Radiology AI Standard of Care: Liability, FDA Devices, and Best Practices

The Frontline of Medical AI

#

Radiology is where artificial intelligence meets clinical medicine at scale. With over 870 FDA-cleared AI algorithms, representing 78% of all medical AI approvals, radiology is both the proving ground and the liability frontier for AI in healthcare. When these algorithms miss cancers, misidentify strokes, or generate false positives that lead to unnecessary interventions, radiologists and healthcare systems face mounting legal exposure.

This guide examines the standard of care for AI use in radiology: what FDA clearance means and doesn’t mean, how courts are approaching liability, what professional societies recommend, and how radiologists can protect themselves while leveraging AI’s genuine benefits.

FDA-Cleared Radiology AI: What It Means

#

The Scale of Approval

#

The FDA has cleared an unprecedented number of radiology AI tools:

Leading Vendors by Clearances:

What FDA Clearance Does NOT Mean

#

Critical for liability analysis:FDA 510(k) clearance does NOT require:

No Prospective Human Trials:

• 97% of radiology AI cleared via 510(k) pathway
• Only need to show “substantial equivalence” to predicate device
• Retrospective data often sufficient

No Generalizability Proof:

• Training on one population doesn’t guarantee performance on another
• A 2024 JAMA study found significant generalizability concerns across patient populations

No Mandatory Post-Market Surveillance:

• Performance in real clinical settings often unknown
• AI drift goes unmonitored without voluntary tracking

No Demographic Validation:

• 95.5% of FDA-cleared AI devices don’t report demographic representation in submissions
• Performance gaps across race, age, and sex may exist undisclosed

2024 Reclassifications

#

The FDA reclassified certain radiology AI products from Class III to Class II:

• Mammography breast cancer detection
• Ultrasound breast lesion analysis
• Radiograph lung nodule detection
• Radiograph dental caries detection

This makes market entry easier but doesn’t indicate improved safety or efficacy.

Clinical Applications and Liability Exposure

#

Mammography AI

#

Current State:

• Largest category of radiology AI by MAUDE adverse event reports (69%)
• Assists in breast cancer detection and density assessment
• Studies show potential for improved sensitivity

Liability Concerns:

• False negatives leading to delayed cancer diagnosis
• False positives causing unnecessary biopsies
• Performance gaps across breast density levels

Notable Development (May 2025): Clairity’s Allix5 received de novo authorization, first AI for breast cancer risk prediction (not detection). Explicitly “not intended to diagnose, detect, or inform the treatment of cancer.”

Stroke Detection AI

#

Current State:

• Time-critical application (every minute matters)
• Tools like Viz.ai alert clinicians to large vessel occlusions
• Can accelerate transfer to stroke-capable facilities

Liability Concerns:

• Critical incident: FDA-cleared AI misidentified ischemic stroke as intracranial hemorrhage, opposite conditions requiring different treatment
• Overreliance may delay human interpretation
• Alert fatigue from false positives

Lung Nodule Detection

#

Current State:

• High volume of CT scans creates screening burden
• AI assists in identifying nodules for follow-up
• Potential to catch early-stage lung cancer

Liability Concerns:

• Missed nodules leading to late-stage diagnosis
• Overdetection leading to unnecessary intervention
• Performance varies by scanner type and patient population

Fracture Detection

#

Current State:

• Emergency department workflow integration
• Assists in identifying subtle fractures
• Can flag studies for radiologist prioritization

Liability Concerns:

• Missed fractures in complex anatomy
• False confidence reducing careful human review
• Performance gaps across age groups and body regions

The Liability Framework

#

Who Is Liable When AI Gets It Wrong?

#

Liability allocation depends on the AI’s role and the human’s response:

AI as Decision Support (Most Current Tools):

• Radiologist makes final determination
• Radiologist bears primary liability
• AI is advisory input, not substitute for judgment

AI as Autonomous System:

• If AI acts independently (rare in radiology currently)
• Product liability against manufacturer strengthens
• Vicarious liability may apply if AI is “subordinate” to radiologist

The “Black Box” Problem: Neural network algorithms cannot be fully understood by manufacturers or clinicians. When AI recommendations cannot be explained:

• Difficult for radiologists to assess soundness
• Challenging to attribute specific errors
• Harder for plaintiffs to prove causation

The Radiologist’s Double Bind

#

AI creates a unique liability trap:

If Radiologist Follows AI and It’s Wrong:

• Physician may be liable for failing to apply independent clinical judgment
• “I followed the AI” is not a defense

If Radiologist Overrides AI and Misses Something:

• Liability for ignoring available technology
• AI recommendation becomes evidence of what should have been seen

Documentation Is Critical:

• Record when AI was consulted
• Document reasoning for agreeing or disagreeing
• Note any AI limitations relevant to specific case

Emerging Product Liability

#

Plaintiffs increasingly add AI vendors as defendants:

Theories Against AI Developers:

