Radiologists Urge Economic Realism in AI Adoption

With no one-size-fits-all ROI, local workflows and scale drive AI's payoff

BY CINDY ZINKOVICH May 26, 2026

Molwicz

Siegel

AI can create measurable economic value in radiology in certain circumstances. However, scientific evidence on cost-effectiveness is sparse, incomplete and often unreflective of real-word conditions and patient populations, according to a systematic review and related commentary published in Radiology: Artificial Intelligence.

“There’s no universal AI dividend—economic value depends strongly on where and how the technology is implemented,” said lead author Isabel Molwitz, MD, MBA, attending radiologist in the Department of Diagnostic and Interventional Radiology and Nuclear Medicine at University Medical Center Hamburg Eppendorf, in Germany. “AI in radiology is a selectively powerful enabler. Its true value emerges when technical excellence meets clinical need and sound economic evaluations. Without that intersection, even brilliant algorithms can become expensive gadgets.”

Imaging providers and payers must demand robust, transparent cost-effectiveness analysis, rather than relying on vendor marketing or time savings claims, Dr. Molwitz said. She noted that of 1,879 AI studies published over a nearly 15-year period, only 1% met review inclusion criteria by explicitly quantifying economic outcomes.

Most of the included studies assessed pixel-focused tools; nearly half involved machine learning and one-third looked at computer-assisted diagnosis. In addition, 10% were modeling studies, which often show lower costs than real-word data, Dr. Molwitz said.

According to Dr. Molwitz and colleagues, the same algorithm can have very different financial outcomes depending on task complexity, examination volume, reimbursement structure, staff costs and willingness-to-pay thresholds for quality of health year gains.

“AI in radiology is a selectively powerful enabler. Its true value emerges when technical excellence meets clinical need and sound economic evaluations. Without that intersection, even brilliant algorithms can become expensive gadgets.”

— ISABEL MOLWITZ, MD, MBA

To make those differences easier to evaluate and compare, the team calls for standardized frameworks, implementation of medical outcomes, ongoing surveillance and interdisciplinary collaboration, including medical and economic experts. “A deep learning model that is economical in a national screening program may not be cost effective in a small private practice,” Dr. Molwitz said.

The commentary authors concur. “It is time for the radiology community to transition from unchecked AI optimism to economic realism,” said Eliot Siegel, MD, professor and vice chair in the Department of Diagnostic Radiology and Nuclear Medicine at the University of Maryland School of Medicine in Baltimore.

Dr. Siegel co-authored the commentary with Alireza Amindarolzarbi, MD, a radiology resident at the University of Maryland School of Medicine. “Current economic models are dangerously narrow and largely ignore the total cost of ownership, including the massive hidden costs of IT integration, cybersecurity, clinician training and local data validation, not to mention the equity cost when models perpetuate algorithmic bias,” Dr. Siegel said.

AI concept image of hand with bag of currency floating above it and other AI symbolism floating around the bag.

When Is AI Creating Value?

Dr. Siegel emphasized that the fastest returns often come from workflow gains, not pixel-level tools. “We must expand our definition of value beyond simple financial metrics to include workforce sustainability, reduced diagnostic errors and truly equitable patient care,” he said.

Drs. Molwitz and Siegel also cautioned against performance drift, a term used to describe when commercial algorithms’ accuracy declines over time across different settings with distinct patient demographics and scanner protocols.

“Large healthcare networks able to create or fine-tune AI models using their own internal datasets will likely improve cost-effectiveness and clinical reliability,” Dr. Siegel said. “However, health disparities will worsen if under-resourced and rural centers are left behind, or if they are forced to rely on commercial models that have been shown to selectively underdiagnose marginalized groups.”

Dr. Molwitz and colleagues encourage strategic AI adoption. Recommendations include running pilot evaluations that track throughput, recall rates and downstream revenue or savings, pursuing fixed price over pay-per-use contracts and monitoring real world performance over time.

“AI tends to be cost effective in high volume, resource intensive workflows such as lung cancer or tuberculosis screening, and when its diagnostic accuracy meets or exceeds human performance,” she said. “Cost savings are most likely when AI replaces constrained staff resources or reduces follow-up imaging through better accuracy and compliance tracking.”

AI provided the clearest economic benefit in a few clinical scenarios, including opportunistic CT analyses (e.g., osteoporosis or cardiovascular risk), MRI acceleration tools that boost scanner capacity, and follow-up tracking systems that generate reimbursable additional imaging.

“For high-volume, high-stakes triage—for example when AI operates in a critical pathway such as acute stroke detection—its value is undeniable,” Dr. Siegel said. “By rapidly detecting large-vessel occlusions, AI accelerates time-to-treatment, directly improving quality-adjusted life years and generating measurable and significant downstream cost savings for the healthcare system.”

The most immediate financial benefit lies outside of pixel-based diagnosis, Dr. Siegel emphasized, encouraging radiologists to prioritize workflow-enhancing tools such as large language models (LLMs) that offer immediate, direct and quantifiable returns.

“Applied to the text-based, non-interpretive parts of our day, such as summarizing complex prior reports or auto-drafting routine negative reports, LLMs save time, increase throughput and directly combat radiologist burnout,” Dr. Siegel said. “By automating the high-volume, repetitive and administrative aspects of our work, AI allows us to spend more time on complex diagnostic problem solving, multidisciplinary consultation and direct patient interaction.”

Will AI Become Essential Infrastructure?

“While the cost-effectiveness of standalone algorithms is a pressing concern today, this is a transient phase,” Dr. Siegel predicted. “Capabilities like AI-powered triage and basic lesion detection will become standard, baseline features built directly into our reporting systems and PACS.”

“Today, vendors and purchasers need reimbursement, so the cost-effectiveness of each solution is vital,” Dr. Molwitz said. “Even if AI solutions are fully integrated into the radiological workflow one day, someone needs to pay. And from their perspective, cost-effectiveness will still matter.”