Detecting the Undetectable: AI in Neuroradiology Shows Promise—But Workflow Friction Persists

Medical Korea 2026 continues to highlight how artificial intelligence is reshaping clinical specialties at a granular level, with radiology emerging as one of the earliest and most data-rich domains for deployment. As imaging volumes surge globally—driven by aging populations, expanded screening programs, and increasingly complex diagnostics—radiologists face mounting pressure to deliver faster, more precise interpretations without compromising accuracy.

Within this context, neuroradiology has become a focal point for AI integration. In a session examining clinical applications and limitations, Professor Leonard Sunwoo of Seoul National University Bundang Hospital explored how AI is already augmenting diagnostic performance—while simultaneously introducing new forms of clinical and operational complexity.

Rather than presenting AI as a linear improvement, his discussion reflected a more nuanced reality: one where performance gains coexist with workflow friction, cognitive bias, and unresolved questions about how radiologists should interact with increasingly capable systems.

Leonard Sunwoo discussed how AI enhances early detection in neuroradiology while introducing new challenges in workflow efficiency and clinical decision-making. Image: GeneOnline

Closing the Diagnostic Gap: When AI Sees What Humans Miss

Sunwoo opened with a comparison that underscores one of AI’s most compelling advantages—its ability to detect patterns that may escape even experienced clinicians, particularly in early-stage disease.

He presented a case of hyperacute stroke in which subtle signal changes on MRI scans were overlooked by human readers but flagged by AI. Follow-up imaging days later confirmed a lesion in precisely that location, demonstrating the system’s sensitivity in identifying abnormalities at a stage where they are nearly imperceptible.

This example aligns with a broader trend in AI-assisted imaging, where algorithms trained on large datasets can recognize faint or diffuse signals that fall below the threshold of human perception, especially under time constraints common in emergency settings.

The implications become clearer at scale. In a study of 4,000 emergency stroke cases, radiologists achieved an already high diagnostic accuracy of 98.4%. However, approximately 50 cases were missed. When AI was applied retrospectively to those false negatives, it correctly identified 40 of them—suggesting that combining human expertise with AI assistance could push sensitivity close to 99.7%.

“AI is not replacing radiologists—but it is changing the ceiling of what is diagnostically possible,” Sunwoo said.

This incremental improvement may appear marginal numerically, but in clinical terms—particularly for time-sensitive conditions such as stroke—the difference can translate into significant patient outcomes.

The Hidden Risk: Automation Bias in Clinical Decision-Making

Yet as AI systems approach or surpass human-level performance in specific tasks, they introduce a less visible but equally critical challenge: how clinicians interpret and respond to machine-generated outputs.

Sunwoo highlighted the growing concern around automation bias, a psychological phenomenon in which users begin to over-rely on automated systems, particularly when those systems demonstrate high accuracy.

“As AI becomes more reliable, there is a risk that clinicians stop questioning it,” he said.

In ambiguous cases, this dynamic can become particularly pronounced. When a physician is uncertain, aligning with AI output may feel like the safer choice—both cognitively and legally—leading to what Sunwoo described as a “conformity effect.” Over time, this could allow AI errors to propagate unchecked, especially if clinicians hesitate to override system recommendations.

The concern extends to medical training. For residents and early-career radiologists, heavy reliance on AI markers may hinder the development of independent diagnostic judgment.

This reflects a broader shift in medical education, where the challenge is no longer just acquiring knowledge, but learning how to critically interact with decision-support systems.

The Efficiency Paradox: When AI Slows Things Down

One of the more counterintuitive findings discussed in the session challenges a central assumption driving AI adoption—that automation inherently improves efficiency.

In a large-scale study involving 60,000 CT scans for brain hemorrhages, the introduction of AI into the workflow actually increased the total time required for radiologists to finalize reports.

The reason lies in the distribution of cases. With only about 3% of scans showing actual hemorrhages, the vast majority of AI alerts occurred in negative cases. Radiologists were therefore required to spend additional time verifying false positives to ensure that no true abnormalities were overlooked.

“AI does not eliminate work—it redistributes it,” Sunwoo noted.

This phenomenon, sometimes referred to as “reversed efficiency,” illustrates a key limitation of current AI systems: while they can enhance sensitivity, they often do so at the cost of specificity, creating additional verification burdens.

For healthcare systems already grappling with workforce shortages, this raises important questions about how AI should be integrated—not just whether it should be used.

Redesigning the Workflow: Toward Human–AI Synergy

Addressing these challenges, Sunwoo emphasized that the future of AI in neuroradiology will depend on workflow design as much as technological advancement.

One proposed approach is the “blind first” method, in which radiologists conduct an initial assessment without AI assistance. Only after forming an independent judgment are AI results introduced, allowing discrepancies to be evaluated consciously rather than subconsciously influencing interpretation.

“The goal is not to remove AI—but to control when and how it enters the decision process.”

This approach reflects a broader shift toward human-in-the-loop systems, where AI serves as a secondary reviewer rather than a primary driver of decisions.

Explainability also plays a critical role. Current AI models often function as “black boxes,” providing outputs without clear reasoning. Sunwoo advocated for the development of explainable AI (XAI), including more interpretable visualizations such as heatmaps that allow clinicians to understand why a particular region was flagged.

At a systems level, he pointed to the need for policy adjustments. Because AI integration currently increases both workload and responsibility for radiologists, reimbursement models may need to evolve.

“If AI adds value and increases effort, that needs to be recognized in how we compensate its use.”

Beyond individual diagnostics, AI also presents opportunities at the population level. In emergency care, for example, it could be used to optimize patient transfers by identifying hospitals with available specialists or capacity—highlighting its potential role in broader healthcare system management.

Beyond the Algorithm: What This Means for Radiology in Korea and Globally

The discussion reflects a pivotal moment for radiology, both in South Korea and internationally. As one of the most digitized healthcare environments, Korea is well-positioned to lead in AI deployment, supported by advanced hospital infrastructure and strong data ecosystems. At the same time, it faces many of the same challenges seen globally—balancing innovation with regulation, efficiency with accuracy, and automation with clinical judgment.

At Medical Korea 2026, these tensions are becoming increasingly visible. The conversation is no longer centered on whether AI works, but on how it reshapes professional roles, workflows, and expectations within medicine.

Sunwoo’s presentation suggests that the future of neuroradiology will not be defined by AI alone, but by how effectively it is integrated into clinical practice. The goal is not full automation, but a redistribution of effort—where AI handles repetitive, data-intensive tasks, allowing radiologists to focus on higher-level interpretation and patient-centered decision-making.

“We are in a transition period,” he said. “The friction we see now is part of building a more sustainable system.”

In that sense, the current challenges—burnout, inefficiencies, and uncertainty—may not signal failure, but rather the early stages of a structural transformation that is still taking shape.

Author

Bernice Lottering