• Design defect (AI trained on biased/limited data)
• Manufacturing defect (specific version bugs)
• Failure to warn (inadequate disclosure of limitations)

Challenges:

• Learned intermediary doctrine (warnings to physicians, not patients)
• FDA clearance as evidence of reasonable care
• Lack of direct physician-patient-style duty

Malpractice Trends and Notable Cases

#

Rising Claims

#

Data indicates significant increase in AI-related malpractice:

2024 Statistics:

• 14% increase in malpractice claims involving AI tools (vs 2022)
• Majority from diagnostic AI in radiology, cardiology, oncology
• Missed cancer diagnoses by AI a central focus

Notable Verdicts

#

Defense Strategies

#

Common defenses in radiology malpractice:

• Standard of care was met at time of interpretation
• AI tool was properly used per manufacturer instructions
• Other factors contributed to patient outcome
• Plaintiff contributed to delay (missed follow-ups)

Professional Society Guidelines

#

American College of Radiology (ACR)

#

The ACR has provided guidance on AI implementation:

Key Recommendations:

• Establish AI oversight committees to review new tools
• Track performance through post-market surveillance
• Ensure quality through ongoing monitoring
• Demand transparency from manufacturers on training data

April 2025 FDA Comment: ACR submitted formal comments to FDA on AI-enabled device software (Docket FDA-2024-D-4488), emphasizing:

• Need for clear validation requirements
• Importance of transparency in training data and ground truth labels
• Role of radiologists in oversight

Radiological Society of North America (RSNA)

#

RSNA emphasizes:

• Radiologists remain ultimately responsible for diagnosis
• AI should augment, not replace, clinical judgment
• Institutions should validate AI performance locally

Specialty-Specific Guidance

#

Professional societies in subspecialties are developing AI-specific guidelines:

• Society of Breast Imaging (mammography AI)
• American Society of Neuroradiology (stroke and brain AI)
• Society of Cardiovascular CT (cardiac imaging AI)

Standard of Care Framework

#

What Reasonable AI Use Looks Like

#

Based on FDA guidance, professional society recommendations, and emerging case law:

Pre-Implementation:

• Validate AI performance in your patient population
• Understand training data demographics and limitations
• Establish clear use case boundaries
• Train radiologists on AI capabilities and limitations

Clinical Use:

• AI recommendations are advisory, not determinative
• Radiologist applies independent clinical judgment
• Document AI use and reasoning for concordance/discordance
• Maintain human oversight of all final interpretations

Quality Assurance:

• Track concordance rates between AI and radiologists
• Monitor for demographic performance gaps
• Report adverse events to FDA MAUDE
• Regularly reassess AI performance

What Falls Below Standard

#

Practices likely to create liability exposure:

Implementation Failures:

• Deploying AI without local validation
• Using AI outside approved indications
• Failing to train staff on limitations
• No quality monitoring program

Clinical Failures:

• Treating AI output as definitive diagnosis
• Ignoring AI recommendations without documented reasoning
• Over-relying on AI in complex or atypical cases
• Failing to consider AI limitations for specific patient

Systemic Failures:

• No AI oversight committee
• Ignoring FDA safety communications
• Failing to update for known issues
• Suppressing concerns about AI performance

Risk Mitigation for Radiologists

#

Documentation Best Practices

#

Every AI-assisted interpretation should document:

1. AI tool used, Name, version, indication
2. AI output, What the AI found/recommended
3. Radiologist assessment, Agreement, disagreement, or modification
4. Clinical reasoning, Why radiologist reached final conclusion
5. Limitations noted, Any factors limiting AI reliability

Institutional Governance

#

Healthcare systems should establish:

AI Oversight Committee:

• Review and approve AI tools before deployment
• Monitor ongoing performance
• Investigate adverse events
• Update policies as technology evolves

Credentialing:

• Require AI training for radiologists
• Document competency in AI-assisted interpretation
• Include AI use in quality review

Contracts:

• Review liability allocation with AI vendors
• Ensure adequate indemnification provisions
• Require performance guarantees

Insurance Considerations

#

Policy Review:

• Check for AI-specific exclusions
• Understand coverage for algorithm-related claims
• Consider AI training requirements for coverage

Emerging Coverage:

• Some insurers now offer AI-specific riders
• Technology E&O may complement malpractice coverage
• Cyber insurance may cover some data-related AI failures

Frequently Asked Questions

#

Related Resources

#

AI Liability Framework

#

• AI Misdiagnosis Case Tracker, Diagnostic failure documentation
• AI Product Liability, Strict liability for AI systems
• AI Medical Device Adverse Events, FDA MAUDE analysis

Healthcare AI

#

• Healthcare AI Standard of Care, Overview of medical AI standards
• AI Insurance Coverage, E&O and malpractice considerations

Emerging Litigation

#

• AI Litigation Landscape 2025, Overview of AI lawsuits
• Section 230 and AI, Platform immunity questions