Tag Archives: Radiology

Navigating AI Decisions: Going Beyond “Buy vs Build”

Conceptual illustration of an AI buy-versus-build crossroads with governance and monitoring decisions

“Should we buy or build?” is usually the first question organizations ask about radiology AI. It is useful, but it is no longer the whole question.

Today, you can choose from cleared clinical algorithms, enterprise AI platforms, generative-AI tools, and internally developed workflows. The real challenge begins after a contract is signed or a prototype works. AI becomes another clinical system: one that can help, distract, confuse, or quietly change behavior.

Anyone who has lived through an enterprise IT rollout, PACS replacement, or hospital merger (let alone lead an aspect of it) knows the pattern. Installation is only one part of deployment. The hard work is making a tool fit the people, data, incentives, exceptions, and failure modes of the place where it will be used.

Continue reading

AI Reimbursement Is Becoming a Workflow Problem

CMS has proposed a new way to recognize and potentially pay for certain software-based clinical services (always start with the fact sheet – far more readable than the full text). This may matter in radiology, but it does not mean every AI tool will suddenly be reimbursed by Medicare.

The proposal calls these tools Software as a Medical Service. Examples include software that extracts useful information from medical images, such as heart blood-flow analysis, fracture-risk scores, or brain MRI comparisons. CMS would give some of these services their own payment category while it learns more about how they are used.

Medicare is beginning to acknowledge that some software can provide clinical information beyond a simple workflow aid, and it augments rather than replace physician work (Radiologists have been saying this for a while, and other -ologies expected to follow as product lines expand). Still, the proposal is temporary and limited. Many AI tools would not qualify for separate payment, and payment will depend on appropriate ordering, documentation, billing, and medical necessity.

Clear language matters – CMS differentiates between assistive, augmentative, and autonomous. Software that helps a radiologist work faster (assistive) is different from software that provides new, clinically useful measurements (augmentative). Both differ from a system that makes a diagnosis without a clinician (autonomous). Administrative tools such as scheduling or drafting messages can be useful, but they are not diagnostic services.

The bigger challenge is workflow. Buying or building a tool is only the beginning. All of it involve some form of cost. Hospitals must integrate it into their systems, review privacy and security, train staff, check its performance, document its use, bill correctly, manage denials, and show that it improves care. On the flip side, a billing code alone does not guarantee revenue.

For radiology leaders, the practical next step is to take inventory: Which software tools are already in use? What do they add to patient care? Are eligible services being documented and billed correctly? What do they cost, and what clinical decisions do they improve?

CMS’s proposal is a meaningful step, but not a windfall. The opportunity is to identify the software services that provide distinct value, and build reliable workflows around them.

FDA AI Guidance and the Hard Part of Transparency

The least glamorous part of AI in radiology may turn out to be the most important: telling people what changed.

That sounds simple. It is not. A diagnostic AI tool may be trained on one dataset, validated on another, deployed inside a PACS or reporting workflow, monitored after release, and then updated when the model, threshold, input, interface, or intended environment changes. Somewhere in that chain, a radiologist is expected to decide whether to trust a box, a score, a contour, a triage flag, or a sentence.

Continue reading

AI Readiness as an Educational Obligation

I wrote a guest editorial in Academic Radiology about AI readiness as a practical educational obligation, rather than an optional informatics side quest.

The piece responds to a study of medical undergraduates and radiology trainees in China, but the issue is broader: enthusiasm and exposure do not equal competence. Radiologists need enough AI literacy to recognize workflow fit, automation bias, governance gaps, and the limits of machine suggestions. Most trainees will not become model developers.

They still need to become safe, skeptical operators who can use AI with confidence and accountability in real clinical environments, worldwide, across diverse resource settings today.

What Radiology Teaches Us About AI and the Workforce

Health systems frequently ask whether artificial intelligence will ultimately reduce the need for clinicians. Radiology provides one of the clearest real-world answers to date. AI changes the mechanics of work far more than it changes the need for expertise.

I had the privilege to speak with CNN on this topic recently. The story is out now.

In imaging operations, AI is already being used to support exam prioritization, improve image reconstruction, and reduce friction in routine workflows. These contributions are meaningful, particularly in high-volume environments. But they do not displace clinical judgment, professional accountability, or responsibility for patient outcomes. In practice, the performance gains attributed to AI are inseparable from expert oversight and careful integration into clinical teams.

We use AI as a capacity and quality multiplier, not as a substitute for our training. That line is difficult to walk and gets thinner as products improve, but the thought matters. Deploying it primarily as a justification for workforce reduction or skill substitution introduces avoidable risk to patient safety and physician trust.

Success ultimately go to those investing in workflow, governance structures, and building solution that define clear roles for both humans and machines. Radiology may be an early example, but the underlying lesson extends well beyond imaging.

The FDA has an Idea, or 10, on Good Machine Learning Practice

Good machine learning practices goes beyond the chip

The U.S. FDA, Health Canada, and the UK’s MHRA have unveiled 10 guiding principles for Good Machine Learning Practice (GMLP) in developing AI/ML medical devices. These principles aim to ensure safety, efficacy, and quality in healthcare innovation. Key focuses include leveraging multi-disciplinary expertise, implementing good software and security practices, ensuring representative clinical study participants and data sets, maintaining independence between training and test data sets, and emphasizing the performance of the human-AI team. These guidelines also highlight the importance of clear user information, robust testing, and ongoing monitoring of deployed models to manage re-training risks and maintain performance.

Read the full GMLP draft on the FDA website.

1-Minute Summary

Here are the ten principles of GMLP.

Continue reading

AI+Human Better than Human in Neurodegenerative Imaging

Recent research underscores a leap in neuroimaging accuracy for Alzheimer’s disease diagnosis, emphasizing the superior performance of AI-assisted radiologists over either AI or humans alone. This collaborative approach marries the meticulous precision of AI with the nuanced understanding of human experts, potentially setting a new standard in the detection of amyloid-related imaging abnormalities. Specifically, it demonstrated superior performance in detecting amyloid-related imaging abnormalities (ARIA), crucial for amyloid-β–directed antibody therapy. This synergy enhances diagnostic precision and underscores the potential of AI-enhanced radiological diagnostics to improve patient care significantly.

How will this synergy between AI and human intelligence redefine the future of medical diagnostics? Can this model be the blueprint for addressing other complex diseases? This breakthrough prompts us to envision a healthcare landscape where technology and human expertise converge to offer unparalleled patient care.

Detailed study can be found in JAMA Network Open.

Your Origin Story for Data Science

There’s an origin story for every superhero; even those without superpowers (like Batman – that’s right) got started somewhere. What we sometimes forget is that there is also an origin story for every regular person, every profession, every hobby.

Source: Wikimedia Commons

 

If you’re a radiologist looking to learn a few things in radiology data science, a simple web search will reveal a seemingly overwhelming amount of material you might have to know.

Fortunately, only a very small subset is necessary to start being productive.  Here are a few resources I used to get started.

Continue reading

The [machine learning] race is on – Don Dennison

Continue reading

Machine learning, real opportunites: Dr. Keith Dreyer’s keynote sets tone for ISC 2016

Dr. Keith Dreyer opens with a keynote during the Intersociety Summer Conference (ISC) with description of data science and overview of how machine learning have evolved over time.

He describes that machines and humans inherently see things differently. Humans are excellent at object classification, recognition of faces, understanding language, driving, and imaging diagnostics. Continue reading