AI Fracture Detection: How Deep Learning Is Reducing Missed Fractures in Emergency Radiology
Emergency departments miss 3–9% of fractures on initial radiograph interpretation. AI-powered fracture detection systems are demonstrating sensitivity improvements from 81% to 92% — here's what clinicians need to know about the technology, commercial tools, and clinical evidence.
The Missed Fracture Problem
Fractures are among the most common diagnoses in emergency departments worldwide — and among the most commonly missed. Published data consistently shows that 3–9% of fractures are not identified on initial radiograph interpretation, with the rate varying by anatomical region, time of day, and clinician experience level.
The clinical consequences of missed fractures extend beyond delayed pain management. A missed scaphoid fracture can progress to avascular necrosis. A missed vertebral compression fracture delays osteoporosis treatment. A missed Salter-Harris fracture in a child can result in growth plate arrest and progressive deformity.
These are not rare events. In a typical emergency department seeing 200 musculoskeletal radiographs per day, 3–9% missed fracture rate translates to 6–18 missed fractures daily. Multiply across a year, and the scale of the problem becomes clear.
AI-powered fracture detection directly addresses this gap — not by replacing clinician judgment, but by providing a systematic second reader that never fatigues, never rushes, and applies consistent criteria to every image.
How AI Fracture Detection Works
The Technical Pipeline
AI fracture detection systems use convolutional neural networks (CNNs) trained on large datasets of annotated radiographs. The fundamental approach involves two steps: localisation (where is the abnormality?) and classification (is it a fracture?).
Two architectural families dominate the field. Classification networks (ResNet, DenseNet, EfficientNet) analyse the entire radiograph and output a binary prediction — fracture present or absent — along with a confidence score. Object detection networks (YOLO, Faster R-CNN) go further by localising the fracture within the image, drawing a bounding box around the suspected abnormality. This localisation is clinically valuable because it directs the clinician's attention to the specific region requiring closer examination.
More recent approaches combine detection with segmentation — not just identifying that a fracture exists and where it is, but delineating its exact extent. This information supports fracture classification (e.g., Salter-Harris typing, AO classification) and surgical planning by providing precise fracture geometry.
What the Evidence Shows
A 2026 systematic review and meta-analysis of AI-assisted fracture detection across multiple anatomical regions found that AI assistance improved clinician sensitivity from approximately 81% to 92% — an absolute improvement of 11 percentage points. Critically, this improvement came without a significant decrease in specificity, meaning the AI was helping clinicians find real fractures rather than generating false alarms.
The improvement was most pronounced for less experienced clinicians (emergency physicians, junior residents) and for fracture types known to be frequently missed — non-displaced fractures, stress fractures, and fractures in anatomically complex regions.
Where AI Performs Best — and Where It Struggles
Strong Performance Areas
Proximal femur fractures are the most clinically impactful application. These fractures occur predominantly in elderly patients, carry significant morbidity and mortality, and can be subtle on standard AP pelvis radiographs — particularly non-displaced femoral neck fractures. AI systems have demonstrated sensitivity approaching 94% with specificity of 96% for proximal femur fracture detection.
Distal radius fractures — the most common adult fracture — show classification accuracies up to 97%. The clinical value is particularly high in paediatric populations, where distinguishing subtle buckle fractures from normal growth plate appearances requires experience that may not be available during overnight emergency shifts.
Spine fractures on lateral radiographs benefit significantly from AI assistance, particularly for compression fractures in osteoporotic patients where vertebral height loss may be gradual and difficult to appreciate without comparison imaging.
Challenging Areas
Stress fractures remain difficult for both human readers and AI. Early stress fractures may show no radiographic abnormality — they are MRI diagnoses. AI can detect cortical thickening and periosteal reaction that suggest evolving stress fractures, but sensitivity for early-stage stress injuries remains lower than for acute fractures.
Fractures with overlapping anatomy — scaphoid fractures obscured by carpal overlap, rib fractures in the context of pleural effusion, pelvic ring fractures — challenge AI systems that rely primarily on bone contour analysis.
Paediatric fractures present unique challenges because normal growth plate anatomy can mimic fracture lines. AI models must be specifically trained on paediatric imaging data to distinguish Salter-Harris injuries from normal physeal appearances. Recent work has shown promising results, but paediatric-specific validation remains less extensive than adult applications.
The Emergency Department Workflow
How AI Fits into Clinical Practice
The integration of AI fracture detection into emergency department workflow follows one of two models.
Concurrent reading: The AI analysis runs automatically when the radiograph is acquired, and results are available to the clinician at the time of initial interpretation. The AI overlay (heatmap or bounding box) appears alongside the original image in the PACS viewer. This model provides immediate decision support but requires that clinicians learn to calibrate their trust in the AI output.
Secondary safety net: The AI analysis runs in parallel or with a slight delay, and discrepancies between the clinician's initial assessment and the AI output trigger an alert. If the clinician reads the radiograph as normal but the AI flags a potential fracture, the case is routed for urgent second review. This model minimises workflow disruption and targets the specific problem of missed fractures.
