In 2020, a study published in Nature Medicine showed that an AI system detected breast cancer in mammograms with greater accuracy than a panel of six radiologists. The researchers didn't present this as proof that radiologists were obsolete — quite the opposite. They saw it as evidence that AI and radiologists working together could catch more cancers than either working alone.
That framing — AI as a collaborator, not a replacement — is still the most accurate way to understand what's happening in radiology today.
What AI Actually Does in a Radiology Workflow
When most people hear "AI in radiology," they picture a computer looking at an X-ray and spitting out a diagnosis. The reality is more nuanced and, honestly, more interesting.
Modern AI systems in radiology typically function as triage tools and flagging systems, not autonomous diagnosticians. Here's what that looks like in practice:
A busy emergency department might receive 400 chest X-rays overnight. An AI system scans each image and flags the 40 that show potential pneumothorax (collapsed lung) or pneumonia. Those 40 go to the top of the radiologist's queue. The radiologist still reads all 400, but they're guaranteed to review the potentially critical cases first — even if they arrived in the queue last.
This isn't about replacing clinical judgment. It's about making sure clinical judgment gets applied to the right images at the right time.
The Three Areas Where AI Is Making a Real Difference
Oncology imaging is where AI has made the most measurable impact so far. Convolutional neural networks trained on millions of mammograms have demonstrated sensitivity rates above 90% for detecting early-stage breast cancer — comparable to experienced radiologists. Similar work is underway for lung nodule detection in CT scans, where the challenge is distinguishing malignant nodules from benign ones at sub-centimeter sizes. Google's DeepMind and Stanford's CheXNet project have both published results showing AI can match or exceed radiologist performance on specific detection tasks.
The important caveat: AI performs well on populations similar to its training data. A model trained on mammograms from one hospital system may not perform as well at a hospital serving a different patient demographic. This is an active area of research, and it's one reason why regulatory agencies require rigorous validation before clinical deployment.
Stroke detection is another area where AI's speed advantage is clinically meaningful. When a patient presents with stroke symptoms, time from onset to treatment directly affects outcomes — the phrase "time is brain" reflects how many neurons die per minute without treatment. AI tools that analyze CT perfusion images and immediately flag large vessel occlusions allow hospital teams to mobilize faster. Viz.ai's software, for example, is FDA-cleared and used in hundreds of hospitals specifically because it can get relevant information to a neurovascular team within minutes of a scan completing.
Workflow efficiency is less dramatic but arguably the most widely implemented application. AI tools that automatically measure bone density, calculate organ volumes, or pre-populate radiologist report templates don't make dramatic diagnostic decisions — but they save radiologists 15-30 minutes per case on routine measurements. In a department reading 300 studies per day, that compounds quickly.
What Still Challenges the Field
The technical challenges in AI radiology are real and worth understanding clearly.
Dataset quality and diversity remains the most fundamental problem. AI systems learn from labeled data — images that a radiologist has already read and annotated. Creating these datasets is expensive, time-consuming, and often biased toward patient populations from large academic medical centers. Models trained in Boston may not generalize to communities with different demographics, access to care patterns, or equipment.
Explainability is a genuine clinical concern. A radiologist making a diagnosis can point to specific findings in an image and explain their reasoning. Many deep learning systems produce a result without being able to articulate why — they generate "saliency maps" that highlight regions of interest, but these don't always align with what a clinician would actually focus on. For a doctor to trust an AI recommendation enough to act on it, some degree of explainability matters.
Regulatory pathways have become clearer since the FDA began developing specific frameworks for AI-based medical devices, but the process remains lengthy. Software that "locks" its algorithm after approval faces the question of how to handle updates as the model improves. Software that continuously learns raises different safety questions. The regulatory science is evolving in parallel with the technology.
Integration with clinical workflow is often underestimated. Even a technically excellent AI tool can fail in practice if it requires radiologists to switch between systems, if its alerts are too frequent and lead to alarm fatigue, or if results don't feed cleanly into existing reporting software. The best-performing tools in controlled studies don't always translate to the best-performing tools in actual hospital environments.
What This Means for Radiology as a Profession
The radiologists who are most engaged with AI right now are not the ones trying to figure out if AI will take their jobs. They're the ones asking how to use these tools to read more images with less fatigue, catch more early-stage findings, and spend their cognitive energy on the cases that genuinely require their expertise.
The workload of radiology has been growing faster than the number of trained radiologists for over a decade — imaging volume has increased while reimbursement has pressured departments to read more cases per day. AI tools that handle routine measurements, flag urgent findings, and streamline report generation address a real operational problem, independent of any concern about diagnostic superiority.
The emerging role isn't "radiologist" versus "AI" — it's "radiologist who uses AI effectively" versus "radiologist who doesn't." In hospitals that have deployed AI triage tools for stroke and pulmonary embolism, time-to-treatment has measurably improved. That's the kind of outcome that drives adoption, regardless of how the technology-versus-human debate resolves theoretically.
The field is still early. Most AI tools in clinical use today are narrow — they do one thing on one image type — and the path to general-purpose diagnostic AI remains unclear. But the trajectory is toward more AI in every imaging department, and the radiologists, technologists, and administrators who understand the technology's actual capabilities will be far better positioned to use it well.
Tags
Taresh Sharan