Between 2008 and 2019, radiologist workload in the United States rose by roughly 80 percent, while the number of radiologists reading cases with trainees grew by about 10 percent.[1] This is happening at a time when the country is facing a growing burden of chronic disease. According to the CDC, an increasing proportion of Americans are living with multiple chronic conditions: 42 percent have two or more, and 12 percent have at least five.[2] As chronic disease becomes more common, imaging volume has continued to rise. That has created a tremendous burden for radiologists, because each study now carries more work around it, from clinical history and prior exams to reporting requirements, follow-up communication, and documentation.
This is where AI may help radiologists manage the rising workload: for example, by helping to identify cases where there may be a high degree of confidence that no cancer is present. Radiologist time may then be reallocated to focus on borderline and potentially high-risk cases that may need immediate intervention.
A structured narrative review of AI in radiology and image-guided interventions reports that, across selected workflows and studies, AI was reported to achieve up to 94 percent segmentation accuracy, 95 percent nodule detection sensitivity, 30 to 75 percent scan time reductions, and 30 to 50 percent faster reporting.[3] However, according to a 2026 report, 76 percent of healthcare leaders say they are not prepared to deploy AI at the speed required.[4]
According to the analyst firm IDC, clinician trust is an important factor in AI adoption in healthcare. Providers rank real-time AI performance feedback highest, followed by access to explainable models and clinical validation evidence.[5]
I recently spoke with a radiologist who offered an interesting analogy. She said radiologists need to be able to trust AI the way they trust a resident. Her comment made me think about what trust actually looks like in clinical practice, and how it is built over time with both medical residents and new tools.
Attending radiologists apply varying degrees of scrutiny to reports generated by residents. That scrutiny depends on the individual radiologist, the case, and the judgment the resident has shown over time. Coaching also plays a role. Radiologists are more likely to build trust when residents demonstrate they can incorporate feedback, and over time, show consistent growth and improvement.
In short, trust is inherently multifactorial, and is built from many signals. That is why it is worth breaking down the habits, safeguards, and evidence that help radiologists decide whether a new tool deserves a place in clinical practice.
Start with work that can be checked quickly
In residency, trust often begins with work that can be reviewed quickly and corrected easily. That work may be shaped by department protocols, triage rules, or the attending’s judgment, but the principle is the same: early confidence is built in places where the work can be checked in the normal flow of care. AI is likely to be adopted in radiology on similar terms.
According to a 2025 study in JAMA Network Open, radiologists using a workflow-integrated generative model across 11,980 live clinical radiograph interpretations improved documentation efficiency by 15.5 percent, with no change in radiologist-rated clinical accuracy or textual quality.[6] That is why early use cases are likely to include report drafting, chart summarization, and retrieval of priors. They may save time, and the radiologist could judge their quality in the ordinary flow of work.
Be consistent
A resident earns trust through reliable work across many cases, and AI systems are likely to be evaluated in a similar way. In screening mammography, small shifts in performance have immediate consequences for patients and for workflow. Breast Cancer Surveillance Consortium data covering 2,301,766 examinations from 2011 through 2018 shows why consistency has to be established early. In that study, some radiologists called back far more patients than others, with recall rates ranging from 6.7 percent to 16.7 percent. Cancer detection also varied, from 2.2 to 9.4 cancers found per 1,000 examinations.[7]
In that kind of environment, even small differences in performance may influence how many patients are called back, how many cancers are found, and how much follow-up work moves through the system.
That is why steadiness matters. In a field where roughly 100 of every 1,000 women may be called back and about two to seven cancers are found, clinicians are more likely to trust systems that perform predictably across thousands of cases, not systems that look impressive on a handful of them.[8]
Make corrections count
Residents typically improve because someone reviews the work, points out what was missed, and expects the next case to be better. Clinical staff will expect the same discipline from AI. A recent interview study of 16 radiologists in the United States and Europe found that post-deployment monitoring of radiology AI is still immature, with the most common method being manual retrospective comparison between AI output and the final radiology report, while automated and statistical monitoring remain far less common.[9] That gap matters because trust weakens when corrections have no visible consequence. Clinicians need to know that edits, overrides, and rejected outputs feed monitoring and future improvement.
The evidence has to match the case
As confidence grows, the amount of explanation can shrink. A seasoned resident can present an answer briefly and expand when asked. A junior trainee usually has to show more of the reasoning. AI will face the same expectation.
