AI in Orthopaedic Radiology

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.

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.

Related pages: Diagnostic Test Statistics is the page to read first and the one that makes every number here interpretable - sensitivity against specificity, why accuracy is a misleading summary when most films are normal, and why a test's predictive value collapses when prevalence falls from the 60-70% of a curated dataset to the few per cent of an emergency department. Levels of Evidence and Evidence-Based Orthopaedics and Critical Appraisal supply the appraisal framework that the retrospective, vendor-authored design of most of this literature fails, and Research Ethics and Integrity covers the conflict-of-interest declarations that decide how much weight an industry-authored evaluation earns. Scaphoid Fractures is the clinical problem these tools are most often aimed at and the reason a negative output never overrides snuffbox tenderness; Scaphoid Nonunion is what a missed one becomes. CT Imaging Principles and MRI Imaging Principles are the reference standards against which occult fracture is defined - and the heterogeneity in which of them each study used is a live source of bias in the pooled estimates. Systematic Approach to Trauma Imaging is the discipline that automation bias erodes: a reader who has stopped searching because the model reported nothing has abandoned the very habit that catches the second injury. Medicolegal Principles and Negligence and Informed Consent and Shared Decision-Making carry the responsibility question - the clinician remains liable for the report, and the patient is rarely told an algorithm was involved. Patient Safety and Wrong-Site Surgery Prevention is the wider literature on how checklists and alerts change behaviour, including how quickly people learn to ignore an alarm that is usually wrong. 3D Printing from Imaging, Robotic and Navigated Arthroplasty and Navigation and Robotics in Spine Surgery are the neighbouring technologies, and the same two tests apply to all of them: is the unaided task genuinely unreliable, and is the thing being made more accurate the outcome or only a surrogate?

• The problem. A deep CNN outputs a prediction without a human-readable reason - the 'black box'.
• The commonest solution. A saliency or class-activation map (for example Grad-CAM, gradient-weighted class-activation mapping) overlays a heatmap on the radiograph showing the region that most drove the model's decision.
• Why it matters. It lets the clinician sanity-check the model - confirming it looked at the fracture and not at a spurious cue such as a laterality marker, text or a cast ('shortcut learning', the Clever-Hans effect) - which builds trust and helps catch out-of-distribution failures. A heatmap shows where, not the true reasoning, so it is a check, not a proof of correctness.

• Three data sets. A model is fit on a training set; hyperparameters and model choice are tuned on a separate validation set; final, unbiased performance is measured on a held-out test set (and ideally an external dataset from a different centre). A published fracture tool, for example, was trained on a 87 percent / 11 percent / 2 percent train-validation-test split.
• Overfitting. If a model memorises the training data (including its noise) rather than learning generalisable features, it performs excellently on training data but poorly on new data - recognised when training accuracy is high but validation/test accuracy is much lower.
• Mitigation. More and more-diverse data, data augmentation (rotations, flips, brightness changes that synthetically expand the dataset), regularisation and early stopping. Any tuning done on the test set leaks information and inflates the reported performance - which is why an untouched, ideally external, test set matters.

Reading these six studies together

The two sensitivities on this page differ by nineteen points, and the reason is the study design, not the bone. The industry-authored standalone evaluation reports a sensitivity of 0.987; the independent meta-analysis pooling seven scaphoid datasets reports 0.80. Whenever you meet an AI performance figure, the first question is who generated it and on whose images: a manufacturer's held-out test set is the most favourable condition a model will ever see, while a pooled estimate across tools, centres and reference standards is the closest available approximation to what happens when someone else installs it. QUADAS-2 rated nine of ten primary studies at high risk of bias or applicability concern, which tells you the direction the published literature errs in.

Convert specificity into workload before you believe a tool is deployable. A standalone specificity of 0.885 sounds excellent and means roughly one normal radiograph in nine is flagged. A department reading a thousand normal extremity films a week would receive about 115 false alarms a week - the mechanism by which screening tools lose the trust of the people using them, and a driver of unnecessary CT, immobilisation and follow-up. Accuracy figures near 0.99 are dominated by true negatives and should never be quoted as the performance of a fracture detector.

Nothing here measures a patient outcome. All five clinical studies are retrospective; none randomised anybody, none followed a patient, and the scaphoid meta-analysis states explicitly that no study has compared AI with clinicians in a clinical setting. What the evidence does support, consistently and across independent groups, is narrower and still useful: AI assistance helps the non-specialist reader more than the specialist (miss rate 20% to 9% for clinicians with limited imaging training, against essentially no change for radiologists) and it shortens reading time by around a quarter. Those are workflow claims, they are reproducible, and they are the honest case for these tools.