High-yield overview
Machine learning for fracture detection, measurement and planning - decision support, not decision replacement
—Fracture Detection
90-95%Sensitivity
—FDA Cleared Tools
—Multiple available
—Common Application
—Wrist, hip fractures
—Role
—Decision support
AI Application Categories
Detection: Fracture identification, abnormality flagging
Measurement: Automated angles, alignment metrics
Planning: Arthroplasty templating, surgical simulation
Prioritisation: Worklist triage by urgency
Key: AI augments clinical capability but requires human oversight
Critical Must-Knows
- AI tools are decision support - clinician remains responsible
- High sensitivity for fracture detection reduces missed injuries
- Best validated for wrist, hip, and chest radiograph applications
- Cannot replace clinical correlation and physical examination
- Regulatory clearance (FDA 510(k), CE/UKCA, or national equivalent) required for clinical use
Clinical Pearls
- "AI assists detection but does not replace clinical decision-making
- "Deep learning uses convolutional neural networks (CNNs)
- "Performance depends on training data quality and diversity
- "Particularly useful for reducing missed fractures in ED
Clinical Warning
AI in radiology is an emerging topic. For fellowship exams, understand the basic concepts (machine learning, deep learning), current validated applications (fracture detection), limitations (training bias, cannot replace clinical judgement), and the medicolegal position (clinician responsibility remains).
Mnemonic
AI TAppraising an Imaging AI Tool
| V | Validation - prospective, peer-reviewed, on a population like yours Validation - prospective, peer-reviewed, on a population like yours |
| A | Approval - FDA/CE/UKCA or national regulator clearance for your jurisdiction Approval - FDA/CE/UKCA or national regulator clearance for your jurisdiction |
| L | Limitations - know the failure modes (occult, out-of-distribution, paediatric) Limitations - know the failure modes (occult, out-of-distribution, paediatric) |
| I | Integration - fits PACS/workflow; who reviews the output? Integration - fits PACS/workflow; who reviews the output? |
| D | Drift & audit - ongoing local monitoring of sensitivity/specificity Drift & audit - ongoing local monitoring of sensitivity/specificity |
| V | Validation - prospective, peer-reviewed, on a population like yours Validation - prospective, peer-reviewed, on a population like yours | I | Integration - fits PACS/workflow; who reviews the output? Integration - fits PACS/workflow; who reviews the output? |
| A | Approval - FDA/CE/UKCA or national regulator clearance for your jurisdiction Approval - FDA/CE/UKCA or national regulator clearance for your jurisdiction | D | Drift & audit - ongoing local monitoring of sensitivity/specificity Drift & audit - ongoing local monitoring of sensitivity/specificity |
| L | Limitations - know the failure modes (occult, out-of-distribution, paediatric) Limitations - know the failure modes (occult, out-of-distribution, paediatric) |
Hook:If a tool is not VALID for your jurisdiction and population, a high published AUC is irrelevant - never deploy on the vendor's numbers alone.
Mnemonic
AI RWhy a 'Negative' AI Result Never Excludes a Fracture
| S | Sensitivity is not 100% - occult fractures are still missed Sensitivity is not 100% - occult fractures are still missed |
| A | Automation bias - do not let a confident output stop your reasoning Automation bias - do not let a confident output stop your reasoning |
| F | Findings clinically - examination and mechanism override the algorithm Findings clinically - examination and mechanism override the algorithm |
| E | Escalate - high suspicion warrants immobilise, repeat imaging or MRI/CT Escalate - high suspicion warrants immobilise, repeat imaging or MRI/CT |
| S | Sensitivity is not 100% - occult fractures are still missed Sensitivity is not 100% - occult fractures are still missed | F | Findings clinically - examination and mechanism override the algorithm Findings clinically - examination and mechanism override the algorithm |
| A | Automation bias - do not let a confident output stop your reasoning Automation bias - do not let a confident output stop your reasoning | E | Escalate - high suspicion warrants immobilise, repeat imaging or MRI/CT Escalate - high suspicion warrants immobilise, repeat imaging or MRI/CT |
Hook:The classic trap: AI says 'no fracture', the patient has snuffbox tenderness - you still treat as a scaphoid fracture. Clinical suspicion always wins.
