Human-AI Collaboration in Radiology: From Artificial Intelligence to Augmented Clinical Intelligence

From Detection to Decision: How Radiologists and Artificial Intelligence Can Work Together Safely and Effectively

Edited by ScholarGen AIHealthCareInsight

Human-AI Collaboration in Radiology: Why the Future Is Not Human vs AI

A radiologist reviewing a chest CT does not simply search for a pulmonary nodule.

The radiologist asks a series of interconnected questions.

Is the finding real? Where exactly is it located? Is it new? Has it changed compared with a previous examination? Could it represent malignancy, infection, inflammation, or a benign process? Does the patient's clinical history alter the interpretation? What additional imaging or clinical evaluation is justified?

Artificial intelligence can contribute to several of these tasks. It can detect abnormalities, segment anatomical structures, quantify measurements, compare examinations, prioritize studies, and generate preliminary report content.

But these capabilities do not eliminate the central clinical problem.

Radiology is not merely image recognition.

It is clinical reasoning under uncertainty.

This distinction explains why the most meaningful future of medical imaging may not be autonomous AI replacing radiologists, but human-AI collaboration in which computational systems and physicians contribute different forms of intelligence.

AI is exceptionally good at repetitive computation, pattern recognition at scale, quantitative measurement, and processing large volumes of imaging data. Radiologists contribute contextual reasoning, clinical integration, uncertainty management, communication, accountability, and the ability to recognize when the apparent imaging problem is not actually the patient's most important problem.

The question, therefore, is changing.

Instead of asking:

“Can AI outperform a radiologist?”

the more clinically useful question is:

“Under what conditions can a radiologist and AI together produce a better clinical decision than either could produce alone?”

That is the central question of Human-AI Collaboration in Radiology.

1. What Is Human-AI Collaboration in Radiology?

Human-AI collaboration in radiology refers to a clinical workflow in which artificial intelligence and radiologists interact during image interpretation, rather than functioning as completely independent systems.

The AI may provide:

abnormality detection
image classification
segmentation
quantitative measurements
image prioritization
comparison with prior examinations
structured findings
preliminary report suggestions
clinical decision-support information

The radiologist then evaluates these outputs in the context of the complete examination and the patient's clinical situation.

The essential concept is augmentation rather than substitution.

A well-designed system should allow the radiologist to accept, modify, reject, or investigate an AI suggestion. The final interpretation remains a clinical judgment rather than an automatic transcription of the algorithm's output.

This distinction becomes particularly important when AI confidence does not correspond to clinical correctness.

An AI model may produce a highly confident output for a case that lies outside its training distribution. A radiologist may recognize that the examination contains unusual anatomy, postoperative change, motion artifact, an uncommon disease pattern, or a technical acquisition problem that makes the AI output unreliable.

Human expertise therefore acts not only as a decision-maker but also as a safety layer around algorithmic inference.

2. Why Is Radiology Particularly Suitable for Human-AI Collaboration?

Radiology generates enormous amounts of structured and visual information.

A single CT examination may contain hundreds or thousands of image slices. MRI examinations can contain multiple sequences, planes, diffusion measurements, contrast-enhanced phases, and quantitative parameters.

This creates a natural environment for computational assistance.

AI can examine large image volumes rapidly and consistently. It can also perform measurements that would otherwise require repetitive manual work.

For example, AI may assist with:

Task	Potential AI Contribution	Human Contribution
Detection	Identify suspicious regions	Confirm whether the finding is clinically real
Segmentation	Outline organs or lesions	Assess anatomical validity
Quantification	Calculate volume, diameter, density or other parameters	Interpret clinical significance
Triage	Prioritize potentially urgent examinations	Confirm urgency and determine appropriate action
Comparison	Detect interval change	Determine whether change is meaningful
Reporting	Suggest structured findings	Produce the final clinically coherent report
Decision support	Present possible diagnostic considerations	Integrate history, examination and laboratory data

The strength of this model is not that AI performs every task.

It is that AI and radiologists can perform different tasks according to their comparative strengths.

3. What Does AI Do Better Than Humans?

AI systems have several characteristics that make them attractive clinical partners.

Scale

An algorithm can process thousands of images without fatigue.

Consistency

A model can apply the same computational operation repeatedly.

Quantification

AI can calculate measurements that may be cumbersome to perform manually.

Pattern recognition

Deep learning systems can identify complex imaging patterns that are difficult to express as conventional rules.

Prioritization

AI can help identify examinations that may require earlier attention.

