Healthcare systems are structurally unprepared for radiology automation

For over a decade, medical imaging has been dominated by a recurring narrative that AI is coming for radiology. It goes something like this: “The radiologist is an endangered species” and “The scan reads itself now.”

Spending over nearly twenty years in the field, I have built AI imaging tools, taken them through FDA clearance and CE marking, deployed them in clinical trials with major pharma companies, and seen them used in real hospitals with real patients. And I want to be honest about what I see, not to comfort anyone, and not to generate alarm, but because the reality is more complex and more interesting than either camp is willing to admit. Let me start with what is often left unsaid.

AI is not hitting a ceiling

The standard reassurance you hear from radiologists, and from well-meaning technologists who want to avoid conflict, is that AI is a “tool,” a “co-pilot,” something that assists but never replaces. Clinical judgment, empathy, and contextual reasoning: these are the moats. The human stays in the loop. That view is becoming harder to defend without serious qualification.

The trajectory of AI in medical imaging is not plateauing. Year on year, models are becoming more accurate, more generalizable, and more capable of detecting findings that human eyes miss. In MRI for prostate cancer detection and diagnosis, in liver fibrosis quantification, and in neuroimaging for the analysis of degenerative disease, we are watching AI close the gap with senior specialist performance and, in some narrow tasks, exceed it. This is not a projection. It is already happening in peer-reviewed literature, and I see it in our own data at our organization.

More importantly, the rate of improvement is not slowing down. Multimodal models that integrate imaging with genomics, proteomics, and clinical history. Foundation models trained on millions of scans. Self-supervised architectures that do not require radiologist annotations to learn. The assumption that there is always going to be a hard ceiling where AI becomes “good enough to assist but not good enough to replace” is not supported by the current evidence curve. 

Within five to ten years, a significant portion of today’s routine cognitive workload in radiology will be automatable to a high standard, and there is no strong scientific reason to assume it will not be.

What AI does well, and where human oversight remains essential

This is where the nuance matters, because “automatable tasks” is not the same as “the profession disappears.”

AI excels at consistent pattern recognition at scale: detecting anomalies, performing precise quantification, comparing with priors, and flagging urgent findings without fatigue. These strengths directly address today’s radiologist shortages and growing imaging volumes. In structured, well-defined tasks with large training datasets, AI is likely to outperform humans. That is the honest assessment.

What AI cannot yet do, and where the real durability of the human expert lies, is navigate genuine uncertainty, carry clinical accountability, integrate findings into a patient’s broader narrative, and make judgment calls where the right answer is not in the training data. The 84-year-old patient with three comorbidities, a contested prior report, and a family asking hard questions: that is not a classification problem.

But this needs careful qualification. The list of things AI “cannot do” has been getting shorter, not longer. Every year, something that was supposedly beyond machines becomes automatable. The list of tasks once considered “uniquely human” continues to shrink. Professional identity should not be built around tasks that may become commoditized.

The job title might change, and that is fine

Here is what is actually happening, and what nobody in the medical establishment wants to say directly: the word “radiologist,” as we know it, may not be the right label for what the expert of 2035 or 2040 actually does.

The radiologist of the future is closer to an imaging data scientist, combining clinical knowledge with technical oversight. They set the parameters within which AI systems operate. They validate outputs and catch edge cases. They integrate quantitative imaging data (biomarkers, phenotype characterizations, longitudinal measurements) into clinical decision-making that no single model can handle end to end. They carry the legal and ethical accountability for AI-assisted diagnoses together with AI manufacturers. And they evolve the science: designing studies, building evidence, and pushing the frontier of what imaging can reveal about disease.

That is a genuinely interesting and important job. It is arguably more intellectually demanding than reading a hundred routine chest CTs in a row. But it requires a very different training, a very different set of skills, and a very different professional identity than what radiology residencies currently produce.

We are not preparing people for that job. We are still largely training radiologists to do what AI will do better, while paying lip service to “AI integration” as an elective skill.

What this means for healthcare systems right now

The danger is not the future. The danger is the gap between where AI capability already is and how healthcare institutions are deploying it.

On one side, you have genuinely regulated, rigorously validated AI tools: FDA-cleared, CE-marked medical devices with clinical evidence behind them, built by people who understand what it means to be accountable for a diagnostic output. On the other side, you have a growing wave of lightweight tools deployed by consulting firms and technology integrators that have no regulatory status, no clinical validation, and no clear answer to the question: who is responsible if this goes wrong? For hospital leaders and healthcare operators, this raises immediate questions around governance, accountability, and the procurement of AI tools.

A call for honesty

The medical imaging community would benefit from two things it has been slow to embrace: honesty about the real trajectory of AI capability, and a serious commitment to redesigning the radiology profession for the future.

Telling radiologists that AI is just a tool and their job is safe as long as they “embrace technology” is not reassuring. It is misleading. The trajectory is real, and pretending otherwise does not protect anyone; it just delays adaptation.

The more useful message is this: the underlying clinical expertise, the anatomical knowledge, the understanding of disease, and the ability to carry responsibility – none of that becomes worthless. But the expression of that expertise is going to change dramatically. The professionals who will thrive are not the ones who resist AI, nor the ones who defer entirely to it. They are the ones who understand it well enough to direct it, challenge it, and ultimately own its outputs.

The role will evolve whether the profession is ready or not. The question is not whether AI changes radiology, but how quickly institutions adapt to that reality.