Overview & Core Principles
AI Terminology
| Term | Definition | Example |
|---|---|---|
| Artificial Intelligence (AI) | Machines performing tasks requiring human intelligence | Any automated image analysis |
| Machine Learning (ML) | Algorithms that improve through experience | Learning from labelled examples |
| Deep Learning (DL) | Neural networks with multiple layers | Convolutional neural networks |
| Convolutional Neural Network (CNN) | Neural network for image analysis | Fracture detection models |
| Training Data | Labelled examples used to teach the algorithm | Radiographs with/without fractures |
| Inference | Applying trained model to new data | Analysing a new patient radiograph |
How AI Learns to Detect Fractures
A deep learning model is trained on thousands of labelled radiographs (fracture vs no fracture). The CNN automatically learns features that distinguish fractures (cortical disruption, angulation, subtle lucent lines) without explicit programming. The model is then validated on a separate test set to assess real-world performance. More diverse training data generally improves generalisation.
Clinical Imaging Applications
AI Fracture Detection Performance
| Body Region | Typical Sensitivity | Clinical Utility |
|---|---|---|
| Wrist/hand | 90-95% | Reduces missed scaphoid, metacarpal fractures |
| Hip | 90-98% | Flags occult neck of femur fractures |
| Chest (ribs) | 85-95% | Detects subtle rib fractures |
| Spine | 85-92% | Identifies vertebral compression fractures |
| Ankle | 88-94% | Assists with subtle malleolar fractures |
| Paediatric elbow | 85-92% | Helps with occult fractures |
ED Workflow Integration
AI fracture detection tools integrate into the ED workflow by automatically analysing radiographs and flagging potential fractures. This can reduce missed fractures (particularly important for trainee coverage and high-volume departments), prioritise urgent cases, and provide a 'second read'. The clinician reviews all AI suggestions and makes the final determination.
Performance Metrics
Understanding AI Performance
| Metric | Definition | Clinical Interpretation |
|---|---|---|
| Sensitivity | True positive rate (detects fractures) | High = few missed fractures |
| Specificity | True negative rate (correct negatives) | High = few false alarms |
| PPV | Positive predictive value | Probability positive result is true |
| NPV | Negative predictive value | Probability negative result is true |
| AUC-ROC | Area under ROC curve | Overall discriminative ability (0.5-1.0) |
| F1 Score | Harmonic mean of precision/recall | Balanced performance measure |
Sensitivity vs Specificity Trade-off
In fracture detection, high sensitivity (few missed fractures) is prioritised over specificity. A sensitive AI tool may generate false positives (overcalling fractures) which are easily dismissed by the clinician. Missing a fracture (false negative) has more serious consequences. Most AI tools are tuned for high sensitivity, accepting some overcalling.
Limitations
Algorithm
AI Limitations in Radiology
| Limitation | Explanation | Mitigation |
|---|---|---|
| Training bias | Model reflects training data characteristics | Diverse, representative datasets |
| Out-of-distribution | Poor performance on unusual cases | Clinical oversight, flag uncertainty |
| Black box | Cannot explain reasoning | Explainability research, heatmaps |
| Data quality | Garbage in, garbage out | Quality training data curation |
| Regulatory lag | Approval slower than development | Use only approved tools clinically |
| Integration challenges | Technical/workflow barriers | PACS integration, user training |
Regulatory and Medicolegal
Regulatory Framework
| Aspect | Requirement | Notes |
|---|---|---|
| Classification | Medical device (software) | SaMD - Software as Medical Device (IMDRF framework) |
| FDA clearance (US) | Required for clinical use in USA | 510(k) pathway most common for fracture AI |
| CE/UKCA marking (EU/UK) | Required in EU (MDR) and UK (UKCA) | Most fracture tools are MDR Class IIa/IIb |
| National regulators | Required in each jurisdiction | TGA (Australia), Health Canada, PMDA (Japan), CDSCO (India) |
| Clinical validation | Performance data required | Prospective, locally representative studies preferred |
| Post-market surveillance | Ongoing monitoring + drift detection | Report adverse events; monitor performance over time |
Medicolegal Position
The clinician remains legally responsible for the clinical decision, whether or not AI was used. AI is a decision support tool, not a decision-maker. If AI misses a fracture, the clinician is still responsible for the missed diagnosis if they did not exercise appropriate clinical judgement. Documentation should reflect that AI was used as an adjunct, not as the sole basis for the decision.
