), and positive predictive value (1.5%-28%). 
  • All the algorithms increased the number of false positives, and with significant variation. One model generated only 10 more false positives than radiologists, while another produced – wait for it – over 2k. 
  • If used to triage patients for follow-up CT exams, one model would generate $1.6k in additional costs while another would produce $327k.

What accounts for the variation? An underlying factor is most likely differences in the datasets used for model training. 

  • In any event, the study underscores the need for more head-to-head comparisons to determine the strengths and weaknesses of individual AI algorithms. 

The Takeaway

This week’s study on how AI performance varies between commercially available algorithms initially seems disturbing and might suggest a need for stronger regulatory oversight. But AI’s diversity could be its strength in a future where every patient case is analyzed by multiple different algorithms, each with its own advantages. This could ultimately produce a more complete picture of the patient than any one algorithm on its own.

AI for PE Detection: ‘Selective but Meaningful’

Posted on by Brian Casey

AI made a “selective but meaningful” contribution to radiologist interpretations of CT pulmonary angiography scans for pulmonary embolism. The study, published in Radiology: Artificial Intelligence, offers valuable insights into real-world implementation of AI on a large scale. 

One of the major criticisms of AI is that algorithms used in real-world clinical situations don’t perform as well as they do in the controlled environments that vendors use to acquire data for regulatory submissions.

  • AI performance can drop off as much as 20 to 30 percentage points for important metrics like sensitivity and specificity. 

The new study sought to investigate this phenomenon by analyzing a real-world implementation of Aidoc’s AI algorithm for PE detection. 

  • Researchers assessed the algorithm’s performance for analyzing CTPA exams across a variety of clinical environments in an integrated health network, including the emergency department and inpatient and outpatient settings. 

Scans of 29.5k patients acquired from 2021 to 2023 were included. AI analyzed images in real time, after which exams were interpreted by radiologists who knew the AI findings. Researchers found…

  • Radiologists using AI had higher sensitivity than the algorithm on its own (99% vs. 85%).
  • Specificity was more or less the same (99.8% vs. 99.5%).
  • Agreement between radiologists and AI was high (98%).
  • Agreement was higher when AI assessed cases as negative rather than positive (98% vs. 94%).
  • Radiologists disagreed with AI in 2.2% of cases. The final determination by a panel of expert thoracic radiologists strongly favored radiologists (89%).
  • Of the 3.3k cases positive for PE, 0.81% were detected only by AI – or 26 cases.

In analyzing the results, the researchers characterized AI’s contribution as “selective but meaningful.”

  • AI-positive results meant scans might require more scrutiny from radiologists, while an AI-negative call might be supportive – but not definitive – for negative PE.

The Takeaway

The new study of AI for PE detection is a fascinating look at real-world AI deployment. While the sensitivity, specificity, and agreement numbers are interesting, what draws our attention is the 26 PE cases caught only by AI over 18 months of use. That boils down to 26 patients whose clinical condition wasn’t missed, and 26 potential malpractice lawsuits that were never filed.

Mammo AI Momentum Builds

Posted on by Brian Casey

Momentum is building toward routine clinical use of AI for breast cancer screening. Several new studies offer even more support for mammography AI, including research published today in Nature Medicine in which AI reduced radiologist workload by over 60% by excluding low-risk studies from human review.

Breast screening has become one of the most promising use cases for AI, with the potential to reduce radiologists’ workload while improving their ability to detect cancer. 

  • For example, the recent MASAI study found that ScreenPoint Medical’s Transpara AI algorithm could replace the second human reader in a double-reading protocol, reducing workload by 44% and improving cancer detection rates by 28%.

The new research in Nature Medicine also used Transpara, as part of the AITIC study in Spain with the goal of seeing if AI could triage low-risk studies so they don’t require review by human radiologists. 

  • AITIC had a prospective design, involving 31k women with screening exams split between 2D mammography (17k) and digital breast tomosynthesis (14k). 

Women in the control arm of the study got conventional double reading by two radiologists – the standard mammography paradigm in Europe.

  • The intervention arm used a partially autonomous AI approach: cases that AI interpreted as low risk were classified as normal and were not reviewed by radiologists, while all other cases were double-read by radiologists using AI support.

In analyzing the results, researchers found…

  • Workload in the AI arm was 64% lower than conventional double reading.
  • AI’s workload reduction was similar between DBT and conventional digital mammography (-66% and -62%, respectively).
  • The AI arm’s cancer detection rate per 1k women was 15% higher (7.3 vs. 6.3 cancers).
  • But the recall rate was also 15% higher.

It’s worth noting that the AITIC study differed from MASAI in its inclusion of DBT screening exams, whereas MASAI only included 2D digital mammography. 

