Radiology, with its structured data and repeatable workflows, is one of the easiest medical domains to automate with AI. And doing so would save money, free up radiologists to address critical, more complex cases, and reduce bottlenecks that prevent patients from receiving necessary care.
This is not a distant scenario. AI tools already on the market can interpret mainstream studies with super-human accuracy—most notably Oxipit’s ChestLink, the first autonomous AI cleared to report normal chest X-rays without humans in the loop. In multicentre data, ChestLink achieved up to 99.8 per cent sensitivity for critical findings, compared to 72–93 per cent for radiologists. For patients with routine cases, the choice between a highly capable AI system and a less accurate human reader is clear.
Once these numbers reach a Treasury Board desk, the moral case for “human eyes on every image” becomes a million-dollar line-item begging for deletion—especially when patients can reasonably ask why a less accurate reader should read their study instead of a proven, higher-accuracy one.
But automation doesn’t come without risk. Without proper guidance, it could drive away the skilled radiologists we still desperately need to clear the backlog of more complex cases. The question for policymakers is no longer whether automation will arrive, but how to effectively capture its savings while protecting the profession’s critical expertise.
High costs, stretched talent
Radiologists in Canada work in both private and public practices, with private imaging services expanding quite a bit since 2020. And it’s a high-volume business. Based on my last four years of practice in high-volume community radiology, roughly nine in ten outpatient X-rays are either clearly normal or so obviously abnormal that a junior trainee can draft the report in under a minute. And Canada performs millions of these each year. These “fast reads” cost taxpayers around $10 per reading in radiologist fees.
As it stands, our most decorated subspecialists still spend hours on routine portable chest films—cases that offer little intellectual challenge and could be safely automated with AI. And this can drive talented radiologists away. They’re leaving the Canadian healthcare system for places where they can better put their talents to use. Meanwhile, there remains a backlog of complex cases that could benefit from their expertise.
My concern about AI in radiology is not that radiologists will run out of work, but that the easy, high-volume X-ray cases are part of a time economy that can shift in either direction. In some settings, they consume hours that could otherwise go towards reducing CT, MRI, and ultrasound backlogs. In others, X-rays are already delayed or handled by trainees and batch-signed by staff so radiologists can prioritise higher-paying cases.
The trajectory is familiar. Nurse practitioners (NPs) began as physician extenders for rural primary care clinics. Now, NPs can safely provide 67-93 per cent of services in primary care settings generally. General practitioners (GPs) did not disappear, but their daily work shifted toward poly-morbid cases and end-of-life care—harder visits with identical pay. Burnout rose among GPs, and fewer trainees entered family medicine.
Without intervention, radiology could experience a parallel funding squeeze: the easy work goes, the complex work stays, but the compensation model doesn’t adjust. For governments and clinics to take full advantage of AI radiology, they need to rethink the funding model.
Policy playbook
With autonomous AI coming to radiology, the common question is whether Canada will shape its arrival or simply react to it. That starts with Health Canada, which should create a fast-track for autonomous AI in imaging — equivalent to the FDA’s “Breakthrough” designation program — not to cut corners on safety, but to shorten the lag between proven performance and public benefit.
From there, three groups need to act:
- Federal and provincial payers. Governments should not wait for perfect data before acting. The cost savings from real-world deployments of autonomous AI in imaging are already clear enough to justify pilot programs. Publishing whatever normal-versus-abnormal data is currently available—and gathering more prospectively—can refine these programs over time, rather than delay them. They should mandate full audit trails—tracking false negatives, algorithm drift, and bias—with public reporting at pilot implementation. And a fixed share of gross AI savings (say, 15 per cent) should be set aside for “Savings-for-Skills” programs, such as retraining radiographers for advanced vascular access, radiologists for short interventional fellowships, and nurses for AI safety oversight roles.
- Private chains and networks. Private imaging providers can turn AI into both a competitive advantage and a partnership tool. That means negotiating “profit-share glide paths” with vendors so that per-study AI fees decline over time, while the savings are shared with payers. They can also market guaranteed turnaround times—sub-five-minute reports for AI-cleared cases—as a premium service. And by using AI triage to move the ambiguous, high-value studies to the top of human worklists, they can protect relationships with clinicians while improving care.
- Radiologists and professional bodies. The profession adjusts by advocating for complexity-indexed billing—higher reimbursement for studies with high algorithmic uncertainty, just as surgery pays more for difficult cases. It also develops AI oversight fellowships that combine informatics, ethics, and regulation. And by publishing their own open guidelines for autonomous reads, fallback protocols, and incident reporting, radiologists can self-regulate in a way that pre-empts restrictive, one-size-fits-all legislation.
A use case, by the numbers
Let’s assume that Immigration, Refugees and Citizenship Canada orders roughly 25,000 chest X-rays each year for visa applicants. Historically, each image is read by a radiologist at about $10 per study—a $250,000 annual cost.
Phase A – Assisted Autonomy (Year 1)
A cautious rollout uses an AI system certified to clear normal films. Radiologists still glance at every normal image, billing $8 instead of $10, and bill the full rate for abnormal cases. The AI vendor is paid $1 for each normal film.
- Status quo cost: $250,000
- Pilot with AI: $200,500
- Annual payer savings: $49,500 — enough to fund a full first-year residency stipend.
- Radiologist workload: Almost nothing on normal cases, freeing capacity for complex studies.
Phase B – Stabilised Autonomy (Year 3+)
Once confidence grows and malpractice insurers are satisfied, normal films require only an institutional sign-off. Human review billing drops to $0.50 each.
- New cost: $14,875
- Annual savings vs. status quo: $235,125
- Radiologist revenue on this work: Down 93 per cent from baseline.
The bottom line is this: the technology will arrive whether we are ready or not. The only choice is whether its savings, speed, and safety are channelled into better care, or lost into the general budget.