Longitudinal analysis

Computational tools can potentially compare current and prior examinations systematically, reducing the cognitive burden of searching for subtle interval changes.

These strengths are particularly valuable in high-volume environments.

However, speed and consistency should not be confused with clinical understanding.

An algorithm can identify a region of interest without understanding why that region matters to the patient.

That distinction remains fundamental.

4. What Does the Radiologist Do Better?

Radiologists do not interpret images in isolation.

They integrate multiple information streams.

A CT finding is interpreted alongside symptoms, age, laboratory results, previous examinations, pathology, treatment history, operative history and the reason for examination.

Consider a pulmonary opacity.

An AI system may classify it as a potential consolidation.

The radiologist must determine whether the opacity is actually:

infection
atelectasis
hemorrhage
pulmonary edema
treatment-related change
organizing pneumonia
malignancy
aspiration
another process

The correct interpretation may depend on information that is not contained in the image itself.

The radiologist also determines when the algorithm should not be trusted.

This ability to recognize uncertainty is one of the most important components of human-AI collaboration.

5. The Concept of Diagnostic Complementarity

The most interesting possibility is not that AI becomes better than humans or humans become better because of AI.

It is that their errors may be different.

Suppose a radiologist detects abnormalities that an AI system misses, while the AI detects subtle findings that the radiologist overlooks.

If the two systems are sufficiently complementary, combining them may improve overall diagnostic performance.

This is known as diagnostic complementarity.

However, complementarity cannot be assumed simply because an algorithm performs well in a validation dataset.

The key question is whether the AI's errors are sufficiently independent from human errors.

If both human and AI systems make the same error for the same reason, collaboration may provide little additional value.

Conversely, if AI identifies a different subset of subtle abnormalities while the radiologist provides contextual correction, the combination may be substantially more useful.

Recent research on human-AI interaction in radiology has emphasized this distinction between evaluating an algorithm independently and evaluating the actual interaction between clinicians and AI in clinical workflow.

The collaboration itself is therefore a clinical technology that requires evaluation.

6. Does AI Always Improve Radiologist Performance?

No.

This is one of the most important conclusions for clinical implementation.

AI assistance can improve diagnostic performance, but its effect is heterogeneous.

Different radiologists may respond differently to the same AI output. The result can depend on the task, the quality of the AI prediction, the user's experience, the interface, the order in which information is presented, and the complexity of the case.

A large study examining AI assistance across chest radiograph tasks found substantial variation in how individual radiologists benefited from AI. Importantly, inaccurate AI predictions could adversely affect radiologist performance. This finding illustrates why “AI-assisted” should never automatically be interpreted as “AI-improved.”

The clinical unit of evaluation should therefore be:

AI + Radiologist + Workflow + Patient

rather than:

AI alone

7. What Is Automation Bias?

Automation bias occurs when people place excessive trust in an automated system and reduce their independent evaluation.

In radiology, this may happen when an AI system displays a lesion candidate and the radiologist unconsciously gives that finding more weight than the image itself deserves.

The opposite phenomenon can also occur.

A radiologist may distrust an AI system even when it provides useful information.

This is sometimes described as algorithmic aversion.

Both extremes are problematic.

The appropriate relationship is neither:

“AI is always correct.”

nor:

“AI cannot be trusted.”

Instead:

“AI provides evidence that must be clinically evaluated.”

A 2026 review specifically highlighted automation bias and overconfidence as emerging concerns as AI becomes more common in clinical radiology.

The solution is not simply to tell radiologists to “trust AI less.”

The better approach is to design systems that support appropriate skepticism.

8. Should the Radiologist See the AI Result Before Reading the Images?

The answer is not universally established.

The order of human-AI interaction can influence diagnostic behavior.

In experimental studies, different collaboration protocols have produced different outcomes. Research involving radiologists and other clinical readers has shown that the sequence in which human and AI information is presented can affect performance.

This creates an important interface-design question.

Should the workflow be:

Human → AI → Reassessment

or:

AI → Human → Verification

or perhaps:

Human and AI in parallel → Integrated decision

There is no single optimal sequence for every imaging task.

For some high-risk applications, independent human interpretation before AI exposure may reduce anchoring. For other applications, AI-first triage may improve efficiency.

Therefore, the interaction protocol itself should be validated rather than selected simply because it is technically convenient.

9. AI in the Radiology Workflow: Where Should It Enter?

AI can potentially participate throughout the imaging pathway.

The most effective implementation may not place AI only at the final diagnostic stage.