Guidelines, Registries & Global Practice
Missed fractures are the single most common diagnostic error in musculoskeletal imaging worldwide, and the burden falls hardest where specialist reporting is scarce - this is the global rationale for fracture-detection AI.
Society & Regulatory Positions on AI in Imaging (Side by Side)
| Body / Region | Position on imaging AI | Practical implication |
|---|---|---|
| FDA (US) | Clears most fracture tools via 510(k); evolving framework for adaptive/locked algorithms | Cleared tools are decision support; predicate-based clearance does not prove outcome benefit |
| ACR (US) | Endorses AI as an adjunct; runs the ACR AI registry (Assess-AI) and Data Science Institute use cases | Encourages local performance monitoring rather than blind adoption |
| RCR / NICE / NHS (UK) | RCR cautious endorsement; NICE early value assessment of fracture-detection AI (e.g. ED use) | Permits conditional use with evidence generation; human report still required |
| EFORT / European radiology bodies | Support AI as augmentation; emphasise CE/MDR compliance and explainability | MDR Class IIa/IIb obligations and post-market surveillance |
| AO Foundation | Promotes AI for classification, templating and surgical planning education | Focus on consistency of fracture classification and pre-operative planning |
| WHO / IMDRF | Frameworks for SaMD and ethics of AI in health, relevant to limited-resource scale-up | Stresses equity, validation in local populations, and governance |
High-Resource vs Limited-Resource Practice Variation
| Dimension | High-resource setting | Limited-resource setting |
|---|---|---|
| Primary role of AI | Second-read / worklist triage to support specialist radiologists | Front-line decision support where no radiologist is available |
| Connectivity | Integrated into PACS/RIS, on-premise or cloud inference | May rely on smartphone capture or intermittent connectivity |
| Main benefit | Efficiency, reduced miss rate, faster turnaround | Access to expertise that would otherwise be absent (task-shifting) |
| Main risk | Automation bias, alert fatigue, over-investigation | Deployment of unvalidated tools, distribution shift, no oversight |
| Governance maturity | Formal validation, audit and surveillance pathways | Often absent - the key barrier to safe scale-up |
Registry & Audit Note
Unlike arthroplasty, AI imaging tools have no single international implant-style registry, but post-market monitoring registries are emerging (for example the ACR Assess-AI registry in the US). The exam-relevant principle: any deployed tool needs continuous local audit of sensitivity, specificity and performance drift in the actual patient population, not a one-off validation figure quoted by the vendor.
Future Directions
Emerging AI Applications
| Area | Application | Potential Impact |
|---|---|---|
| Natural language processing | Automated report generation | Efficiency, consistency |
| Multimodal AI | Combined imaging and clinical data | More holistic assessment |
| Federated learning | Training without sharing data | Privacy-preserving improvement |
| Foundation models | Pre-trained, adaptable models | Faster development of new tools |
| Real-time guidance | Intraoperative AI assistance | Surgical precision |
| Outcome prediction | Predict treatment success | Personalised medicine |
Radiologist-AI Collaboration
The future is likely radiologist-AI collaboration rather than replacement. AI handles routine detection and measurement tasks, freeing radiologists for complex interpretation, clinical correlation, and communication. Studies suggest radiologist + AI outperforms either alone for many tasks.
Systematic Approach: A Negative AI Result Is Not a Differential
The most dangerous error is treating an AI "no fracture" output as a clinical answer. AI flags a pattern; the clinician must still work through the differential of why a region hurts despite a reassuring algorithm. The table below contrasts the entities a fracture-detection model is and is not built to resolve.
Painful Region with a 'Negative' or Equivocal AI Output - Differential
| Entity | Why AI may miss or mislabel it | Clinical action that overrides AI |
|---|---|---|
| Occult scaphoid fracture | Often invisible on initial radiograph (AI trained on radiographs cannot see what is not yet visible) | Snuffbox tenderness - immobilise, repeat imaging or MRI at 10-14 days |
| Occult hip / femoral neck fracture | Subtle trabecular disruption, frequent false negatives in osteopenic bone | Inability to weight-bear - CT or MRI regardless of AI output |
| Stress / insufficiency fracture | Radiographically silent for 2-3 weeks; outside most training distributions | History (load change, metabolic risk) - MRI or bone scan |
| Pathological fracture / bone lesion | Model trained on traumatic fractures may not flag underlying lesion | Atraumatic mechanism, lytic/blastic clues - cross-sectional imaging, oncology referral |
| Non-accidental injury (paediatric) | AI detects the fracture, not the pattern or social context | Recognise inconsistent history, multiple ages of injury - safeguarding pathway |
| Soft-tissue / ligamentous injury | Fracture model has no class for ligament or tendon | Examination, stress views, MRI / ultrasound |
| Out-of-distribution image | Unusual projection, hardware, paediatric physis, rare anatomy degrades performance | Treat AI output as unreliable; rely on clinical reasoning |
The Automation-Bias Trap
The validated harm of AI is not the false negative itself but automation bias - the tendency to defer to a confident-looking output and stop thinking. Studies show readers can be anchored by an incorrect AI suggestion, occasionally being talked OUT of a correct call. The safe posture: read the image first, then consult the AI, and let a discordant high clinical suspicion always win.