  • While 2D mammography is the norm in Europe, much of the U.S. has switched to DBT for breast screening, so the AITIC results offer good news for U.S. breast imaging practices considering AI adoption.

The Takeaway

The AITIC study’s new results are powerful confirmation of findings from the recent MASAI trial and support broader clinical deployment of mammography AI. Taken together with positive findings from last week’s Nature Cancer articles (see The Wire section in this newsletter), they paint a picture of a technology that’s ready for prime time.

FDA Updates AI List with New Clearances

Posted on by Brian Casey

The FDA last week updated its list of cleared AI-enabled medical devices, with the new list showing AI marketing authorizations through the end of 2025. The updated list reveals that radiology is maintaining its lead as the medical specialty with the most clearances.

The FDA’s previous update featured data through the end of September 2025, and showed the number of AI-enabled medical devices for radiology crossed the 1k mark. The new numbers show continued momentum for medical imaging.

  • The agency’s data go all the way back to 1995 (the first cleared radiology device on the list was ImageChecker from R2 Technology/Hologic in 1998). 

The new list tracks authorizations through the end of December 2025, and indicates the agency has…

  • Authorized 1,451 AI-enabled medical devices since it began keeping track in 1995.
  • Approved 1,104 radiology devices, or 76% of total AI-enabled medical authorizations.
  • In the fourth quarter of 2025, the FDA cleared 72 AI-enabled medical devices, of which 55 (76%) were radiology devices. 
  • For all of 2025, radiology secured 75% of authorizations, compared to 73% for all of 2024 and 80% for 2023. 
  • GE HealthCare retained the top spot as the company with the most radiology AI authorizations at 120 (including acquisitions Bay Labs, BK Medical, Caption Health, MIM Software, icometrix, and Spectronic Medical).
  • Next is Siemens Healthineers at 89 (including Varian), then Philips at 50 (including DiA Analysis and TomTec), Canon at 45 (including Vital Images and Olea), United Imaging at 38, Aidoc at 31, and DeepHealth at 28 (including Quantib and iCAD). 

As we’ve noted in the past, the FDA’s list includes not only standalone software applications, but also imaging hardware with embedded AI applications, such as a mobile X-ray system with AI algorithms for detecting emergent conditions. 

The Takeaway

The new FDA list shows radiology’s continued dominance when it comes to AI-enabled medical device technology. But an interesting subtext is the ongoing consolidation in the radiology AI space, which could mean that some firms may be climbing the list quickly.

Agentic AI for Radiology Follow-Up

Posted on by Brian Casey

Agentic AI has quickly become one of the hottest topics in radiology. But what is it really good for? Texas researchers offer one possible use case in a new study in NEJM Catalyst: scouring radiology reports to identify patients who require follow-up. 

Agentic AI is a new flavor of artificial intelligence that’s capable of working autonomously to complete tasks with minimal human supervision.

  • In healthcare, it’s being applied to a wide range of tasks, from improving health system operations to clinical and administrative jobs.

In the current study, researchers from Parkland Health in Dallas assigned agentic AI to one of the trickiest tasks in radiology: making sure patients with suspicious findings comply with recommendations for follow-up procedures.  

  • Previous studies have documented low rates of adherence to radiologist recommendations for follow-up imaging (possibly as low as 50%), creating the uncomfortable possibility of missed opportunities that could have major patient-care ramifications.

The dilemma can be compounded with the use of structured note templates in EHRs, as improper use or modification of these macros can lead to missed notifications. 

  • To address the problem, Parkland clinicians developed an AI agent based on a pretrained open-source large language model (Meta’s LLM Llama 3 70B) that reviews clinical impressions, extracts important details for follow-up, and integrates its findings into departmental workflow to enable patient outreach.

In tests on 10k radiologist notes, Parkland researchers found that their AI agent…

  • Had an overall detection rate of ~5.1%, slightly lower than other published studies (8% to 12%).
  • Had far higher sensitivity than Parkland’s previous macro-based follow-up notification system (99% vs. 16%), correctly flagging 6X more cases (513 vs. 83).
  • Achieved higher accuracy (99% vs. 58%), and 94% accuracy for characterizing follow-up timing, recommended procedure, and underlying abnormality. 

Considering Parkland’s annual volume of 500k imaging studies, the AI agent could identify 21.5k follow-up cases a year. 

  • Many of these could be serious issues, such as new cancer diagnoses or pathologies that require surgical intervention. 

The Takeaway

The new study shows that agentic AI isn’t some technogeek’s far-off dream – it’s a useful tool on the verge of real-world implementation, with the potential to improve patient care without overburdening radiology staff.

Get every issue of The Imaging Wire, delivered right to your inbox.