For example, AI may assist with:

protocol selection
image quality assessment
reconstruction
worklist prioritization
incidental finding detection
lesion segmentation
quantitative analysis
reporting
follow-up recommendation support

This means that the future radiology department may be better understood as an AI-enabled information system rather than a workstation with an AI button.

10. AI-Assisted Detection: A Second Pair of Eyes

One of the most established conceptual applications is AI as a second reader.

The system searches for predefined abnormalities and highlights suspicious regions.

Chest radiography provides a useful example.

A radiologist may miss a subtle pulmonary nodule because it overlaps a rib, clavicle, heart, or diaphragm. An AI system may flag the region and prompt closer inspection.

Importantly, the AI finding is not the diagnosis.

It is a visual cue.

The radiologist still needs to determine whether the opacity corresponds to a genuine pulmonary lesion and whether it has clinical significance.

Studies of AI-assisted chest radiography have reported improvements in sensitivity and reading efficiency under specific experimental conditions, although the magnitude of benefit varies by pathology and reader. One study reported a reduction in mean reading time from 81 seconds to 56 seconds with AI assistance, while specificity did not improve uniformly across all abnormalities.

This illustrates a broader principle:

Improving one performance metric does not necessarily mean improving the entire diagnostic process.

11. AI as a Quantitative Partner

One of AI's most valuable roles may be measurement rather than diagnosis.

Radiologists routinely perform measurements such as:

lesion diameter
organ volume
tumor burden
ventricular volume
vascular dimensions
bone density
lung volume
fat distribution

Automated quantification can improve reproducibility and reduce repetitive manual work.

The clinical value becomes greater when measurements are performed longitudinally.

For example:

Current volume – Previous volume → Quantitative change

This may help transform qualitative impressions into objective trajectories.

But automated measurement still requires validation.

Segmentation errors can occur because of:

postoperative anatomy
infiltrative disease
adjacent structures
poor contrast
motion
artifacts
unusual anatomy

The radiologist must therefore understand not only the numerical result but also how the number was produced.

12. Generative AI and Radiology Reporting

The next phase of collaboration increasingly involves generative AI and large vision-language models.

These systems may assist with:

structured reporting
draft report generation
summarization
comparison with prior examinations
terminology normalization
communication of findings
extraction of relevant clinical information

The key word is draft.

A generated report is not equivalent to a signed medical report.

Recent proof-of-concept work has examined collaborative large vision-language model assistance for chest radiograph reporting. In that study, AI assistance produced efficiency gains particularly for complex reports, while the radiologists retained the ability to accept, modify, or reject suggestions.

This interaction model is clinically more appropriate than fully automatic report generation because it keeps the physician inside the decision loop.

13. Why Report Generation Is More Difficult Than It Appears

A radiology report is not simply a description of an image.

It is a clinical communication document.

The report must answer:

What is abnormal?
Where is it?
How confident are we?
What is the most likely diagnosis?
What alternatives remain?
Does the finding require action?
What should the referring clinician know immediately?

A generative AI model can produce grammatically excellent prose while still generating an incorrect clinical statement.

This creates a dangerous asymmetry:

Linguistic quality can conceal diagnostic error.

A fluent report can therefore be more dangerous than an obviously poor report because users may be less likely to question it.

Human verification remains essential.

14. Explainability: Does the Radiologist Need to Know Why AI Thinks Something Is Abnormal?

Ideally, yes—but explainability itself is not a guarantee of correctness.

Heatmaps, saliency maps, bounding boxes, confidence scores, segmentation overlays, and attention visualizations may help the radiologist understand what the model is focusing on.

However, an explanation can be visually persuasive without accurately representing the underlying reasoning of the model.

Therefore, explainability should be treated as a supporting interface rather than definitive evidence.

A clinically useful AI explanation should help answer:

Where did the AI look?

What did it detect?

How confident is it?

What are the known limitations?

Was the examination within the validated population and technical range?

The ultimate objective is not to make AI appear understandable.

It is to make AI appropriately verifiable.

15. Human Oversight Is a Clinical Safety Mechanism

Human oversight is sometimes presented as an ethical requirement.

It is more than that.

It is a clinical safety mechanism.

The radiologist is responsible for evaluating whether:

the examination is technically adequate
the AI output is applicable
the highlighted finding is genuine
the clinical context is consistent
additional imaging is required
the AI has encountered an unusual case
the final diagnosis is sufficiently supported

The WHO has emphasized that responsibility and accountability must remain clearly defined when AI is used in healthcare, together with attention to equity, transparency, safety and human rights.