Controversies & Areas of Uncertainty
Evidence Base
Deep Learning Assistance Closes the Accuracy Gap in Fracture Detection Across Clinician Types
Anderson PG, Baum GL, Keathley N, Sicular S, Venkatesh S, Sharma A, Daluiski A, Potter H, Hotchkiss R, Lindsey RV, Jones RM • Clin Orthop Relat Res (2022)
Key Findings:
- Multi-reader multi-case study: 24 clinicians (radiologists, orthopaedic surgeons, PAs, primary care and emergency physicians) read 175 cases across 12 anatomical regions, aided and unaided by an FDA-cleared deep learning tool
- Reader accuracy rose with AI aid: AUC 0.90 unaided to 0.94 aided (difference 0.04, 95% CI 0.01 to 0.07)
- Sensitivity improved from 82% to 90% and specificity from 89% to 92% with AI assistance
- Clinicians with limited MSK imaging training reduced their fracture miss rate from 20% to 9%, matching radiologist performance (10%)
Clinical Implication: AI assistance most benefits the non-specialist reader (ED physician, PA, junior trainee) who interprets a large share of musculoskeletal radiographs - exactly the after-hours setting where fractures are most often missed. It narrows, but does not eliminate, the expertise gap.
Limitation: Level III retrospective multi-reader design; reader behaviour in a study may not reflect real-world workflow, and the tool was developed by the sponsoring company.
Deep Learning Tool to Improve Fracture Detection by Radiologists and Emergency Physicians on Extremity Radiographs
Fu T, Viswanathan V, Attia A, Zerbib-Attal E, Kosaraju V, Barger R, Vidal J, Bittencourt LK, Faraji N • Acad Radiol (2023)
Key Findings:
- Standalone deep learning performance on 2626 extremity radiographs: accuracy 0.986, sensitivity 0.987, specificity 0.885, with accuracy over 0.95 across body part, age, sex, view and scanner
- Multi-reader study (24 readers): with AI aid, accuracy rose by 0.047 (95% CI 0.034 to 0.061) and sensitivity improved from 0.865 to 0.955
- Average interpretation time fell by 7.1 seconds (27%) per examination
- Diagnostic gain was largest for emergency physicians and non-MSK radiologists
Clinical Implication: Beyond accuracy, AI can shorten reading time and raise throughput - relevant to high-volume emergency departments and resource-limited settings with few specialist readers.
Limitation: Single-institution test set; retrospective reader study with a one-month washout, not a prospective outcome trial. Industry co-authorship.
Clinical Decision Scenarios
Use these scenarios to practise clinical reasoning and management decisions
CLINICAL SCENARIOChallenging
CLINICAL PROMPT
"Your hospital is considering implementing an AI tool for fracture detection on emergency department radiographs. What factors would you consider?"
PRACTICAL APPROACH
Key considerations: (1) Evidence base - Is there published validation data? What is the sensitivity and specificity? Was it validated on a population similar to ours? (2) Regulatory - Does it carry the appropriate clearance for our jurisdiction (FDA 510(k) in the US, CE/UKCA mark in Europe/UK, or the relevant national regulator such as TGA, Health Canada, PMDA or CDSCO)? This is mandatory for clinical use. (3) Integration - Can it integrate with our PACS and workflow? Who reviews the AI output? (4) Clinical governance - Who is responsible for the final decision? How is AI use documented? What happens if AI misses a fracture? (5) Training - Do clinicians understand how to interpret AI results, including limitations? (6) Cost-benefit - What is the cost versus expected reduction in missed fractures and potential litigation? (7) Quality improvement - How will we audit AI performance in our population? (8) Bias - Does it perform equally across all patient demographics? The clinician always remains responsible for the final clinical decision.
