Mitchell Katz doesn’t mince words. The CEO of NYC Health + Hospitals — the largest public hospital system in the United States, serving more than a million patients annually across 11 acute care hospitals and dozens of clinics — says he’s ready to replace radiologists with artificial intelligence. Not supplement them. Not assist them. Replace them.
“I am ready right now to replace radiologists with AI,” Katz told an audience at a health conference, as reported by Slashdot. The statement landed like a grenade in the medical community, triggering fierce debate among physicians, hospital administrators, and AI researchers about whether the technology is truly ready — and whether the economics of American healthcare will force the issue regardless.
Katz’s argument is grounded in something deeply practical: access. NYC Health + Hospitals is a safety-net system. It treats the uninsured, the undocumented, the poorest residents of New York City. It operates on thin margins and faces chronic staffing shortages. Radiologists are expensive and hard to recruit, particularly for a public system that can’t match private-sector compensation. If an AI model can read a chest X-ray or CT scan with accuracy comparable to or exceeding a human radiologist, Katz’s calculus is straightforward. Why not use it?
The question is whether that “if” has actually been satisfied.
Where the Technology Actually Stands
AI’s performance in medical imaging has improved dramatically over the past five years. Models trained on millions of labeled images can now detect certain conditions — lung nodules, breast cancer on mammograms, diabetic retinopathy — with sensitivity and specificity that rival experienced radiologists in controlled studies. The FDA has cleared hundreds of AI-enabled medical devices, many of them in radiology. Companies like Aidoc, Viz.ai, and Rad AI have built substantial businesses around augmenting radiologists’ workflows, flagging critical findings, and prioritizing worklists.
But augmentation and replacement are very different things.
Radiologists don’t just read images. They correlate findings with clinical history. They communicate with referring physicians. They perform and supervise procedures — biopsies, drains, angiograms. They handle ambiguity, the cases where the image doesn’t clearly say yes or no, where judgment and experience matter enormously. Current AI systems excel at pattern recognition within narrow, well-defined tasks. They struggle with the integrative reasoning that defines much of radiology practice.
Dr. Curtis Langlotz, a professor of radiology and biomedical informatics at Stanford, has been studying AI in medical imaging for years. He’s noted repeatedly that while AI will transform radiology, the notion of wholesale replacement oversimplifies what radiologists do. The technology works best, he and others have argued, when it handles the routine — freeing human radiologists to focus on complex cases, quality assurance, and clinical consultation.
Katz seems aware of these nuances but appears willing to push the boundary further than most health system leaders. His public posture suggests a belief that the technology curve is steep enough, and the staffing crisis acute enough, that waiting for perfect AI means accepting preventable harm from delayed diagnoses.
There’s a certain logic to this in the safety-net context. When you don’t have enough radiologists to begin with, the comparison isn’t AI versus a well-rested, fellowship-trained subspecialist. It’s AI versus a scan sitting unread for days. Or weeks. In that framing, even imperfect AI looks attractive.
Still, the legal and regulatory framework isn’t set up for radiologist-free imaging. The practice of medicine requires a licensed physician. AI systems currently operate under physician supervision — the FDA’s clearances are for decision-support tools, not autonomous diagnostic agents. Removing radiologists from the loop would require regulatory changes that no state or federal agency has signaled it’s prepared to make anytime soon.
And then there’s liability. Who’s responsible when an AI misses a tumor? The hospital system? The software vendor? The referring physician who relied on the AI’s output? These aren’t hypothetical questions. They’re the kind of issues that keep hospital general counsels awake at night and that plaintiff’s attorneys would love to test in court.
The American College of Radiology has pushed back firmly against the replacement narrative for years, arguing that AI is a tool for radiologists, not a substitute. The organization has invested heavily in data registries and quality initiatives designed to demonstrate the irreplaceable value of physician oversight. Radiologists themselves have watched the “AI will replace you” predictions cycle through since at least 2016, when Geoffrey Hinton, the godfather of deep learning, famously said medical schools should stop training radiologists because AI would outperform them within five to ten years. That deadline has come and gone. Radiology residency programs remain full. Radiologist salaries remain high.
But Katz’s comments carry a different weight than an academic’s prediction. He runs a system with a $12 billion annual budget and 43,000 employees. He’s not speculating about the future from a conference stage for sport. He’s signaling operational intent — or at least a willingness to move aggressively if the regulatory environment permits.
The economics are hard to ignore. A full-time radiologist in the U.S. commands total compensation between $350,000 and $600,000 or more, depending on subspecialty and geography. A large hospital system might employ dozens. AI software licenses, even at enterprise scale, cost a fraction of that. The math gets even more compelling when you factor in overnight coverage, where teleradiology firms charge premium rates for after-hours reads that an AI system could theoretically handle around the clock without fatigue or shift differentials.
Private equity has noticed. Investment in healthcare AI has surged, with radiology remaining one of the most active sectors. Investors see the combination of high labor costs, measurable performance benchmarks, and massive data availability as ideal conditions for AI disruption. Whether that disruption means replacing radiologists or simply restructuring their work — fewer radiologists doing more with AI assistance — remains the central question.
Some radiologists are already adapting. Younger physicians entering the field increasingly view AI fluency as a core competency, not a threat. They’re training with AI tools, publishing research on human-AI collaboration, and positioning themselves as the physicians who can interpret what the algorithms produce and catch what they miss. The radiologist of 2035, many in the field believe, won’t look like the radiologist of 2015 — but there will still be a radiologist.
Katz may ultimately prove right that AI can handle a substantial portion of routine imaging interpretation. Chest X-rays for pneumonia. Screening mammograms. CT scans for pulmonary embolism. These are high-volume, pattern-heavy tasks where AI has demonstrated strong performance. But the full scope of radiology — the interventional procedures, the tumor boards, the phone calls to emergency physicians at 2 a.m. explaining why a finding changes the surgical plan — that’s harder to automate than a conference soundbite suggests.
What’s most significant about Katz’s statement isn’t the technology assessment. It’s the willingness of a major health system CEO to say the quiet part out loud. Hospitals across the country are struggling with workforce shortages, rising costs, and growing patient volumes. AI offers a potential release valve. And the leaders of these institutions, even if they couch it in more diplomatic language than Katz, are thinking the same thing.
The radiologists who thrive will be the ones who make themselves indispensable in ways that algorithms can’t replicate. The ones who don’t may find that Mitchell Katz wasn’t an outlier — he was just early.
WebProNews is an iEntry Publication