Radiology therefore needs AI systems that are not merely accurate but governable.

16. What Happens When AI Is Wrong?

AI errors can be divided conceptually into two broad categories.

False Negative

The AI fails to identify a real abnormality.

False Positive

The AI identifies something that is not clinically meaningful.

Both have consequences.

A false negative may contribute to a missed diagnosis.

A false positive may increase:

additional imaging
unnecessary follow-up
patient anxiety
radiologist workload
healthcare expenditure

The appropriate threshold depends on the clinical task.

For a life-threatening emergency, high sensitivity may be prioritized.

For population screening, specificity and downstream consequences may become particularly important.

Therefore, AI performance should never be summarized by a single accuracy number.

Clinical implementation requires examination of:

sensitivity
specificity
positive predictive value
negative predictive value
calibration
subgroup performance
workflow impact
false-positive burden
false-negative consequences
time to diagnosis
patient outcomes

17. Why External Validation Matters

An AI model can perform exceptionally well in development data and still perform poorly elsewhere.

Reasons include differences in:

scanners
acquisition protocols
patient populations
disease prevalence
image quality
demographics
clinical referral patterns
annotation practices
institutional workflow

This is the problem of generalizability.

A model validated in one hospital should not automatically be assumed to perform identically in another.

Each stage answers a different question.

Retrospective testing asks whether the model can work on historical cases.

Prospective silent testing asks whether it behaves appropriately in the actual workflow without influencing decisions.

Live evaluation asks whether the technology improves clinical performance.

Post-deployment monitoring asks whether performance remains stable over time.

18. AI Drift: The Problem After Deployment

AI validation does not end when a model enters clinical practice.

Performance can change.

This phenomenon is often described as model drift or performance drift.

A hospital may replace a CT scanner.

A new imaging protocol may be introduced.

Patient demographics may change.

Disease prevalence may change.

The AI vendor may update the model.

Radiologists may change how they interact with the system.

Any of these factors can alter real-world performance.

This is why continuous monitoring is essential.

The American College of Radiology's AI quality framework emphasizes governance, inventory, testing, workflow integration, and ongoing monitoring of clinical AI.

The question after deployment is therefore not:

“Does this AI work?”

It is:

“Does this AI continue to work safely in this clinical environment?”

19. The Radiologist's New Skill Set

Human-AI collaboration will change radiology training.

Future radiologists will need traditional imaging expertise plus a working understanding of AI.

They should understand concepts such as:

sensitivity and specificity
ROC analysis
calibration
dataset shift
external validation
bias
algorithmic failure
explainability
workflow integration
cybersecurity
data governance
regulatory considerations

They do not need to become machine-learning engineers.

But they need enough AI literacy to ask intelligent clinical questions.

For example:

What population was this model trained on?

Was it externally validated?

What happens when image quality is poor?

Which abnormalities does it reliably detect?

What are its known failure modes?

How frequently is the model monitored?

These questions are becoming part of clinical competence.

20. AI Literacy Should Begin With Failure, Not Marketing

An effective AI education program should not begin with a demonstration of how impressive a model can be.

It should begin with how it fails.

Radiologists should see:

false positives
false negatives
atypical anatomy
poor-quality images
out-of-distribution cases
misleading heatmaps
incorrect report suggestions

Understanding failure creates appropriate expectations.

The objective is not to make clinicians enthusiastic about AI.

It is to make them competent users of AI.

21. The Future: From AI Tools to Clinical Teammates

The next generation of radiology AI may become more integrated.

Instead of isolated algorithms for individual findings, future systems may combine:

imaging
prior imaging
radiology reports
laboratory data
pathology
clinical notes
genomic information
patient history

This creates a multimodal clinical intelligence layer.

A future system might identify a lesion on CT, compare it with previous imaging, quantify interval growth, summarize relevant clinical information, and draft a structured interpretation.

But the final question remains human:

What does this finding mean for this patient?

That question cannot be reduced to image classification alone.

22. Human-AI Collaboration and Patient Trust

Patients may not care whether a diagnosis was produced by a neural network, a radiologist, or both.

They care whether the diagnosis is correct, understandable, timely and safe.

This makes transparency important.

Patients should not be left with the impression that an AI system independently made a definitive medical decision if a physician actually reviewed and interpreted the case.

The future relationship should therefore be:

AI assists.

Radiologist verifies.

Clinical team decides.

Patient participates.