KEY CLINICAL POINTS
Requires regulatory clearance for the jurisdiction
Published validation data essential
PACS integration and workflow design
Clinician remains legally responsible
Local validation and ongoing audit
COMMON PITFALLS
Assuming AI eliminates missed fractures
Using unapproved tools clinically
Over-reliance without clinical correlation
FURTHER QUESTIONS
"What would you do if the AI says 'no fracture' but you have high clinical suspicion? How would you explain AI use to a patient?"
CLINICAL SCENARIOCritical
CLINICAL PROMPT
"An ED registrar reviews a wrist X-ray and the AI tool reports 'no fracture detected'. The patient has snuffbox tenderness."
PRACTICAL APPROACH
The registrar should proceed based on clinical judgement, NOT solely on AI output. Management: (1) Snuffbox tenderness is a clinical indicator for suspected scaphoid fracture. (2) AI 'no fracture' does not exclude a fracture - scaphoid fractures are notoriously difficult to detect on initial radiographs (sensitivity approximately 70-80% even for experienced readers). (3) Apply standard scaphoid protocol: immobilise in scaphoid cast/splint, arrange follow-up X-ray in 10-14 days or MRI if available. (4) Document clinical findings and management reasoning. Key principles: AI is decision support, not decision replacement. Clinical correlation is essential. A negative AI result with positive clinical findings requires conservative management. The clinician is responsible for the final decision. Document that AI was reviewed but clinical judgement guided management.
KEY CLINICAL POINTS
Clinical findings override AI output
Scaphoid fractures often occult initially
Treat clinically suspected fracture appropriately
AI sensitivity is not 100%
Document clinical reasoning
COMMON PITFALLS
Relying solely on AI result
Discharging without appropriate follow-up
Not documenting clinical reasoning
FURTHER QUESTIONS
"What is the sensitivity of initial radiographs for scaphoid fracture? How would you explain to the patient why you're treating despite a 'normal' X-ray?"
CLINICAL SCENARIOStandard
CLINICAL PROMPT
"You are asked to give a presentation on AI in orthopaedic imaging to your department. What key messages would you convey?"
PRACTICAL APPROACH
Key messages: (1) Current state - AI is increasingly validated for fracture detection (particularly wrist, hip), automated measurements (Cobb angle, alignment), and surgical templating. Multiple tools have regulatory approval. (2) Performance - AI achieves 90-95% sensitivity for fracture detection in validated applications, potentially reducing missed diagnoses. (3) Role - AI is a decision support tool, not a replacement for clinical judgement. The combination of AI + clinician often outperforms either alone. (4) Limitations - AI cannot examine patients, consider mechanism, or integrate the full clinical picture. Performance depends on training data and may not generalise to all populations. (5) Responsibility - The clinician remains legally and ethically responsible for decisions, regardless of AI use. (6) Future - Expect increased integration into workflows, more sophisticated applications (outcome prediction, surgical guidance), and AI-radiologist collaboration models. (7) Implementation - Requires regulatory approval, clinical governance, training, and ongoing audit.
KEY CLINICAL POINTS
AI is decision support, not replacement
High sensitivity but not perfect
Clinical correlation always required
Clinician remains responsible
AI + clinician often outperforms either alone
COMMON PITFALLS
Over-promising AI capabilities
Underestimating implementation challenges
Ignoring regulatory requirements
FURTHER QUESTIONS
"How might AI change radiology training? What ethical considerations exist with AI in healthcare?"
AI in Orthopaedic Radiology Quick Reference
Clinical summary
Core Concepts
- •Deep learning uses CNNs for image analysis
- •Trained on labelled examples
- •Validated on separate test data
- •Regulatory clearance required for clinical use
Current Applications
- •Fracture detection (wrist, hip common)
- •Automated measurements (Cobb angle)
- •Arthroplasty templating
- •Worklist prioritisation
Performance
- •Sensitivity 90-95% for fracture detection
- •High sensitivity prioritised (few missed)
- •May have lower specificity (overcalling)
- •AI + clinician better than either alone
Key Principles
- •Decision support, not replacement
- •Clinical correlation essential
- •Clinician remains legally responsible
- •Negative AI does not exclude pathology