AI Aids in Detection but Fails to Reduce Workload
Artificial intelligence has made meaningful progress in radiology, particularly in detection. For example, products that examine long nodules improve sensitivity and have helped radiologists to identify findings that might otherwise be missed. At first glance, this seems like exactly what the field needs, as better detection should translate into faster reads, increased efficiency, and improved patient outcomes.
However, recent findings suggest something more complicated. A study examining AI-assisted lung nodule detection showed that while radiologists identified more findings, their interpretation time did not decrease. In other words, AI improved what radiologists could see, but not how quickly they could work. This result is surprising, but it reflects a broader reality in radiology today.
The Problem Isn’t Detection
Detection is only one step in the radiology workflow. Radiologists must still review prior studies, assess progression, interpret findings in a clinical context, and synthesize everything into a report. These steps are not automated, and most current AI tools do not meaningfully accelerate them. In fact, better detection can sometimes increase the workload, as more findings require more validation, comparison, and documentation. This is why AI can improve accuracy without improving efficiency; it enhances only one part of the process.
Volume and Complexity are Rising
Imaging volumes continue to rise across modalities, driven by advances in diagnostics and the expansion of preventive and value-based care. Radiologists are also managing a growing number of responsibilities tied to each study. Following the 21st Century Cures Act, reports are now released directly to patients, often immediately. This has increased the need for clarity in reporting and introduced new downstream tasks, including patient communication and follow-up. The result is not just more studies, but also more work per study. The systems radiologists rely on were simply not designed to support this level of complexity.
The Real Bottleneck: Reporting and Workflow
Most radiology systems were built in silos. PACS, reporting tools, and AI applications operate as separate layers, each with its own data model and workflow. As a result, radiologists spend significant time navigating between systems, reconstructing context, and manually assembling reports. Information does not flow naturally from images to interpretation to communication.
This explains why AI has not yet delivered on its promise to reduce workload. Many tools are added on top of existing infrastructure rather than integrated into it. AI cannot meaningfully reduce the effort required to complete a case, as there is no shared context across systems.
Rethinking Reporting
If radiology is to improve efficiency, the focus must shift. This means rethinking reporting.
Traditionally, reporting has been treated as the final step in the process, a place where findings are documented after interpretation. However, reporting sits at the center of the workflow, where imaging, clinical context, and communication converge.
A new generation of reporting systems is emerging, built around imaging data and workflow context rather than text alone. In this model, reporting is pixel-powered, which means the system starts with the images themselves, along with the clinical indication, prior studies, and relevant patient history. Prior exams can be automatically surfaced and summarized, and clinical context can be integrated directly into the reporting process. Voice interaction can move beyond dictation to trigger actions across the workflow.
Rather than simply generating text, the system begins to observe, reason, and act within the reporting process itself. This is the foundation of what can be described as agentic reporting.
Rethinking What AI Should Do
The takeaway from the lung nodule study is not that AI has failed, but rather that the problem has been misdefined. Improving what radiologists can see does not reduce the time required to interpret, and communicate those findings. Real gains will come from systems that understand context, connect workflows, and support reporting as an integrated process.