A 2024 simulated-use study shows why the evidence has to match the case. In the study, 20 radiologists from seven Dutch medical centers reviewed lung nodule assessments with AI recommendations available. After seeing the AI output, radiologists changed the number of nodules they reported in 27 of 140 assessments, changed their malignancy predictions in 32 of 140 assessments, and changed follow-up advice in 12 of 140 assessments.[10] In other words, AI did not simply sit in the background and was observed to influence clinical judgment in ways that could affect what gets reported, how risk is understood, and what happens next for the patient.
That is why radiologists need more than an answer. They typically need to see sufficient supporting evidence to decide whether the recommendation fits the image, the clinical context, and the level of risk. What clinicians need is support that matches the case, with enough evidence to inspect the result and more depth when the stakes are higher.
Responsibility stays with the radiologist
A resident can contribute substantially to the work, and the attending still owns the report. The same line of responsibility should be maintained once AI enters the workflow. A 2024 RSNA and MICCAI expert report argued that radiologists must be able to trust a system’s design, receive adequate training, and work within clear lines of clinical accountability before AI can be safely integrated into practice.[11]As AI takes on a larger role in report drafting, triage, and workflow support, the final interpretation remains the responsibility of the physician.
The most promising vision of radiology AI is a connected workflow in which image analysis, chart summarization, prior retrieval, report drafting, advanced visualization, and downstream clinical preparation move together rather than as separate tasks. That is the kind of system many health systems are seeking, because it reduces the small delays and repeated handoffs that accumulate across the day.
Trust will be built there, in the ordinary discipline of the reading room, where a tool proves that it may save time, fits cleanly into practice, holds up over repeated use, and responds to correction. If AI can meet that standard, radiology may offer one of the earliest examples of serious clinical integration, with the potential to ease pressure in the places where pressure is greatest and support confidence through dependable use.
[1] Burns J, et al. “Evolving Trainee Participation in Radiologists’ Workload Using a National Medicare Dataset.” Journal of the American College of Radiology, 2024. Public summary: https://radiologybusiness.com/topics/healthcare-management/healthcare-staffing/rising-workloads-spur-academic-radiologists-spend-less-time-training-residents
[2] Centers for Disease Control and Prevention. Benavidez GA, Zahnd WE, Hung P, Eberth JM. “Chronic Disease Prevalence in the US: Sociodemographic and Geographic Variations by ZIP Code Tabulation Area.” Preventing Chronic Disease, 2024. https://www.cdc.gov/pcd/issues/2024/23_0267.htm
[3] Friebe M. “AI in radiology and interventions: a structured narrative review of workflow automation, accuracy, and efficiency gains of today and what’s coming.” International Journal of Computer Assisted Radiology and Surgery, published online November 17, 2025; issue year 2026. https://link.springer.com/article/10.1007/s11548-025-03547-2
[4] Incredible Health. “Healthcare Leaders Are Betting on AI, While 76% Say Their Organizations Can’t Keep Up.” Business Wire press release, March 25, 2026. https://www.businesswire.com/news/home/20260325565819/en/
[5] IDC Data
[6] Huang J, Wittbrodt MT, Teague CN, et al. “Efficiency and Quality of Generative AI-Assisted Radiograph Reporting.” JAMA Network Open, 2025. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2834943
[7] Breast Cancer Surveillance Consortium. “Performance Measures for 2,301,766 Screening Mammography Examinations from 2011–2018.” https://www.bcsc-research.org/index.php/statistics/screening-performance-benchmarks/performance-measures
[8] DenseBreast-info.org. “For Providers: Cancer Detection by Screening Method in Dense Breasts.” https://densebreast-info.org/screening-technologies/cancer-detection-by-screening-method/
[9] Chow J, Lee R, Wu H. “How Do Radiologists Currently Monitor AI in Radiology and What Challenges Do They Face? An Interview Study and Qualitative Analysis.” Journal of Imaging Informatics in Medicine, published online April 8, 2025; issue year 2026. https://link.springer.com/article/10.1007/s10278-025-01493-8
[10] “The Impact of Expectation Management and Model Transparency on Radiologists’ Trust and Utilization of AI Recommendations for Lung Nodule Assessment on Computed Tomography: Simulated Use Study.” JMIR AI, 2024. https://ai.jmir.org/2024/1/e52211
[11] Radiological Society of North America. “Experts Outline Considerations to Deploy AI in Radiology.” July 10, 2024. https://www.rsna.org/media/press/i/2517