That structure preserves the human dimension of healthcare while allowing technology to contribute computational capabilities.

23. What Should Hospitals Consider Before Deploying Radiology AI?

A hospital should evaluate more than model accuracy.

A practical implementation checklist includes:

Domain	Key Question
Clinical validity	Has the model been appropriately validated?
Generalizability	Does validation represent the local population and imaging environment?
Workflow	Where does AI enter the radiology workflow?
Human oversight	Who reviews and acts on AI output?
Safety	What happens when the AI fails?
Monitoring	How is performance monitored after deployment?
Governance	Who owns the AI lifecycle?
Cybersecurity	How are imaging and clinical data protected?
Usability	Does the interface support rather than distract from interpretation?
Accountability	Are responsibilities clearly defined?
Education	Are users trained in strengths and failure modes?
Equity	Does performance vary across patient groups?

A technically excellent model can still fail as a clinical product if these questions are ignored.

24. What Is the Most Important Principle for Radiologists?

The most important principle may be surprisingly simple:

Never allow the AI output to replace your observation of the image.

AI should direct attention, not dictate perception.

When the AI highlights a lesion, inspect the lesion.

When the AI reports “normal,” review the examination.

When the AI produces a confident classification, ask whether the case is appropriate for that model.

When the AI-generated report sounds convincing, verify every clinically meaningful statement.

This is not distrust.

It is professional responsibility.

25. A Radiologist's Practical Human-AI Checklist

Before accepting an AI output, ask:

1. Is the examination technically adequate?

If the image quality is poor, AI performance may not be reliable.

2. Is the AI tool appropriate for this examination?

A model designed for one acquisition protocol or pathology should not automatically be generalized to another.

3. What exactly did the AI detect?

Identify the anatomical location and imaging correlate.

4. Does the finding exist?

Review the original images independently.

5. Is the finding clinically meaningful?

Not every detected abnormality represents disease requiring intervention.

6. Does the clinical context support the interpretation?

Integrate symptoms, history, and previous imaging.

7. Could the AI be wrong?

Actively consider the principal failure modes.

8. Would additional imaging change the decision?

AI should not eliminate appropriate multimodality evaluation.

9. Can the final report be defended clinically?

The radiologist should be able to explain the final interpretation without relying solely on the algorithm.

26. What Will AI Not Replace?

AI may automate tasks.

It may change workflows.

It may reduce repetitive work.

It may improve consistency.

But several responsibilities remain fundamentally clinical:

integrating imaging with the patient
managing uncertainty
determining clinical relevance
communicating critical findings
deciding when additional investigation is justified
recognizing unexpected diagnoses
balancing competing possibilities
participating in multidisciplinary decision-making
accepting professional responsibility

The radiologist of the future may therefore spend less time on repetitive visual search and more time on higher-order interpretation and clinical communication.

That is not a reduction of radiology.

It is a potential evolution of radiology.

Conclusion

Human-AI collaboration in radiology should not be framed as a competition between biological and artificial intelligence.

The clinically meaningful future is more nuanced.

AI can process images at scale, identify patterns, perform quantitative measurements, prioritize examinations, and assist with reporting. Radiologists bring contextual reasoning, clinical judgment, uncertainty management, communication, and accountability.

The strongest systems will connect these capabilities rather than attempt to eliminate one of them.

The real innovation is therefore not simply a more accurate algorithm.

It is a better human-AI system.

Such a system must be clinically validated, integrated into workflow, continuously monitored, transparent about limitations, and designed around meaningful human oversight.

For radiologists, the essential skill will not be learning how to compete with AI.

It will be learning how to work intelligently with AI without surrendering clinical judgment to it.

The ultimate objective is not artificial intelligence replacing human intelligence.

It is augmented clinical intelligence serving the patient.

Key Takeaways

Human-AI collaboration is fundamentally different from autonomous AI diagnosis.
AI and radiologists may provide complementary strengths and complementary errors.
AI assistance does not guarantee improved diagnostic performance.
Automation bias and excessive trust in AI are important clinical risks.
Human oversight remains central to safe radiology AI deployment.
AI should be evaluated within real clinical workflows, not only as a standalone algorithm.
External validation and prospective evaluation are essential before broad implementation.
Post-deployment monitoring is necessary because AI performance can change over time.
Radiologists need practical AI literacy, including understanding model limitations and failure modes.
The future of radiology is likely to emphasize augmented clinical intelligence rather than simple human replacement.