Both models have demonstrated clinical value. The safety net approach is particularly well-suited to emergency departments because it addresses the highest-risk scenario (missed fracture) without requiring the clinician to change their primary workflow.
The Night Shift Problem
Fracture missed rates increase during overnight shifts, when junior residents may be interpreting radiographs without immediate senior supervision. A 2025 study in an emergency department setting demonstrated that AI-aided diagnosis increased fracture detection accuracy from 79.5% (clinician alone) to 89.3% (clinician with AI) — a statistically and clinically significant improvement.
AI fracture detection provides consistent performance regardless of time of day, fatigue level, or case volume — effectively functioning as an always-available second reader.
Commercial Tools: What's Available
Several AI fracture detection tools have achieved regulatory clearance and are deployed in clinical practice.
BoneView (Gleamer) — CE-marked and FDA-cleared. Detects fractures across multiple anatomical regions (wrist, hip, ankle, shoulder, spine, knee). Provides fracture localisation with bounding boxes and generates structured reports. Extensively validated in multi-reader studies.
SmartUrgences (Milvue) — CE-marked. Designed specifically for emergency radiology workflow. Demonstrated 91% sensitivity and 95% specificity in a prospective emergency department evaluation, with CT as the gold standard reference.
Other cleared tools include products from Imagen (OsteoDetect, specifically for distal radius fractures), Zebra Medical Vision, and several regional players.
The market is maturing rapidly, with increasing evidence from prospective clinical deployments rather than retrospective research studies alone.
What Clinicians Should Know
Before Adopting an AI Fracture Detection Tool
Regulatory status is non-negotiable. Only CE-marked or FDA-cleared tools should be used in clinical practice. Research Use Only tools may be valuable for evaluation but must not be the basis for clinical decisions.
Understand the performance envelope. No AI system detects 100% of fractures. Know the tool's sensitivity and specificity for each anatomical region and fracture type. A tool with 94% sensitivity for hip fractures but 78% sensitivity for scaphoid fractures has a very different clinical profile depending on your patient population.
AI does not replace clinical judgment. A negative AI result does not exclude fracture. Patients with high clinical suspicion should receive appropriate follow-up imaging (CT, MRI) regardless of the AI output.
Explainability supports trust. Tools that show their reasoning — heatmaps, bounding boxes, or attention maps highlighting the suspected fracture location — enable clinicians to assess whether the AI's conclusion is anatomically plausible.
Privacy architecture matters. Know whether the radiograph is processed locally or transmitted to an external server. For emergency department integration, processing speed is also critical — a tool that takes 5 minutes to return results has less clinical utility than one that runs in seconds.
The Training Effect
An underappreciated benefit of AI fracture detection is its educational value. A 2025 study demonstrated that an AI-based training module significantly improved trainee paediatric fracture detection accuracy. By consistently highlighting subtle fracture patterns, AI tools serve as a continuous teaching resource — particularly valuable for junior clinicians developing their radiograph interpretation skills.
Beyond Detection: Fracture Classification and Planning
Fracture detection — identifying that a fracture exists — is the first step. The clinical pathway continues with classification, reduction planning, and treatment decision-making.
Fracture classification (AO/OTA classification, Neer classification for proximal humerus, Garden classification for femoral neck) remains largely a human task in 2026, though AI systems are being developed to automate classification from radiographs and CT.
3D fracture analysis from CT data enables precise fragment identification, virtual reduction simulation, and patient-specific fixation planning. This represents the natural extension of detection AI into the surgical planning workflow.
Salnus's Position
While Salnus's current clinical focus is knee OA assessment rather than fracture detection, our underlying technology stack — DICOM processing, deep learning inference, explainable AI, and client-side privacy architecture — is applicable to fracture detection workflows.
Our nnU-Net bone segmentation model, currently achieving Dice scores of 0.964 across four knee bones, demonstrates the platform's capability for precise musculoskeletal image analysis. Extension to fracture-specific applications is a natural development pathway as the platform matures.
For clinicians interested in evaluating our platform or discussing potential fracture detection research collaborations, contact our team.
Disclaimer: This article is for educational and research purposes only. Salnus tools are designated for Research Use Only (RUO) and are not cleared medical devices. Clinical decisions should be made by qualified physicians based on comprehensive patient assessment.
References:
- Abdellatif N, et al. Assessment of AI-aided X-ray in diagnosis of bone fractures in emergency setting. Egypt J Radiol Nucl Med. 2025;56:160.
- Enhanced fracture detection on radiographs with AI assistance for clinicians: systematic review and meta-analysis. Ann Med. 2026;2610079.
- Elkohail A, et al. AI in Bone Fracture Detection: A Review. Cureus. 2025;17(11):e97674.
- AI for Fracture Diagnosis: Overview of Current Products. AJR. 2023.
- Xue F, et al. FAD-MIL: weakly supervised fracture detection. Sci Rep. 2026.
Reviewed by the Salnus biomedical engineering team.