FAQ

1. Will AI replace radiologists?

AI is more likely to transform radiologists' work than simply eliminate the profession. Many tasks can be automated or assisted, but clinical interpretation requires integration of imaging, history, prior examinations, uncertainty, and patient-specific context. The more realistic model is radiologist-plus-AI collaboration, with the physician retaining responsibility for the final clinical interpretation.

2. Can AI improve radiologist diagnostic accuracy?

Yes, AI assistance can improve performance for particular imaging tasks, but the effect is not uniform. Benefits depend on the algorithm, pathology, reader, workflow, and quality of AI predictions. Some studies demonstrate improved sensitivity or efficiency, while other research shows that incorrect AI outputs can negatively influence clinicians.

3. What is automation bias in radiology?

Automation bias occurs when a clinician gives excessive weight to an automated recommendation and reduces independent evaluation. In radiology, this may cause a radiologist to accept an AI-generated finding without adequately reviewing the original images. Appropriate AI design and education should encourage verification rather than passive acceptance.

4. Should radiologists trust AI?

Radiologists should neither blindly trust nor automatically reject AI. AI should be treated as an additional source of evidence. The radiologist should understand the model's intended use, validation population, limitations, and failure modes and independently verify clinically important outputs.

5. Why is external validation important for radiology AI?

Performance can change when an AI model is used with different scanners, protocols, populations, disease prevalence, or clinical workflows. External validation helps determine whether performance generalizes beyond the development environment. Prospective evaluation and post-deployment monitoring provide additional evidence about real-world reliability.

6. Can generative AI write radiology reports?

Generative AI can assist with draft reporting and information synthesis, but generated text should not automatically be considered a final medical report. Incorrect statements can be expressed with fluent and convincing language. Radiologist verification remains essential, particularly for abnormal and clinically complex examinations.

7. What is the future of human-AI collaboration in radiology?

The future is likely to involve increasingly multimodal systems that combine imaging, prior studies, clinical information, and structured data. The radiologist will remain central to clinical reasoning while AI increasingly supports detection, quantification, workflow management, and communication.

Medical Disclaimer

This column is intended for educational and professional information purposes. It does not provide individualized medical diagnosis, treatment, or clinical advice. AI performance varies according to the model, imaging modality, patient population, acquisition protocol, clinical environment, and workflow. Clinical decisions should remain under appropriate supervision by qualified healthcare professionals.

References

Geis JR, Brady AP, Wu CC, Spencer J, Ranschaert E, Jaremko JL, et al. Ethics of Artificial Intelligence in Radiology: Summary of the Joint European and North American Multisociety Statement. Radiology. 2019;293(2):436-440. DOI: 10.1148/radiol.2019191586.
McKinney SM, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature. 2020;577(7788):89-94. DOI: 10.1038/s41586-019-1799-6.
Ardila D, Kiraly AP, Bharadwaj S, et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med. 2019;25(6):954-961. DOI: 10.1038/s41591-019-0447-x.
Lee JH, Hong H, Nam G, Hwang EJ, Park CM. Effect of Human-AI Interaction on Detection of Malignant Lung Nodules on Chest Radiographs. Radiology. 2023;307. DOI: 10.1148/radiol.222976.
Kocak B, Cuocolo R. Human-AI interaction and collaboration in radiology: from conceptual frameworks to responsible implementation. Diagn Interv Radiol. 2026. DOI: 10.4274/dir.2026.263780.
Mezrich JL. Automation bias and overconfidence in artificial intelligence and associated legal implications. Clin Imaging. 2026;132:110745. DOI: 10.1016/j.clinimag.2026.110745.
Huang HY, Huang YH, Lin CH, et al. AI-assisted chest radiograph interpretation enhances diagnostic confidence and standardizes diagnostic accuracy across radiologists: a multi-reader study. Clin Imaging. 2026;130:110694. DOI: 10.1016/j.clinimag.2025.110694.
Pellegrini C, et al. Enhancing radiology workflows through collaborative AI-assisted chest X-ray reporting using large vision-language models: a proof-of-concept study. Insights Imaging. 2026. DOI: 10.1186/s13244-026-02292-7.
Impact of human and artificial intelligence collaboration on workload reduction in medical image interpretation. npj Digit Med. 2024. DOI: 10.1038/s41746-024-01328-w.
Artificial intelligence as a simultaneous second reader in diagnostic radiology: An umbrella review of systematic reviews and meta-analyses. Clin Imaging. 2026.
World Health Organization. Ethics and governance of artificial intelligence for health. Geneva: WHO; 2021.